Power grid cascading failure propagation path identification method and system based on graph neural network

Abstract

This application relates to the field of data processing technology and discloses a method and system for identifying the propagation path of cascading faults in power grids based on graph neural networks. The method includes: mapping buses, lines, and transformers as graph nodes and connecting edges to construct a power grid topology graph carrying real-time power flow data, protection action signals, and switch status; identifying fault-immune nodes using dual criteria and removing them, then redistributing power flow loads to residual connecting edges according to capacity margin ratios to obtain a contracted subgraph; inputting the contracted subgraph into a graph neural network for multi-layer neighborhood aggregation, outputting risk scores for each node and filtering the set of active leading-edge nodes; extracting the differences between nodes at adjacent times as propagation increments according to time sequence, reconstructing the cascading fault propagation trajectory, and outputting the fault propagation path and equipment risk rating. This application improves the accuracy of fault propagation path identification and equipment risk rating in complex cascading fault scenarios.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method and system for identifying the propagation path of cascading faults in power grids based on graph neural networks. Background Technology

[0002] A cascading failure in a power grid refers to a chain reaction process in which a failure in one piece of equipment triggers a successive tripping of adjacent equipment due to power flow shifts, voltage fluctuations, or protection malfunctions, leading to a continuous expansion of the fault's scope. Existing methods for identifying cascading failure propagation paths mainly fall into two categories: one is based on probabilistic models, statistically inferring propagation paths by establishing a device failure probability matrix; the other is based on graph theory methods, using the power grid topology as a foundation and employing shortest path algorithms or maximum flow algorithms to identify fault propagation links. Both methods take a static power grid topology as input, perform a one-time inference of the complete graph structure after the fault occurs, and output a fixed propagation path result.

[0003] However, the above methods have significant shortcomings in practical applications. First, the power grid topology continuously changes structurally during cascading faults due to circuit breaker actions. A one-time inference based on a static graph cannot reflect the dynamic contraction of the topology during fault evolution, causing physically isolated devices to still participate in subsequent inferences, leading to misjudgments of fault propagation paths. Second, existing methods rely solely on a single protection action signal to determine whether a device has exited the propagation link. However, in actual engineering, protection action signals suffer from issues such as false transmission and delayed uploading. A single criterion cannot effectively distinguish between actual tripping isolation and signal anomalies, thus affecting the accuracy of topology updates. Third, static graph methods fail to correlate edge weights with the physical process of power flow transfer during node feature extraction. The carrying state of residual connecting edges after removing faulty nodes cannot be updated in real time, causing a deviation between the graph structure and the actual operating state of the power grid. Summary of the Invention

[0004] This application provides a method and system for identifying propagation paths of cascading faults in power grids based on graph neural networks. It solves the problems of inaccurate propagation path identification caused by static graph-based cascading fault propagation path identification methods in the prior art, which suffer from the inability to dynamically update the topology, the single node stripping criterion, and the failure to synchronously map power flow transfer to edge weights. This improves the accuracy of fault propagation path identification and equipment risk rating in complex cascading fault scenarios.

[0005] Firstly, this application provides a method for identifying the propagation path of cascading faults in a power grid based on a graph neural network, the method comprising: Step S1: Map the busbars, lines, and transformers to graph nodes and connecting edges according to their electrical coupling relationships. Use the real-time power flow data, protection action signals, and switch status of each node as node feature vectors to construct the power grid topology graph. Step S2: Using the tripping triggered by the protection action signal and the opening confirmed by the switch status as dual criteria, mark the fault-immune node from the power grid topology diagram, remove the fault-immune node from the power grid topology diagram, and redistribute the power flow load it carries to the edge weight of the residual connection edge according to the capacity margin ratio of each neighboring node to obtain the shrinking subgraph. Step S3: Input the contracted subgraph into the graph neural network, using the node feature vector as the initial input, and output the risk score of each node after multi-layer neighborhood aggregation. The active frontier node set is obtained by filtering the risk scores. Step S4: Arrange the contracted subgraphs of each time step in chronological order, use the node difference of the active frontier node set at adjacent times as the propagation increment, reconstruct the cascading fault propagation trajectory, take the node with the highest risk score in the first time step as the fault origin point, and output the fault propagation path and equipment risk rating.

