Power grid transient fault mode classification method and system based on artificial intelligence

Abstract

The application provides a power grid transient fault mode classification method and system based on artificial intelligence, relates to the field of power system faults, and comprises the following steps: collecting multi-dimensional physical quantity data, extracting transient response strength and topological connectivity coupling calculation node state values, generating diffusion trajectories and extracting morphological features to determine fault types, source positions and boundaries, calculating time sequence gradient to trace propagation paths, identifying path intersection nodes to calculate cumulative contribution degrees to determine core nodes, configuring isolation trigger conditions for the core nodes to output a network segmentation scheme. The application can quickly and accurately identify faults, reduce power outage ranges, and improve power grid stability.

Description

Technical Field

[0001] This invention relates to power system fault technology, and more particularly to a method and system for classifying transient fault modes in power grids based on artificial intelligence. Background Technology

[0002] As a critical infrastructure, the safe and stable operation of the power grid system is vital to the national economy and social life. With the continuous expansion and increasing complexity of the power grid, transient faults occur frequently, potentially leading to widespread power outages and severe economic losses. Therefore, it is crucial to quickly and accurately identify the types of transient faults, locate their sources, and take effective measures. Traditional power grid fault analysis methods mainly rely on expert experience and simplified physical models, which are insufficient to cope with the complexity and dynamics of modern power grid systems. In recent years, with the development of artificial intelligence technology, data-driven fault analysis methods have been gradually applied to the field of power grid fault classification and location, but many challenges still remain.

[0003] Existing methods often focus on a single physical quantity or ignore grid topology information, failing to comprehensively reflect the propagation characteristics and impact range of faults, resulting in low accuracy in identifying complex faults. Traditional fault classification methods struggle to track fault propagation paths in real time and cannot effectively identify key nodes in the fault propagation process, thus hindering precise fault isolation and control. Furthermore, existing technologies lack a quantitative assessment of the contribution of fault propagation when determining network segmentation schemes, making it difficult to find the optimal isolation point location, potentially leading to over-isolation or under-isolation, impacting the stability and reliability of the power grid system.

[0004] As power grids develop towards intelligence and distribution, there is an urgent need for advanced technologies that can comprehensively utilize multi-dimensional physical quantity information and topological characteristics to accurately classify and locate transient faults and provide optimized isolation solutions, thereby improving the fault recovery capability and overall reliability of power grid systems. Summary of the Invention

[0005] This invention provides an artificial intelligence-based method and system for classifying transient fault modes in power grids, which can solve the problems in the prior art.

[0006] A first aspect of this invention provides an artificial intelligence-based method for classifying transient fault modes in power grids, comprising:

[0007] Collect time-series data of multidimensional physical quantities during transient disturbances in the power grid, extract the transient response intensity of the time-series data of multidimensional physical quantities, obtain the topological connectivity of the power grid topology, couple the transient response intensity with the topological connectivity to obtain the node state value, and generate the diffusion trajectory based on the time evolution of the node state value.

[0008] Extract the morphological features of the diffusion trajectory, map the morphological features to fault type identifiers, and determine the location of the fault source and the boundary of the fault impact based on the diffusion trajectory;

[0009] The temporal gradient of the node state value is calculated. The propagation path is obtained by tracing back from the fault source location along the temporal gradient to the affected nodes within the fault influence boundary. The propagation contribution of the propagation path is calculated based on topological connectivity and temporal gradient.

[0010] Identify the path intersection nodes of the propagation path in the power grid topology, sum up the propagation contribution of each path intersection node to obtain the cumulative contribution, and determine the core node set based on the cumulative contribution.

[0011] Configure isolation trigger conditions for the core node set based on the fault type identifier. When the isolation trigger conditions are detected, disconnect the topology connection of the corresponding node in the core node set and output the network segmentation scheme.

[0012] Collect time-series data of multidimensional physical quantities during transient disturbances in the power grid, extract the transient response intensity of the time-series data of multidimensional physical quantities, and obtain the topological connectivity of the power grid topology, including:

[0013] Multi-scale time-frequency decomposition is performed on the time series data of multidimensional physical quantities to obtain transient components in different frequency bands. The energy concentration degree and frequency drift trajectory of the transient components within the transient window are extracted. The transient response intensity is calculated based on the coupling strength of the energy concentration degree and frequency drift trajectory.

[0014] Read the physical connection relationship between power grid nodes, construct a weighted topology graph that reflects electrical distance and impedance transmission characteristics, calculate the effective conduction path and branch redundancy between node pairs in the weighted topology graph, and generate a dynamic propagation factor matrix as topology connectivity;

[0015] The transient response intensity is used as the excitation source input, and the dynamic propagation factor matrix is ​​used as the propagation medium parameter. The node state value is obtained by iterative calculation through the diffusion equation.

[0016] The numerical evolution of node state values ​​at continuous time points is tracked, the set of nodes whose state values ​​exceed the perturbation threshold is extracted, the activation time of each node in the set is recorded, and the set of nodes is connected in chronological order of the activation times to form a diffusion trajectory.

[0017] Extracting the morphological features of the diffusion trajectory and mapping these features to fault type identifiers, the determination of the fault source location and fault influence boundary based on the diffusion trajectory includes:

[0018] The activation time and topological position of each node in the diffusion trajectory are constructed into a set of spatiotemporal coordinate points. The spatiotemporal coordinate points are morphologically clustered to obtain multiple diffusion clusters. The spatial compactness and temporal extension of each diffusion cluster are extracted, and the cross ratio of spatial compactness and temporal extension is used as a morphological feature.

[0019] Establish interval division rules for the cross ratio of spatial density and temporal extension. When the cross ratio shows a spatial density-dominant pattern, it is identified as a localized concentrated fault. When the cross ratio shows a temporal extension-dominant pattern, it is identified as a chain propagation fault. When the cross ratio shows a balanced distribution pattern, it is identified as an oscillation-spreading fault. The fault type identifier is determined based on the dominant pattern of the cross ratio.

[0020] In each diffusion cluster, identify the node with the earliest activation time, calculate the temporal difference between each node and the other nodes in its diffusion cluster, and select the node with the largest temporal difference as the source of the fault.

[0021] Simultaneously, the decay gradient distribution of node state values ​​in the diffusion trajectory in the topological space is extracted, and nodes where the decay gradient distribution changes abruptly are connected to form the fault impact boundary.

[0022] The temporal gradient of node state values ​​is calculated, and the propagation path is obtained by tracing back along the temporal gradient from the fault source location to the affected nodes within the fault influence boundary. The propagation contribution of the propagation path is calculated based on topological connectivity and temporal gradient, including:

[0023] An evolution surface of the state values ​​of each node in the diffusion trajectory on the time axis is constructed. The evolution surface is decomposed locally into a linearized tangent plane to obtain the temporal change plane. The normal vector of the temporal change tangent plane is extracted as the temporal gradient.

[0024] Establish a vector field topology structure for temporal gradients, identify the source singular position corresponding to the fault source position and the sink singular position corresponding to each node within the fault influence boundary in the vector field topology structure, and extract the streamline trajectory connecting the source singular position and each sink singular position as the propagation path.

[0025] Calculate the topological betweenness and the divergence of the temporal gradient at each node along the propagation path. Use the nonlinear coupling value of the topological betweenness and the divergence as the node propagation potential energy. Integrate the node propagation potential energy along the propagation path to obtain the propagation contribution of the propagation path.

[0026] A vector field topology structure for temporal gradients is established. Within this structure, the singular positions of the source points corresponding to the fault origin and the singular positions of the sink points corresponding to each node within the fault influence boundary are identified. The streamline trajectories connecting the source point singular positions and each sink point singular position are extracted as propagation paths, including:

[0027] The distribution of temporal gradients in the power grid topology space is constructed as a vector field. The vector field is decomposed topologically to obtain the set of singular points and the manifold structure. The global topological invariants of the vector field are calculated based on the topological index of the set of singular points. The global topological invariants are mapped to the topology type of the vector field topology structure.

[0028] In the vector field topology, singular points with positive topological exponents are extracted as candidate source points. The topological distance between each candidate source point and the location of the fault source is calculated. The candidate source point with the smallest topological distance is selected as the singular location of the source point.

[0029] In the vector field topology, singular points with negative topological indices are extracted as candidate sinks. It is then determined whether each candidate sink is located within the fault influence boundary. Candidate sinks located within the fault influence boundary are designated as singular sink locations.

[0030] In the vector field topology, a family of streamlines is constructed with the singular position of the source point as the initial point. Each streamline in the family of streamlines is topologically extended to the singular position of the sink point. The streamlines extended from the singular position of the source point to the singular position of each sink point are extracted to form a set of streamline trajectories.

[0031] Calculate the trajectory deviation of each streamline in the streamline trajectory set under topological perturbation, and take the streamline with trajectory deviation lower than the median value of trajectory deviation in the streamline trajectory set as the propagation path.

[0032] Identify the path intersection nodes in the power grid topology, sum the propagation contributions of each path intersection node to obtain the cumulative contribution, and determine the core node set based on the cumulative contribution.

[0033] Construct a path coverage topology map of the propagation path in the power grid topology, count the path coverage frequency of each node, and extract nodes whose path coverage frequency exceeds the coverage frequency of a single propagation path as candidate intersection nodes;

[0034] Calculate the angular dispersion between the propagation paths at each candidate intersection node, and take the candidate intersection node whose angular dispersion is lower than the median value of the angular dispersion in the candidate intersection node set as the path intersection node;

[0035] Extract the propagation contribution of the propagation path passing through each path intersection node, calculate the deviation between the topological angle and the right angle of the propagation path at the path intersection node as the path weight factor, and weight and accumulate the propagation contribution and the path weight factor to obtain the cumulative contribution of the path intersection node.

[0036] A distribution field of the cumulative contribution of path intersection nodes on the power grid topology is constructed. The distribution field is decomposed into a gradient field to obtain a gradient vector field of cumulative contribution. Local maxima of the gradient vector magnitude are identified in the gradient vector field of cumulative contribution. The path intersection nodes corresponding to the local maxima are taken as the core node set.

[0037] Based on the fault type identifier, configure isolation trigger conditions for the core node set. When the isolation trigger conditions are detected, disconnect the topology connection of the corresponding node in the core node set and output the network segmentation scheme, including:

[0038] Determine the corresponding fault propagation direction based on the fault type identifier, trace the downstream propagation links of each node in the core node set along the fault propagation direction, and identify the load node density on each downstream propagation link.

[0039] Mark the core nodes of downstream propagation links whose load node density exceeds the average load node density in the core node set as priority isolation nodes, and configure isolation trigger conditions for priority isolation nodes.

[0040] Real-time monitoring of the fault current flow direction of each node in the core assembly, determining whether the fault current flow direction is consistent with the fault propagation direction, and extracting the flow rate change rate of the fault current when the fault current flow direction is consistent with the fault propagation direction, and matching the flow rate change rate with the isolation trigger condition to determine whether to trigger the isolation operation.

[0041] When an isolation operation is triggered, the upstream and downstream nodes of the priority isolation node in the power grid topology are located along the fault propagation direction. The topological connection between the priority isolation node and the upstream node is disconnected to block the fault propagation path. A network segmentation scheme is generated based on the disconnected topology.

[0042] A second aspect of the present invention provides an artificial intelligence-based power grid transient fault mode classification system, comprising:

[0043] The first unit is used to collect time-series data of multidimensional physical quantities during transient disturbances of the power grid, extract the transient response intensity of the time-series data of multidimensional physical quantities, obtain the topological connectivity of the power grid topology, couple the transient response intensity with the topological connectivity to obtain the node state value, and generate the diffusion trajectory based on the time evolution of the node state value.

[0044] The second unit is used to extract the morphological features of the diffusion trajectory, map the morphological features to the fault type identifier, and determine the location of the fault source and the boundary of the fault impact based on the diffusion trajectory.

[0045] The third unit is used to calculate the temporal gradient of the node state value. It traces back from the fault source location along the temporal gradient to the affected nodes within the fault influence boundary to obtain the propagation path. Based on topological connectivity and temporal gradient, it calculates the propagation contribution of the propagation path.

[0046] The fourth unit is used to identify the path intersection nodes of the propagation path in the power grid topology, accumulate the propagation contribution of each path intersection node to obtain the cumulative contribution, and determine the core node set based on the cumulative contribution.

[0047] The fifth unit is used to configure isolation trigger conditions for the core node set according to the fault type identifier. When the isolation trigger conditions are detected, the topology connection of the corresponding node in the core node set is disconnected and the network segmentation scheme is output.

[0048] In this embodiment, by coupling physical quantity time-series data with topology, a fault propagation trajectory is generated and morphological features are extracted, achieving accurate classification of transient fault modes in the power grid. This effectively improves the accuracy of fault identification and provides reliable assurance for the safe operation of the power grid. A gradient-based tracing method is used to determine the fault propagation path, and key nodes are identified by calculating propagation contribution, enabling precise location of the fault source and its impact range. This significantly shortens fault diagnosis time and reduces system operation and maintenance costs. Isolation trigger conditions are configured for core nodes based on fault type, achieving intelligent isolation and zoned control of power grid faults. This effectively prevents fault propagation and cascading failures, enhances the resilience and reliability of the power grid system, and provides technical support for the safe operation of the smart grid. Attached Figure Description

[0049] Figure 1 This is a flowchart illustrating the artificial intelligence-based power grid transient fault mode classification method according to an embodiment of the present invention.

[0050] Figure 2 This is a flowchart of the topology vector field analysis for identifying power grid fault propagation paths according to an embodiment of the present invention. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0052] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0053] Figure 1 This is a flowchart illustrating the artificial intelligence-based power grid transient fault mode classification method according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0054] Collect time-series data of multidimensional physical quantities during transient disturbances in the power grid, extract the transient response intensity of the time-series data of multidimensional physical quantities, obtain the topological connectivity of the power grid topology, couple the transient response intensity with the topological connectivity to obtain the node state value, and generate the diffusion trajectory based on the time evolution of the node state value.

[0055] Extract the morphological features of the diffusion trajectory, map the morphological features to fault type identifiers, and determine the location of the fault source and the boundary of the fault impact based on the diffusion trajectory;

[0056] The temporal gradient of the node state value is calculated. The propagation path is obtained by tracing back from the fault source location along the temporal gradient to the affected nodes within the fault influence boundary. The propagation contribution of the propagation path is calculated based on topological connectivity and temporal gradient.

[0057] Identify the path intersection nodes of the propagation path in the power grid topology, sum up the propagation contribution of each path intersection node to obtain the cumulative contribution, and determine the core node set based on the cumulative contribution.

[0058] Configure isolation trigger conditions for the core node set based on the fault type identifier. When the isolation trigger conditions are detected, disconnect the topology connection of the corresponding node in the core node set and output the network segmentation scheme.

[0059] In one optional implementation, collecting time-series data of multidimensional physical quantities during power grid transient disturbances, extracting the transient response intensity of the multidimensional physical quantity time-series data, and obtaining the topological connectivity of the power grid topology include:

[0060] Multi-scale time-frequency decomposition is performed on the time series data of multidimensional physical quantities to obtain transient components in different frequency bands. The energy concentration degree and frequency drift trajectory of the transient components within the transient window are extracted. The transient response intensity is calculated based on the coupling strength of the energy concentration degree and frequency drift trajectory.