[0006] Secondly, this application provides a power grid cascading fault propagation path identification system based on graph neural networks, the power grid cascading fault propagation path identification system based on graph neural networks comprising: The module is used to map the busbars, lines, and transformers into graph nodes and connecting edges according to their electrical coupling relationships. The real-time power flow data, protection action signals, and switch status of each node are used as node feature vectors to construct the power grid topology graph. The calibration module is used to use the tripping triggered by the protection action signal and the opening confirmed by the switch status as dual criteria to calibrate the fault-immune node from the power grid topology diagram, remove the fault-immune node from the power grid topology diagram, and redistribute the power flow load it carries to the edge weight of the residual connection edge according to the capacity margin ratio of each neighboring node to obtain a contracted subgraph. The filtering module is used to input the shrinking subgraph into the graph neural network, take the node feature vector as the initial input, and output the risk score of each node after multi-layer neighborhood aggregation. The active frontier node set is obtained by filtering the risk scores. The output module is used to arrange the contracted subgraphs of each time step in chronological order, use the node difference of the active frontier node set at adjacent time points as the propagation increment, reconstruct the cascading fault propagation trajectory, and take the node with the highest risk score in the first time step as the fault origin point to output the fault propagation path and equipment risk rating.

[0007] In the technical solution provided in this application, a power grid topology diagram is constructed based on the electrical coupling relationship between buses, lines, and transformers. Real-time power flow data, protection action signals, and switch status are uniformly encoded into node feature vectors, enabling the graph structure to simultaneously carry two types of information: the power grid topology connection relationship and the real-time operating status of each device. This provides a dual-dimensional input foundation of structure and state for subsequent neighborhood aggregation of the graph neural network. Furthermore, fault-immune nodes are calibrated using a dual criterion of tripping triggered by protection action signals and opening confirmed by switch status. Compared to a single signal criterion, the dual criterion, through the collaborative verification of timing rationality and signal consistency, limits the impact of engineering interference factors such as signal mistransmission and delayed transmission to the node calibration stage, preventing miscalibrated nodes from entering the subsequent stripping process. After the fault-immune node is stripped, its power flow load is redistributed to the edge weights of the remaining connection edges according to the capacity margin ratio of each neighboring node. This ensures that the edge weight distribution of the contracted subgraph maintains physical consistency with the actual power flow transfer process of the power grid, and each contraction of the graph structure corresponds to the actual topology evolution state of the power grid.

[0008] The graph neural network drives multi-layer neighborhood aggregation on the shrinking subgraph by updating edge weights. This gives higher weight to neighboring nodes with tight electrical coupling and large power flow transfers in the feature aggregation. The aggregation result directly reflects the propagation risk level of each node under the current topology shrinkage state. The risk score output after scoring mapping layer and Sigmoid normalization has a clear physical correspondence. Furthermore, a graph shrinkage rate-triggered node feature vector normalization update mechanism is introduced to promptly correct the numerical range of feature vectors when large-scale node stripping causes power flow data distribution shift, ensuring that the feature input of the graph neural network is always within the effective numerical range under different shrinkage degrees. Finally, the propagation increment is extracted by the difference set of active leading node sets at each time step, and the risk level of the equipment is comprehensively evaluated by the product of the cumulative risk score and the occurrence frequency. The dynamic evolution process of cascading failures is restored to a propagation path with a temporal structure, and the continuous risk level of each node in the path is quantitatively ranked, providing dispatchers with a complete propagation link and risk classification basis from the fault origin point to each level of diffusion node. Attached Figure Description

[0009] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a schematic diagram of an embodiment of the power grid cascading fault propagation path identification method based on graph neural networks in this application. Figure 2This is a schematic diagram comparing the error rate of fault-immune nodes with signal noise level under dual criteria and single protection signal criteria in the embodiments of this application. Detailed Implementation