[0061] Read the physical connection relationship between power grid nodes, construct a weighted topology graph that reflects electrical distance and impedance transmission characteristics, calculate the effective conduction path and branch redundancy between node pairs in the weighted topology graph, and generate a dynamic propagation factor matrix as topology connectivity;

[0062] The transient response intensity is used as the excitation source input, and the dynamic propagation factor matrix is ​​used as the propagation medium parameter. The node state value is obtained by iterative calculation through the diffusion equation.

[0063] The numerical evolution of node state values ​​at continuous time points is tracked, the set of nodes whose state values ​​exceed the perturbation threshold is extracted, the activation time of each node in the set is recorded, and the set of nodes is connected in chronological order of the activation times to form a diffusion trajectory.

[0064] In implementation, the first step is to collect multi-dimensional physical quantity time-series data during power grid disturbances. This data includes continuous time series of physical quantities such as voltage, current, frequency, and phase angle before and after the disturbance. The sampling frequency can be set to 10kHz, and the time window is from 0.2 seconds before the fault to 1 second after the fault, totaling 1.2 seconds of data. Data acquisition equipment can be deployed at substations and important line nodes to record and upload data in real time. Multi-scale time-frequency decomposition is then performed on the collected multi-dimensional physical quantity time-series data. Wavelet packet transform is used to decompose the original time-series data into transient components in three frequency bands: low frequency (0.1-5Hz), mid-frequency (5-50Hz), and high frequency (50-500Hz). Taking voltage data as an example, wavelet packet decomposition of the voltage waveform at a certain measurement point yields transient components in different frequency bands. In a real-world case, the bus voltage data of a 110kV substation was processed, and the decomposed low-frequency transient component exhibited obvious oscillation characteristics, with an amplitude of about 3% of the base value and a duration of about 200ms.

[0065] Extract the energy concentration and frequency drift trajectory of transient components within the transient window. Taking the moment of disturbance occurrence as the center, a 200ms time window is selected before and after it, and the energy concentration of transient components in each frequency band is calculated. The energy concentration is characterized by the ratio of the energy of the transient component within this time window to the energy of the entire observation time window. Simultaneously, the trajectory of the dominant frequency of the transient component within the transient window is tracked, and the frequency drift is recorded. The transient response intensity is calculated based on the coupling strength between the energy concentration and the frequency drift trajectory. The energy concentration value and the frequency drift are weighted and combined. The weighting coefficients are determined through historical case optimization; the weight for energy concentration is 0.7, and the weight for frequency drift is 0.3. For example, the calculation result for a certain node is: Transient response intensity = 0.7 × 78.6% + 0.3 × (0.4 / 1.2) = 0.55 + 0.10 = 0.65.

[0066] The physical connections between power grid nodes are read to construct a weighted topology graph. Physical connection information between nodes is extracted from the power grid management system, including parameters such as line type, rated capacity, and line length. For connections between two nodes, the electrical distance (considering line impedance) is calculated instead of the simple geometric distance. Taking a regional power grid as an example, it contains 15 nodes and 20 lines, with line impedance ranging from 0.012 to 0.089 ohms / km and line lengths between 5 and 30 kilometers.

[0067] Calculate the effective conductance path and branch redundancy between node pairs in the weighted topology graph. The effective conductance path is the comprehensive admittance value considering all possible paths, and the branch redundancy represents the quantity and quality of alternative paths between two nodes. For the above-mentioned regional power grid, the effective conductance path value between node 1 and node 5 is 8.76 Siemens, and the branch redundancy is 2.3.

[0068] A dynamic propagation factor matrix is ​​generated to represent topological connectivity. Based on the effective conductance path and branch redundancy, the dynamic propagation factor matrix is ​​established. Matrix element values ​​represent the degree of attenuation of a disturbance propagating from one node to another. In the example matrix, the propagation factor for nodes 1 to 5 is 0.83, indicating that 83% of the disturbance intensity is retained during propagation.

[0069] The transient response intensity is used as the excitation source input, and the dynamic propagation factor matrix is ​​used as the propagation medium parameter. The node state values ​​are calculated iteratively using the diffusion equation. Initially, all node state values ​​are set to 0. The calculated transient response intensity is used as the external excitation source intensity for each node and substituted into the diffusion equation for iterative calculation. During the iteration process, the node state values ​​change with time, exhibiting nonlinear diffusion characteristics. In the test case, a time step of 10ms was used, with a total of 100 steps.

[0070] The algorithm tracks the numerical evolution of node state values ​​over consecutive time points and extracts the set of nodes that exceed a perturbation threshold. A perturbation threshold of 0.35 is set; a node is considered activated when its state value exceeds this threshold. The activation time of each node is recorded; for example, the activation times of the node sequence {3, 5, 7, 2, 9} are {0ms, 30ms, 50ms, 70ms, 90ms}, respectively.

[0071] The nodes are connected in chronological order of activation to form a diffusion trajectory. A diffusion trajectory diagram is created based on the activation sequence of the nodes, with directed line segments connecting adjacent activated nodes. The direction of the line segments indicates the direction of disturbance propagation. For example, if the disturbance source is located at node 3, the diffusion trajectory is 3→5→7→2→9, which is consistent with the disturbance propagation sequence recorded in the actual fault waveform data of the power grid.

[0072] Based on the above technical solution, transient response features can be accurately extracted from multi-dimensional physical quantity time-series data. Combined with the weighted connectivity of the power grid topology, the disturbance propagation path can be dynamically characterized, enabling the tracking of the temporal evolution of node state values. This method can identify abnormal nodes and their activation order, reconstruct disturbance propagation trajectories, improve the accuracy and response speed of transient fault detection, and provide high-precision decision-making basis for power grid stability analysis, fault location, and emergency dispatch.

[0073] In one optional implementation, extracting the morphological features of the diffusion trajectory, mapping these features to a fault type identifier, and determining the fault source location and fault impact boundary based on the diffusion trajectory includes:

[0074] The activation time and topological position of each node in the diffusion trajectory are constructed into a set of spatiotemporal coordinate points. The spatiotemporal coordinate points are morphologically clustered to obtain multiple diffusion clusters. The spatial compactness and temporal extension of each diffusion cluster are extracted, and the cross ratio of spatial compactness and temporal extension is used as a morphological feature.

[0075] Establish interval division rules for the cross ratio of spatial density and temporal extension. When the cross ratio shows a spatial density-dominant pattern, it is identified as a localized concentrated fault. When the cross ratio shows a temporal extension-dominant pattern, it is identified as a chain propagation fault. When the cross ratio shows a balanced distribution pattern, it is identified as an oscillation-spreading fault. The fault type identifier is determined based on the dominant pattern of the cross ratio.

[0076] In each diffusion cluster, identify the node with the earliest activation time, calculate the temporal difference between each node and the other nodes in its diffusion cluster, and select the node with the largest temporal difference as the source of the fault.

[0077] Simultaneously, the decay gradient distribution of node state values ​​in the diffusion trajectory in the topological space is extracted, and nodes where the decay gradient distribution changes abruptly are connected to form the fault impact boundary.

[0078] First, the diffusion trajectory data in the network system is acquired, including the activation time and topological location information of each node. This information is then used to construct a spatiotemporal coordinate point set, where each point is represented by both the spatial location coordinates and the time coordinates of the node. For example, in a power network, the fault activation time of substation A can be recorded as 12:30:45, with location coordinates of (35.6, 78.2), thus forming the spatiotemporal coordinate point (35.6, 78.2, 12:30:45).