[0011] This application provides a method and system for identifying the propagation path of cascading faults in a power grid based on a graph neural network. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data used can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or device that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0012] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of the power grid cascading fault propagation path identification method based on graph neural networks in this application includes: Step S1: Map the busbars, lines, and transformers to graph nodes and connecting edges according to their electrical coupling relationships. Use the real-time power flow data, protection action signals, and switch status of each node as node feature vectors to construct the power grid topology graph. The power grid topology is constructed using three types of primary equipment—busbars, lines, and transformers—as basic units. Busbars handle the electrical convergence of equipment at the same voltage level, lines are responsible for power transmission between nodes, and transformers convert electrical energy between different voltage levels. After determining the node set and connection edge set based on the actual electrical coupling relationships of these three types of equipment, each connection edge is assigned an initial edge weight based on the line impedance value between its two endpoints. The smaller the impedance, the larger the edge weight, reflecting the tightness of the electrical coupling. The node feature vector of each node is constructed by concatenating three types of data in the order of node number: real-time power flow data reflects the direction and magnitude of active and reactive power flow at the node at the current moment; protection action signals record the action status of the corresponding protection device at the node; and switch status records the open and closed positions of circuit breakers and disconnectors. These three types of data together describe the operating status of the node within the current sampling period.

[0013] Step S2: Using the tripping triggered by the protection action signal and the opening of the switch status as dual criteria, mark the fault-immune nodes from the power grid topology diagram, remove the fault-immune nodes from the power grid topology diagram, and redistribute the power flow load they carry to the edge weights of the residual connection edges according to the capacity margin ratio of each neighboring node to obtain the shrinking subgraph. Specifically, the identification of fault-immune nodes relies on the collaborative verification of dual criteria. A single protection action signal may be falsely transmitted or delayed in transmission; judging a node's exit from the propagation link solely based on the protection action signal will lead to misidentification. Therefore, a switching status is introduced as a second criterion, requiring both the protection action signal triggering a trip and the switching status confirming the opening of the circuit breaker to be simultaneously satisfied before a node can be identified as a fault-immune node. The redistribution of power flow is based on the capacity margin ratio of each neighboring node. The capacity margin is the difference between the current rated capacity and the actual load carried by each neighboring node. A larger margin indicates a stronger ability of the neighboring node to absorb transferred power flow, resulting in a larger increment in the edge weight allocated to the corresponding residual connection edge. This ensures that the edge weight distribution of the contracted subgraph truly reflects the physical process of power flow transfer after the fault-immune node is removed.

[0014] Step S3: Input the contracted subgraph into the graph neural network, using the node feature vector as the initial input. After multi-layer neighborhood aggregation, the risk score of each node is output, and the active frontier node set is obtained by filtering the risk scores. In this method, the multi-layer neighborhood aggregation of the graph neural network uses the node feature vectors of each node in the contracted subgraph as the input to layer 0. During each aggregation layer, each node sums the hidden vectors of all its neighboring nodes from the previous layer according to the edge weights of the corresponding connecting edges. This sum is then concatenated with its own hidden vector from the previous layer and processed by a linear transformation and activation function to obtain the hidden vector for that layer. After L layers of aggregation, the hidden vectors of each node have incorporated the state information of all nodes within its L-hop neighborhood. The difference in edge weights means that electrically tightly coupled neighboring nodes contribute more to the aggregation result. Finally, the L-layer hidden vectors of each node are linearly transformed by the scoring mapping layer and normalized by Sigmoid to output a risk score. The score ranges from 0 to 1, reflecting the likelihood that the node will participate in the propagation of a cascading failure at the current time step.

[0015] Step S4: Arrange the contracted subgraphs of each time step in chronological order, use the difference between the active leading edge nodes in adjacent time steps as the propagation increment, reconstruct the cascading fault propagation trajectory, take the node with the highest risk score in the first time step as the fault origin point, and output the fault propagation path and equipment risk rating.

[0016] Specifically, the propagation increment is extracted based on the difference set of the active frontier nodes in adjacent time steps. The newly added nodes in the difference set are the devices that have newly entered the propagation frontier in the current time step. The newly added nodes in each time step are connected sequentially to form a cascading failure propagation trajectory. The failure origin point is the node with the highest risk score in the first time step. This node exhibits the highest propagation risk in the initial stage of the cascading failure, which is consistent with the physical meaning of the failure's initial location. The device risk rating is a combination of the cumulative risk score of each node across all time steps and its frequency of appearance in the active frontier node set. Nodes with high cumulative scores and high frequency of appearance remain at the propagation front throughout the cascading failure process, and their device risk level is the highest.