[0079] A density-based clustering algorithm was applied to the constructed spatiotemporal coordinate point set for morphological clustering, resulting in multiple diffusion clusters. During density-based clustering, a spatiotemporal distance threshold of 8 units and a density threshold of 5 points were set. Through clustering, the original 234 nodes were divided into four distinct diffusion clusters, containing 78, 56, 62, and 38 nodes respectively. For each diffusion cluster, its spatial compactness and temporal spread were calculated. Spatial compactness is characterized by the reciprocal of the standard deviation of the spatial coordinates of the nodes within the cluster; a larger value indicates a more concentrated spatial distribution. Temporal spread is characterized by the ratio of the maximum span of activation times of nodes within the cluster to the number of nodes; a larger value indicates more significant temporal propagation.

[0080] Taking the first diffusion cluster as an example, the standard deviation of its 78 nodes' spatial coordinates is 2.3 units, resulting in a spatial compactness of 0.435. The earliest activation time is 12:30:45, and the latest is 12:42:18, spanning 11 minutes and 33 seconds, which translates to 693 seconds. Therefore, the temporal stretch is 693 / 78 = 8.885. The crossover ratio between spatial compactness and temporal stretch is then calculated. For the first diffusion cluster, the crossover ratios include compactness / stretch = 0.049 and stretch / compactness = 20.425. These two crossover ratios together constitute the morphological feature vector of this diffusion cluster [0.049, 20.425].

[0081] Based on morphological feature vectors, fault type identification rules are established. When density / extension > 1 and extension / density < 0.5, it is identified as a localized fault; when density / extension < 0.1 and extension / density > 10, it is identified as a cascading fault; when 0.1 ≤ density / extension ≤ 1 and 0.5 ≤ extension / density ≤ 10, it is identified as an oscillating diffusion fault. For the first diffusion cluster, its morphological feature vector [0.049, 20.425] satisfies the criteria for a cascading fault, and therefore it is identified as a cascading fault. Similarly, the morphological feature vector of the second diffusion cluster is [1.25, 0.32], and it is identified as a localized fault; the morphological feature vector of the third diffusion cluster is [0.45, 2.22], and it is identified as an oscillating diffusion fault.

[0082] When determining the source of the fault in each diffusion cluster, the first step is to identify the set of nodes with the earliest activation times within each cluster. For example, in the first diffusion cluster, three nodes have activation times of 12:30:45: A15, B23, and C47. The temporal dissimilarity between these nodes and other nodes in the cluster is calculated. The temporal dissimilarity is defined as the sum of the squares of the differences in activation times between the node and all other nodes. The temporal dissimilarity of node A15 is 86452s. 2 The timing difference of node B23 is 92641s. 2 The timing difference of node C47 is 76329s.2 The node with the largest timing difference was selected as the source of the fault.

[0083] To determine the fault impact boundary, the attenuation gradient distribution of node state values ​​in the topological space is extracted from the diffusion trajectory. Node state values ​​can be physical quantities such as voltage deviation, flow anomalies, or signal strength. The rate of change of state values ​​between adjacent nodes is calculated to form an attenuation gradient field. When the rate of change of state values ​​between two adjacent nodes exceeds a preset threshold of 0.25, it is determined to be a gradient abrupt change point. In this example, a total of 47 gradient abrupt change point pairs were detected. These node pairs are connected to form a closed loop, constituting the fault impact boundary. The boundary morphology varies significantly depending on the fault type: the impact boundary of a locally concentrated fault is approximately circular with a radius of about 12 units; the impact boundary of a cascading fault is band-shaped, with a length of about 35 units and a width of about 8 units; the impact boundary of an oscillating diffusion fault is irregularly shaped, with a maximum diameter of about 26 units.

[0084] Based on the above technical solutions, fault types can be accurately identified by the spatiotemporal morphological characteristics of the diffusion trajectory, distinguishing between localized, chain-propagation, and oscillating diffusion faults. At the same time, the fault source node can be located, and the fault impact boundary can be determined by using the decay gradient of the node state value. This enables visualized analysis and refined positioning of power grid disturbance propagation, providing reliable technical support for rapid fault diagnosis, risk assessment, and precise dispatching.

[0085] In one optional implementation, the temporal gradient of the node state values ​​is calculated, and the propagation path is obtained by tracing back along the temporal gradient from the fault source location to the affected nodes within the fault influence boundary. The propagation contribution of the propagation path is calculated based on topological connectivity and the temporal gradient, including:

[0086] An evolution surface of the state values ​​of each node in the diffusion trajectory on the time axis is constructed. The evolution surface is decomposed locally into a linearized tangent plane to obtain the temporal change plane. The normal vector of the temporal change tangent plane is extracted as the temporal gradient.

[0087] Establish a vector field topology structure for temporal gradients, identify the source singular position corresponding to the fault source position and the sink singular position corresponding to each node within the fault influence boundary in the vector field topology structure, and extract the streamline trajectory connecting the source singular position and each sink singular position as the propagation path.

[0088] Calculate the topological betweenness and the divergence of the temporal gradient at each node along the propagation path. Use the nonlinear coupling value of the topological betweenness and the divergence as the node propagation potential energy. Integrate the node propagation potential energy along the propagation path to obtain the propagation contribution of the propagation path.

[0089] During implementation, the first step is to acquire the time-series status data of each node in the network system. This data can include performance indicators, failure rates, or other quantifiable status parameters of each node during system operation. For example, in a distributed system with 100 nodes, the system collects status data such as CPU utilization, memory usage, and response time of each node at 5-minute intervals over a continuous 72-hour period.

[0090] To construct the evolution surface of node state values ​​along the time axis in the diffusion trajectory, the specific implementation involves mapping the node state values ​​into a three-dimensional space. The x-axis and y-axis represent the two-dimensional coordinates of the nodes in the network topology, respectively, while the z-axis represents the node state value. The time dimension is represented through dynamic evolution. Taking a mesh network with 50 nodes as an example, during a fault, observation shows that the state value of node 25 sharply increases from the normal value of 0.2 to 0.9 and gradually spreads to surrounding nodes. These state values ​​change over time, forming a dynamically evolving surface.

[0091] When performing local linearization decomposition on the evolution surface, a time window of 30 minutes is selected. At each time point t, a plane fitting is performed on the state values ​​of each node and its neighboring nodes. Specifically, in the local region of node i, a local plane is fitted using the state values ​​of itself and its directly connected neighboring node j through the least squares method. In practice, for a node with 6 neighbors, the state values ​​of these 7 points at time t are collected to construct a local linear model. Thus, at each time point t, each node i corresponds to a local tangent plane. In the process of extracting the normal vector of the temporally changing tangent plane as the temporal gradient, the normal vector of each node i at time t is calculated for its local tangent plane. The direction of the normal vector indicates the direction of the fastest change in the node's state value, and the magnitude of the normal vector represents the magnitude of the rate of change. The temporal gradient vector of the fault source node 25 has a magnitude of 0.15 and points out of the network, while the temporal gradient vector of the affected nodes points in the direction of the source node.

[0092] When establishing the vector field topology of temporal gradients, the temporal gradient vectors of all nodes are organized into a vector field. Special topologies, particularly singularities, are identified within this vector field. Singularities are points in the vector field where the magnitude of the vector is zero or the direction is uncertain. According to vector field theory, singularities can be classified into source points, sink points, saddle points, etc. Source points are points where all vectors point outwards, corresponding to fault origins; sink points are points where all vectors point inwards, corresponding to affected terminal nodes.

[0093] In identifying the singular locations of source points corresponding to the fault origin, the loop integral value around each node is calculated. When the loop integral value is positive and large, the point is likely to be the source point. In practice, the loop integral value of node 25 is 1.57, much higher than other nodes, and it is identified as the fault source point. For the singular locations of sink points corresponding to each node within the fault influence boundary, the loop integral is also calculated. When the value is negative and large, the point is the sink point. In the example, the loop integrals of nodes 31, 42, and 19 are -1.32, -1.28, and -1.45, respectively, and they are identified as sink points, representing the main affected terminals of the fault.