[0017] In one specific embodiment, step S1 includes: Using the busbar as the voltage aggregation node, the line as the power transmission node, and the transformer as the voltage transformation node, a set of nodes and a set of connecting edges are established according to the electrical coupling relationship between each device to obtain the initial topology framework; Based on the impedance values ​​between the nodes at both ends of each connecting edge in the initial topology framework, each connecting edge is assigned an initial edge weight to obtain a weighted topology graph. The power grid topology is obtained by concatenating the real-time power flow data, protection action signals, and switch status of each node in the current sampling period according to the node number order, constructing the node feature vector for each node in the weighted topology diagram.

[0018] Specifically, the initial topology framework is established based on the primary wiring diagram of the power grid. Buses, as voltage convergence nodes, handle the electrical connections of various devices at the same voltage level. Lines, as power transmission nodes, record the power flow transmission relationships between nodes. Transformers, as voltage transformation nodes, connect buses of different voltage levels. The three types of devices determine the node set and connection edge set according to their actual electrical coupling relationships. The node set includes all graph nodes corresponding to buses, lines, and transformers in the power grid, while the connection edge set includes edges with direct electrical connections between devices. Together, they constitute the initial topology framework. Based on this, each connection edge is assigned an initial edge weight according to the line impedance value between the nodes at both ends. The impedance value is inversely proportional to the edge weight; the smaller the impedance, the tighter the electrical coupling between the nodes at both ends, and the larger the corresponding edge weight, resulting in a weighted topology graph.

[0019] The node feature vector is constructed from three types of data for each node within the current sampling period: real-time power flow data, which includes the current measured values ​​of active and reactive power of the corresponding equipment, reflecting the real-time load status of the node; protection action signals, which record whether the corresponding protection device of the node has acted and the type of action within the current sampling period; and switch status, which records the current open / closed position of the corresponding circuit breaker and disconnector. These three types of data are sequentially concatenated into a one-dimensional vector according to node number, serving as the node feature vector for that node. The node feature vectors of all nodes, together with the node set, connection edge set, and initial edge weights of the weighted topology graph, constitute the power grid topology graph. The power grid topology graph fully describes the current topology of the power grid and the operating status of each node.

[0020] In one specific embodiment, step S2 uses the tripping triggered by the protection action signal and the opening of the circuit breaker confirmed by the switch status as dual criteria, including: The first criterion is whether the time difference between the triggering time of the protection action signal of each node and the opening confirmation time of the switch status falls within a preset time window. Nodes whose time difference falls within the preset time window are marked as candidate immune nodes. Based on the consistency verification results of the circuit breaker position signal and remote signaling confirmation signal corresponding to each candidate immune node as the second criterion, the candidate immune nodes that pass the consistency verification are finally confirmed as fault immune nodes.

[0021] Specifically, the preset time window is determined based on the physical response time range from the protection device's action to the circuit breaker's completion of tripping. The time difference between the triggering moment of the protection action signal and the tripping confirmation moment of the switching quantity reflects the time interval from the protection device issuing a tripping command to the circuit breaker actually completing the tripping action. Under normal tripping conditions, this time difference falls within a reasonable physical response time range. If the time difference exceeds the upper limit of the preset time window, it indicates that the circuit breaker has an abnormal situation of failure to operate or delayed operation, and the corresponding node should not be marked as a candidate immune node. If the time difference is negative or extremely small, it indicates that the change in the switching quantity state precedes the transmission of the protection action signal, indicating an abnormal signal transmission sequence, which also does not meet the first criterion. Only when the time difference falls within the preset time window can the node be marked as a candidate immune node.

[0022] The circuit breaker position signal originates from the auxiliary contacts of the circuit breaker body, directly reflecting the actual open / closed state of the circuit breaker's mechanical contacts. The remote signaling confirmation signal originates from the remote acquisition results of the circuit breaker position by the substation monitoring system. Both should indicate an open state for the consistency check to pass. If the circuit breaker position signal shows open while the remote signaling confirmation signal shows closed, or if the two states are opposite, it indicates an abnormal signal acquisition or communication transmission error, and the corresponding candidate immune node will not be confirmed as a fault-immune node. The fault-immune nodes finally confirmed after dual-criteria collaborative verification have their protection actions and circuit breaker open state both checked for timing rationality and signal consistency to ensure that the calibration results are consistent with the actual physical isolation state of the power grid.