[0094] When extracting the streamline trajectories connecting the singular locations of the source and each sink as propagation paths, we start from each sink and trace step-by-step along the opposite direction of the temporal gradient vector until we reach the source. This yields the propagation path from the source to each sink. For example, the propagation path from source 25 to sink 31 passes through nodes 25→26→27→28→31, indicating that the fault propagated from node 25 to node 31 along this path.

[0095] In calculating the topological betweenness number of each node along the propagation path, all shortest paths in the network are statistically analyzed, and the percentage of times each node appears on these paths is calculated. Nodes with high topological betweenness numbers play a crucial role in network communication. In the example network, node 27 has a topological betweenness number of 0.42, meaning that 42% of the shortest paths pass through this node, indicating that it is a critical node for fault propagation.

[0096] For calculating the divergence of temporal gradients at nodes, we examine the difference between the temporal gradient vector of each node and the temporal gradient vectors of all its neighbors. A positive divergence indicates that the node is a diffusion source, and a negative divergence indicates that the node is a convergence point. The divergence of node 25 is 0.38, confirming it as a fault source; the divergences of nodes 31, 42, and 19 are -0.29, -0.32, and -0.27, respectively, verifying that they are influencing the terminals.

[0097] When using the nonlinear coupling value of topological betweenness and divergence as the node propagation potential energy, a weighted product of the two values ​​is employed. For example, the propagation potential energy of node 27 is calculated as 0.42 × 0.15 = 0.063. The propagation contribution of the propagation path is obtained by integrating the node propagation potential energy along the path, which is the sum of the propagation potential energy of all nodes on the path. The path propagation contribution from 25 to 31 is 0.247, from 25 to 42 is 0.215, and from 25 to 19 is 0.231, indicating that the fault has the greatest impact on node 31.

[0098] Using the above methods, the fault source node 25 was successfully identified, and the main affected nodes 31, 42, and 19 were determined. The fault propagation path and its contribution were quantified, providing a reliable basis for fault location and prevention. Visual reconstruction and impact analysis of the fault propagation path were achieved, providing quantitative evidence for research on fault transmission mechanisms, identification of key nodes, and optimization of power grid protection strategies.

[0099] like Figure 2 The diagram illustrates the topology vector field analysis process for identifying power grid fault propagation paths in this embodiment.

[0100] In one optional implementation, a vector field topology of temporal gradients is established. Within this topology, the source singularity corresponding to the fault origin and the sink singularity corresponding to each node within the fault influence boundary are identified. The streamline trajectories connecting the source singularity and each sink singularity are extracted as propagation paths, including:

[0101] The distribution of temporal gradients in the power grid topology space is constructed as a vector field. The vector field is decomposed topologically to obtain the set of singular points and the manifold structure. The global topological invariants of the vector field are calculated based on the topological index of the set of singular points. The global topological invariants are mapped to the topology type of the vector field topology structure.

[0102] In the vector field topology, singular points with positive topological exponents are extracted as candidate source points. The topological distance between each candidate source point and the location of the fault source is calculated. The candidate source point with the smallest topological distance is selected as the singular location of the source point.

[0103] In the vector field topology, singular points with negative topological indices are extracted as candidate sinks. It is then determined whether each candidate sink is located within the fault influence boundary. Candidate sinks located within the fault influence boundary are designated as singular sink locations.

[0104] In the vector field topology, a family of streamlines is constructed with the singular position of the source point as the initial point. Each streamline in the family of streamlines is topologically extended to the singular position of the sink point. The streamlines extended from the singular position of the source point to the singular position of each sink point are extracted to form a set of streamline trajectories.

[0105] Calculate the trajectory deviation of each streamline in the streamline trajectory set under topological perturbation, and take the streamline with trajectory deviation lower than the median value of trajectory deviation in the streamline trajectory set as the propagation path.

[0106] When a power system fault occurs, the first step is to collect power grid operation status monitoring data, including time series data of key parameters such as voltage, current, and frequency before and after the fault. This monitoring data is then sorted by timestamp and preprocessed to eliminate outliers and noise interference, ensuring data quality. The rate of change between adjacent time points is calculated from the preprocessed time series data, forming a time series gradient matrix. This matrix reflects the rate and direction of change of parameters at each node of the power grid over time, providing a foundation for subsequent analysis.

[0107] When constructing a vector field in the power grid topology space, the physical topology of the power grid is mapped to a two-dimensional plane coordinate system, with each node corresponding to a coordinate point, and the connection relationships between nodes remaining unchanged. At each coordinate point, a vector is determined based on the temporal gradient value of that node. The magnitude of the vector represents the gradient magnitude, and the direction represents the gradient direction. For example, if the voltage at node A drops from 220V to 215V, and the calculated gradient is -5V / s, then a vector of length 5 pointing in the direction of voltage decrease is drawn at node A. In this way, a continuous vector field distribution is formed across the entire plane.

[0108] A topological decomposition is performed on the constructed vector field to identify singularities. Singularities are special locations in a vector field where the magnitude of a vector is zero or its direction is uncertain. These singularities can be located by calculating the curl and divergence of the vector field. For example, when vectors around a point are radially distributed, that point is a source point; when vectors are convergent, that point is a sink point. A topological index is calculated for each identified singularity; the topological index for the source point is positive, and the topological index for the sink point is negative. By summing the topological indices of all singularities in the entire vector field, a global topological invariant is obtained. This invariant can be used to determine the topological type of the vector field, such as saddle point, vortex, or nodal type.

[0109] When identifying the singular location of the source point corresponding to the fault origin, all singular points with positive topological indices are extracted from the vector field topology as candidate source points. For example, in a power grid with 50 nodes, five candidate source points with positive topological indices may be identified. For each candidate source point, the topological distance between it and the known fault origin location is calculated. The topological distance can be determined by calculating the number of edges on the shortest path between the two points. The shortest path from candidate source point S1 to the fault origin location contains 3 edges, so the topological distance is 3. The candidate source point with the smallest topological distance is selected as the final determined singular location of the source point.

[0110] When identifying singular locations of sinks for each node within the fault influence boundary, all singular points with negative topological indices are extracted from the vector field topology as candidate sinks. In the same power grid, up to eight candidate sinks with negative topological indices may be identified. The fault influence boundary is determined based on fault monitoring data; node parameters within this boundary deviate from normal values ​​exceeding a preset threshold. Nodes whose voltage deviates from the rated value by more than 5% are considered to be located within the fault influence boundary. Each candidate sink is then determined to be within this boundary, and those located within the boundary are retained as the final determined singular locations of the sinks.

[0111] When extracting streamline trajectories connecting the singular locations of the source and sinks, a family of streamlines is constructed along the vector field direction, using the determined singular location of the source as the initial point. A streamline is a curve along the vector field direction, with its tangent vector aligning with the vector field direction at each point. Starting from the source S, a small step is taken along the vector direction of the current point, and the trajectory is recorded until the vector field boundary or sink is reached. Each streamline in the streamline family is topologically extended to ensure it extends to the singular locations of the sinks. All streamlines extended from the singular location of the source to the singular locations of the sinks are extracted, forming a set of streamline trajectories.

[0112] To determine the most probable fault propagation path, the trajectory deviation of each streamline in the streamline trajectory set was calculated under topological perturbation. A small random perturbation was introduced into the vector field, and the degree of change in each streamline trajectory was observed. The lower the trajectory deviation, the more stable the streamline is, and the higher its probability of being a propagation path. The trajectory deviation was calculated for 10 streamlines, yielding values ​​of {0.12, 0.08, 0.15, 0.25, 0.05, 0.18, 0.09, 0.21, 0.11, 0.14}, with a median value of 0.13. The five streamlines with a trajectory deviation below 0.13, corresponding to {0.12, 0.08, 0.05, 0.09, 0.11}, were selected as the final determined fault propagation paths.

[0113] The propagation paths determined by the above methods intuitively demonstrate the way faults spread in the power grid, providing a scientific basis for power grid operation and maintenance personnel to formulate fault isolation and system recovery strategies. This method can accurately identify singular locations of source and sink points in the power grid's time-series gradient vector field, construct a set of streamline trajectories from the fault source to the influence boundary, and select stable streamlines as propagation paths. This method achieves high-precision reconstruction and topological stability analysis of fault propagation paths, providing reliable technical support for fault propagation law research, critical node location, and power grid risk prevention and control.

[0114] In one optional implementation, the path intersection nodes of the propagation path in the power grid topology are identified, and the propagation contribution of each path intersection node is accumulated to obtain a cumulative contribution. The core node set is determined based on the cumulative contribution, including:

[0115] Construct a path coverage topology map of the propagation path in the power grid topology, count the path coverage frequency of each node, and extract nodes whose path coverage frequency exceeds the coverage frequency of a single propagation path as candidate intersection nodes;

[0116] Calculate the angular dispersion between the propagation paths at each candidate intersection node, and take the candidate intersection node whose angular dispersion is lower than the median value of the angular dispersion in the candidate intersection node set as the path intersection node;

[0117] Extract the propagation contribution of the propagation path passing through each path intersection node, calculate the deviation between the topological angle and the right angle of the propagation path at the path intersection node as the path weight factor, and weight and accumulate the propagation contribution and the path weight factor to obtain the cumulative contribution of the path intersection node.

[0118] A distribution field of the cumulative contribution of path intersection nodes on the power grid topology is constructed. The distribution field is decomposed into a gradient field to obtain a gradient vector field of cumulative contribution. Local maxima of the gradient vector magnitude are identified in the gradient vector field of cumulative contribution. The path intersection nodes corresponding to the local maxima are taken as the core node set.

[0119] This embodiment identifies a set of core nodes based on propagation paths within the power grid topology. During power grid propagation, specific nodes often become the intersection points of multiple propagation paths, and these intersection nodes have a significant impact on the overall propagation effect. By analyzing the distribution and intersection characteristics of propagation paths in the topology, the set of core nodes is identified, providing a basis for power grid monitoring and optimization.

[0120] When constructing a path coverage topology map of propagation paths in a power grid topology, the first step is to acquire power grid topology data, including node coordinates, connectivity, and power transmission parameters. Assuming a power grid contains 100 nodes and 150 edges, 20 main propagation paths are obtained through simulation. Each propagation path is marked, recording all nodes it covers. Path 1 covers nodes {1, 5, 8, 12, 15, 20}, and path 2 covers nodes {3, 5, 8, 10, 15, 18}. The frequency of each node being covered by a path is counted; for example, node 5 is covered by 10 paths, node 8 by 12 paths, and node 15 by 15 paths. Nodes whose path coverage frequency exceeds the coverage frequency of a single propagation path are selected as candidate intersection nodes. In this example, if a single path covers an average of 6 nodes, nodes with a coverage frequency greater than 6, such as nodes 5, 8, and 15, are marked as candidate intersection nodes, resulting in a total of 25 candidate intersection nodes.

[0121] Calculate the angular dispersion between propagation paths at each candidate intersection node. For each candidate intersection node, analyze the direction vectors of all propagation paths passing through that node. Taking node 8 as an example, calculate the direction vectors of the 12 paths passing through this node, which can be represented as unit vectors pointing from node 8 to adjacent nodes. Calculate the angles between each pair of these direction vectors to obtain a set of angles. Calculate the standard deviation of these angles as the angular dispersion, representing the degree of dispersion of path directions. Calculate the angular dispersion for all candidate intersection nodes; for example, the angular dispersion of node 5 is 0.8, that of node 8 is 0.4, and that of node 15 is 0.6. Calculate the median value of the angular dispersion for all candidate intersection nodes. Candidate intersection nodes with angular dispersion lower than the median value, such as nodes 8 and 15, are identified as path intersection nodes, resulting in a total of 12 path intersection nodes.

[0122] The propagation contribution of each propagation path passing through each path intersection node is extracted. The propagation contribution, based on power flow analysis results, represents the degree of influence of the path on system stability. Taking node 8 as an example, the propagation contributions of the 12 propagation paths passing through this node are {0.75, 0.82, 0.68, 0.91, 0.78, 0.85, 0.73, 0.79, 0.81, 0.76, 0.83, 0.72}. The deviation between the topological angle and the right angle of the propagation path at the path intersection node is calculated as the path weight factor. Taking path 1 on node 8 as an example, its angle with the horizontal direction is 65 degrees, and its deviation from the right angle (90 degrees) is 25 degrees. After normalizing this deviation, the path weight factor is 0.72. Weighting factors were calculated for all 12 paths, resulting in {0.72, 0.85, 0.78, 0.92, 0.65, 0.83, 0.79, 0.88, 0.71, 0.81, 0.75, 0.69}. The propagation contribution was weighted and accumulated with the path weighting factors, yielding a cumulative contribution of 7.53 for node 8. Similarly, the cumulative contributions of all intersecting nodes of the paths were calculated.

[0123] A distribution field of the cumulative contribution of path intersection nodes on the power grid topology is constructed. The cumulative contribution values ​​of 12 path intersection nodes are mapped onto the power grid topology to form a spatial distribution field. A continuous distribution field is generated using interpolation methods; for example, the cumulative contribution of node 8 is 7.53, and the cumulative contribution of node 15 is 8.12. The cumulative contributions of adjacent regions are obtained through linear interpolation. The distribution field is decomposed into a gradient field, and the rate of change and direction of the cumulative contribution at each location are calculated to obtain the cumulative contribution gradient vector field. Local maxima of the gradient vector magnitude in the gradient vector field are identified; these points represent the locations where the cumulative contribution changes most drastically. Five local maxima are identified, corresponding to nodes 8, 15, 23, 42, and 67. These five nodes constitute the core node set, which plays a crucial role in the power grid propagation process and can be used as key monitoring and protection targets.

[0124] Through the above implementation method, key intersection nodes of propagation paths can be identified in the power grid topology, the cumulative propagation contribution of each node can be quantified, a contribution gradient field can be constructed, and local maxima points can be extracted, thereby forming a core node set. This method enables precise location of key nodes in power grid fault propagation, reveals fault transmission hubs, and improves the accuracy and reliability of decision-making in fault protection, risk assessment, and scheduling optimization.

[0125] In one optional implementation, configuring isolation trigger conditions for the core node set based on the fault type identifier, and disconnecting the topology connection of the corresponding node in the core node set and outputting a network segmentation scheme when the isolation trigger conditions are detected to be met, includes:

[0126] Determine the corresponding fault propagation direction based on the fault type identifier, trace the downstream propagation links of each node in the core node set along the fault propagation direction, and identify the load node density on each downstream propagation link.

[0127] Mark the core nodes of downstream propagation links whose load node density exceeds the average load node density in the core node set as priority isolation nodes, and configure isolation trigger conditions for priority isolation nodes.

[0128] The fault current flow direction of each node in the core section is monitored in real time to determine whether the fault current flow direction is consistent with the fault propagation direction. When the fault current flow direction is consistent with the fault propagation direction, the flow rate change rate of the fault current is extracted. The flow rate change rate is matched with the isolation trigger condition to determine whether to trigger the isolation operation. When the isolation operation is determined to be triggered, the upstream and downstream nodes of the priority isolation node in the power grid topology are located along the fault propagation direction. The topology connection between the priority isolation node and the upstream node is disconnected to block the fault propagation path. A network segmentation scheme is generated based on the disconnected topology.

[0129] The direction of fault propagation is determined based on the fault type identifier in the power grid system. For example, for a single-phase ground fault, the fault propagation direction is usually from the fault point outwards; for a three-phase short-circuit fault, the fault propagation direction is along the path of least impedance. When a three-phase short-circuit fault is detected in a substation, the direction of fault propagation is determined based on the distribution characteristics of the fault current, mainly spreading downstream along the main line to the load area.

[0130] The system traces the downstream propagation links of each node in the core node set along a defined fault propagation direction. The core node set can be a critical substation, important switching equipment, or connection point in the power grid. The system tracks the downstream links of each core node using the power grid topology map and identifies the load node density on each downstream propagation link. Load node density is calculated by statistically analyzing the number of users connected per unit length of link or the load capacity per unit area. In a certain distribution network, the downstream link L1 of node N1 covers an area containing 150 user nodes with a link length of 5 kilometers, resulting in a load node density of 30 nodes / kilometer; while the downstream link L2 of node N2 covers 200 user nodes with a link length of 10 kilometers, resulting in a load node density of 20 nodes / kilometer.

[0131] Calculate the mean load node density in the core node set. Assuming the core node set contains 10 nodes, the load node densities of their downstream links are 30, 20, 25, 15, 18, 22, 28, 19, 24, and 21 nodes / km, respectively, and the calculated mean is 22.2 nodes / km. The core nodes corresponding to downstream propagation links with load node densities exceeding the mean are marked as priority isolation nodes. In the example above, the load node densities of nodes N1(30), N3(25), N7(28), and N9(24) exceed the mean of 22.2 nodes / km, therefore these four nodes are marked as priority isolation nodes.

[0132] Isolation trigger conditions are configured for the marked priority isolation nodes. These conditions are primarily based on the rate of change of fault current. Different thresholds are set for different types of faults. For example, for a three-phase short-circuit fault, isolation is triggered when the rate of change of fault current exceeds 200% / second; for a single-phase ground fault, isolation is triggered when the rate of change of fault current exceeds 150% / second. These thresholds can be adjusted based on historical fault data and grid characteristics. In the actual configuration, the isolation trigger condition for priority isolation node N1 is set to a rate of change of fault current exceeding 180% / second.

[0133] The system monitors the fault current flow direction of each node in the core node set in real time. Current sensors installed on each node collect current data to determine if the fault current flow direction matches the pre-determined fault propagation direction. When the fault current flow direction matches the fault propagation direction, the system begins extracting the fault current flow rate change data. In node N1, the fault current suddenly increases from 10A to 40A within a 0.1-second interval, resulting in a calculated flow rate change rate of 300% / second, significantly exceeding the set isolation trigger threshold of 180% / second. The extracted flow rate change rate is matched with the isolation trigger condition to determine whether to trigger an isolation operation. The system compares the calculated flow rate change rate with the preset trigger threshold; if the flow rate change rate exceeds the threshold, an isolation operation is triggered. In the example above, the flow rate change rate of node N1 (300% / second) exceeds the trigger threshold of 180% / second, indicating that isolation operation is required for node N1.

[0134] When an isolation operation is triggered, the system locates the upstream and downstream nodes of the priority isolation node in the power grid topology along the fault propagation direction. For example, the upstream node of node N1 is substation A, and the downstream nodes are distribution boxes B and C. The system commands to disconnect the topological connection between node N1 and the upstream node substation A, and achieves physical isolation by controlling the corresponding switching equipment, thereby blocking the path of fault propagation from substation A to node N1 and its downstream nodes.

[0135] A network segmentation scheme is generated based on the disconnected topology. The segmentation scheme includes isolation boundaries, the extent of the affected area, and power restoration recommendations. After segmentation, two independent subnets are formed: Subnet 1 includes substation A and its connected unaffected areas; Subnet 2 includes node N1 and the area covered by its downstream distribution boxes B and C. It is recommended to provide emergency power to subnet 2 via a backup power source, while simultaneously adjusting the load distribution of other lines within subnet 1 to ensure the overall stable operation of the power grid.

[0136] Based on the above technical solution, core nodes can be intelligently identified and isolation trigger conditions can be configured according to the fault type. Fault current changes can be monitored in real time, and the topology connections of critical nodes can be disconnected first to block fault propagation paths and achieve local grid isolation. This effectively prevents fault spread, ensures stable grid operation, and generates network segmentation schemes, providing accurate decision-making basis for fault emergency dispatch and load optimization.

[0137] A second aspect of this invention provides an artificial intelligence-based classification system for transient fault modes in power grids, the system comprising:

[0138] The first unit is used to collect time-series data of multidimensional physical quantities during transient disturbances of the power grid, extract the transient response intensity of the time-series data of multidimensional physical quantities, obtain the topological connectivity of the power grid topology, couple the transient response intensity with the topological connectivity to obtain the node state value, and generate the diffusion trajectory based on the time evolution of the node state value.

[0139] The second unit is used to extract the morphological features of the diffusion trajectory, map the morphological features to the fault type identifier, and determine the location of the fault source and the boundary of the fault impact based on the diffusion trajectory.

[0140] The third unit is used to calculate the temporal gradient of the node state value. It traces back from the fault source location along the temporal gradient to the affected nodes within the fault influence boundary to obtain the propagation path. Based on topological connectivity and temporal gradient, it calculates the propagation contribution of the propagation path.

[0141] The fourth unit is used to identify the path intersection nodes of the propagation path in the power grid topology, accumulate the propagation contribution of each path intersection node to obtain the cumulative contribution, and determine the core node set based on the cumulative contribution.

[0142] The fifth unit is used to configure isolation trigger conditions for the core node set according to the fault type identifier. When the isolation trigger conditions are detected, the topology connection of the corresponding node in the core node set is disconnected and the network segmentation scheme is output.

[0143] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0144] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A power grid transient fault mode classification method based on artificial intelligence, characterized in that, include: Collect time-series data of multidimensional physical quantities during transient disturbances in the power grid, extract the transient response intensity of the time-series data of multidimensional physical quantities, obtain the topological connectivity of the power grid topology, couple the transient response intensity with the topological connectivity to obtain the node state value, and generate the diffusion trajectory based on the time evolution of the node state value. Extract the morphological features of the diffusion trajectory, map the morphological features to fault type identifiers, and determine the location of the fault source and the boundary of the fault impact based on the diffusion trajectory; The temporal gradient of the node state value is calculated. The propagation path is obtained by tracing back from the fault source location along the temporal gradient to the affected nodes within the fault influence boundary. The propagation contribution of the propagation path is calculated based on topological connectivity and temporal gradient. Identify the path intersection nodes of the propagation path in the power grid topology, sum up the propagation contribution of each path intersection node to obtain the cumulative contribution, and determine the core node set based on the cumulative contribution. Configure isolation trigger conditions for the core node set according to the fault type identifier. When the isolation trigger conditions are detected, disconnect the topology connection of the corresponding node in the core node set and output the network segmentation scheme. Identify the path intersection nodes in the power grid topology, sum the propagation contributions of each path intersection node to obtain the cumulative contribution, and determine the core node set based on the cumulative contribution. Construct a path coverage topology map of the propagation path in the power grid topology, count the path coverage frequency of each node, and extract nodes whose path coverage frequency exceeds the coverage frequency of a single propagation path as candidate intersection nodes; Calculate the angular dispersion between the propagation paths at each candidate intersection node, and take the candidate intersection node whose angular dispersion is lower than the median value of the angular dispersion in the candidate intersection node set as the path intersection node; Extract the propagation contribution of the propagation path passing through each path intersection node, calculate the deviation between the topological angle and the right angle of the propagation path at the path intersection node as the path weight factor, and weight and accumulate the propagation contribution and the path weight factor to obtain the cumulative contribution of the path intersection node. A distribution field of the cumulative contribution of path intersection nodes on the power grid topology is constructed. The distribution field is decomposed into a gradient field to obtain a gradient vector field of cumulative contribution. Local maxima of the gradient vector magnitude are identified in the gradient vector field of cumulative contribution. The path intersection nodes corresponding to the local maxima are taken as the core node set.

2. The method according to claim 1, characterized in that, Collect time-series data of multidimensional physical quantities during transient disturbances in the power grid, extract the transient response intensity of the time-series data of multidimensional physical quantities, and obtain the topological connectivity of the power grid topology, including: Multi-scale time-frequency decomposition is performed on the time series data of multidimensional physical quantities to obtain transient components in different frequency bands. The energy concentration degree and frequency drift trajectory of the transient components within the transient window are extracted. The transient response intensity is calculated based on the coupling strength of the energy concentration degree and frequency drift trajectory. Read the physical connection relationship between power grid nodes, construct a weighted topology graph that reflects electrical distance and impedance transmission characteristics, calculate the effective conduction path and branch redundancy between node pairs in the weighted topology graph, and generate a dynamic propagation factor matrix as topology connectivity; The transient response intensity is used as the excitation source input, and the dynamic propagation factor matrix is ​​used as the propagation medium parameter. The node state value is obtained by iterative calculation through the diffusion equation. The numerical evolution of node state values ​​at continuous time points is tracked, the set of nodes whose state values ​​exceed the perturbation threshold is extracted, the activation time of each node in the set is recorded, and the set of nodes is connected in chronological order of the activation times to form a diffusion trajectory.

3. The method according to claim 1, characterized in that, Extracting the morphological features of the diffusion trajectory and mapping these features to fault type identifiers, the determination of the fault source location and fault influence boundary based on the diffusion trajectory includes: The activation time and topological position of each node in the diffusion trajectory are constructed into a set of spatiotemporal coordinate points. The spatiotemporal coordinate points are morphologically clustered to obtain multiple diffusion clusters. The spatial compactness and temporal extension of each diffusion cluster are extracted, and the cross ratio of spatial compactness and temporal extension is used as a morphological feature. Establish interval division rules for the cross ratio of spatial density and temporal extension. When the cross ratio shows a spatial density-dominant pattern, it is identified as a localized concentrated fault. When the cross ratio shows a temporal extension-dominant pattern, it is identified as a chain propagation fault. When the cross ratio shows a balanced distribution pattern, it is identified as an oscillation-spreading fault. The fault type identifier is determined based on the dominant pattern of the cross ratio. In each diffusion cluster, identify the node with the earliest activation time, calculate the temporal difference between each node and the other nodes in its diffusion cluster, and select the node with the largest temporal difference as the source of the fault. Simultaneously, the decay gradient distribution of node state values ​​in the diffusion trajectory in the topological space is extracted, and nodes where the decay gradient distribution changes abruptly are connected to form the fault impact boundary.

4. The method according to claim 1, characterized in that, The temporal gradient of node state values ​​is calculated, and the propagation path is obtained by tracing back along the temporal gradient from the fault source location to the affected nodes within the fault influence boundary. The propagation contribution of the propagation path is calculated based on topological connectivity and temporal gradient, including: An evolution surface of the state values ​​of each node in the diffusion trajectory on the time axis is constructed. The evolution surface is decomposed locally into a linearized tangent plane to obtain the temporal change plane. The normal vector of the temporal change tangent plane is extracted as the temporal gradient. Establish a vector field topology structure for temporal gradients, identify the source singular position corresponding to the fault source position and the sink singular position corresponding to each node within the fault influence boundary in the vector field topology structure, and extract the streamline trajectory connecting the source singular position and each sink singular position as the propagation path. Calculate the topological betweenness and the divergence of the temporal gradient at each node along the propagation path. Use the nonlinear coupling value of the topological betweenness and the divergence as the node propagation potential energy. Integrate the node propagation potential energy along the propagation path to obtain the propagation contribution of the propagation path.

5. The method according to claim 4, characterized in that, A vector field topology structure for temporal gradients is established. Within this structure, the singular positions of the source points corresponding to the fault origin and the singular positions of the sink points corresponding to each node within the fault influence boundary are identified. The streamline trajectories connecting the source point singular positions and each sink point singular position are extracted as propagation paths, including: The distribution of temporal gradients in the power grid topology space is constructed as a vector field. The vector field is decomposed topologically to obtain the set of singular points and the manifold structure. The global topological invariants of the vector field are calculated based on the topological index of the set of singular points. The global topological invariants are mapped to the topology type of the vector field topology structure. In the vector field topology, singular points with positive topological exponents are extracted as candidate source points. The topological distance between each candidate source point and the location of the fault source is calculated. The candidate source point with the smallest topological distance is selected as the singular location of the source point. In the vector field topology, singular points with negative topological indices are extracted as candidate sinks. It is then determined whether each candidate sink is located within the fault influence boundary. Candidate sinks located within the fault influence boundary are designated as singular sink locations. In the vector field topology, a family of streamlines is constructed with the singular position of the source point as the initial point. Each streamline in the family of streamlines is topologically extended to the singular position of the sink point. The streamlines extended from the singular position of the source point to the singular position of each sink point are extracted to form a set of streamline trajectories. Calculate the trajectory deviation of each streamline in the streamline trajectory set under topological perturbation, and take the streamline with trajectory deviation lower than the median value of trajectory deviation in the streamline trajectory set as the propagation path.

6. The method according to claim 1, characterized in that, Based on the fault type identifier, configure isolation trigger conditions for the core node set. When the isolation trigger conditions are detected, disconnect the topology connection of the corresponding node in the core node set and output the network segmentation scheme, including: Determine the corresponding fault propagation direction based on the fault type identifier, trace the downstream propagation links of each node in the core node set along the fault propagation direction, and identify the load node density on each downstream propagation link. Mark the core nodes of downstream propagation links whose load node density exceeds the average load node density in the core node set as priority isolation nodes, and configure isolation trigger conditions for priority isolation nodes. Real-time monitoring of the fault current flow direction of each node in the core assembly, determining whether the fault current flow direction is consistent with the fault propagation direction, and extracting the flow rate change rate of the fault current when the fault current flow direction is consistent with the fault propagation direction, and matching the flow rate change rate with the isolation trigger condition to determine whether to trigger the isolation operation. When an isolation operation is triggered, the upstream and downstream nodes of the priority isolation node in the power grid topology are located along the fault propagation direction. The topological connection between the priority isolation node and the upstream node is disconnected to block the fault propagation path. A network segmentation scheme is generated based on the disconnected topology.

7. An artificial intelligence-based power grid transient fault mode classification system, used to implement the method of any one of claims 1-6, characterized in that, include: The first unit is used to collect time-series data of multidimensional physical quantities during transient disturbances of the power grid, extract the transient response intensity of the time-series data of multidimensional physical quantities, obtain the topological connectivity of the power grid topology, couple the transient response intensity with the topological connectivity to obtain the node state value, and generate the diffusion trajectory based on the time evolution of the node state value. The second unit is used to extract the morphological features of the diffusion trajectory, map the morphological features to the fault type identifier, and determine the location of the fault source and the boundary of the fault impact based on the diffusion trajectory. The third unit is used to calculate the temporal gradient of the node state value. It traces back from the fault source location along the temporal gradient to the affected nodes within the fault influence boundary to obtain the propagation path. Based on topological connectivity and temporal gradient, it calculates the propagation contribution of the propagation path. The fourth unit is used to identify the path intersection nodes of the propagation path in the power grid topology, accumulate the propagation contribution of each path intersection node to obtain the cumulative contribution, and determine the core node set based on the cumulative contribution. The fifth unit is used to configure isolation trigger conditions for the core node set according to the fault type identifier. When the isolation trigger conditions are detected, the topology connection of the corresponding node in the core node set is disconnected and the network segmentation scheme is output.

Images