[0023] Figure 2 This is a schematic diagram comparing the error rate of fault-immune nodes with signal noise levels under dual criteria and single protection signal criteria in embodiments of this application. Figure 2 As shown, the horizontal axis represents the signal noise level, and the vertical axis represents the false labeling rate of the fault-immune node. The solid line corresponds to the false labeling rate change curve using only a single protection action signal criterion, while the dashed line corresponds to the false labeling rate change curve using dual criteria in this application. When the signal noise level is 0, the false labeling rate of the single criterion is 0.02, and the false labeling rate of the dual criterion is 0.01, with little difference between the two. As the signal noise level increases to 0.15, the false labeling rate of the single criterion rises to 0.43, while the false labeling rate of the dual criterion only rises to 0.10. This indicates that under the condition of noise interference in the protection action signal, the dual verification mechanism, which uses the time difference between the protection action signal triggering time and the switch status opening confirmation time falling into a preset time window as the first criterion and the consistency verification between the circuit breaker position signal and the remote signaling confirmation signal passing as the second criterion, has a significant suppressive effect on the false labeling rate of the fault-immune node in high-noise scenarios compared to the single protection signal criterion.

[0024] In one specific embodiment, step S2, removing the fault-immune node from the power grid topology diagram, includes: Remove the connection edges between each fault-immune node and all its neighboring nodes from the power grid topology graph to obtain the stripped edge set; Based on the power flow load value carried by each fault-immune node before stripping, the power flow load value is allocated to the edge weight of the corresponding residual connection edge in the stripping edge set according to the proportion of the capacity margin of each neighboring node connected to it to the total capacity margin of all neighboring nodes, so as to obtain the updated edge weight of each residual connection edge. The original edge weights of the corresponding connecting edges in the power grid topology graph are replaced with updated edge weights to obtain a contracted subgraph.

[0025] Specifically, the stripping edge set is obtained by taking fault-immune nodes as the operation object. All connection edges between each fault-immune node and all its neighboring nodes are removed from the power grid topology graph. The set of connection edges after removal is the stripping edge set. The fault-immune nodes themselves are simultaneously removed from the node set. The remaining node set and the remaining connection edge set together form the shrunk graph structure skeleton. The power flow load value carried by each fault-immune node before stripping is taken from the real-time power flow data component in its node feature vector. This value reflects the amount of active power actually carried by the node at the time of stripping. The capacity margin of each neighboring node is the difference between its rated capacity and its current actual load. The sum of the capacity margins is the sum of the capacity margins of all neighboring nodes connected to the fault-immune node. The proportion of the capacity margin of each neighboring node to the sum of the capacity margins is the power flow load weight proportion allocated to the residual connection edge of that neighboring node.

[0026] The calculation of updated edge weights is based on the original edge weights of each residual connection edge. The proportionally allocated power flow load increment is superimposed on the original edge weight of the corresponding residual connection edge to obtain the updated edge weight. The original edge weight is determined by the impedance value between the nodes at both ends of the connection edge. The updated edge weight is based on the original edge weight with the load increment brought about by the power flow transfer, so that the edge weight of the residual connection edge reflects both the degree of electrical coupling and the actual power flow carrying capacity change. After replacing the original edge weight of the corresponding connection edge in the power grid topology with the updated edge weight, the edge weight distribution of the contracted subgraph corresponds to the physical state of the actual power flow redistribution of the power grid after the fault-immune node is removed. The larger the capacity margin of the neighboring node, the more power flow transfer its corresponding residual connection edge undertakes, and the larger the updated edge weight.

[0027] In one specific embodiment, step S2, obtaining the contracted subgraph, further includes: The ratio of the number of fault-immune nodes stripped at each time step to the total number of nodes in the power grid topology graph is used as the graph shrinkage rate. When the graph shrinkage rate exceeds the preset shrinkage threshold, the node feature vectors of each residual node in the shrinkage subgraph are renormalized according to the current real-time power flow data of each node to obtain the updated shrinkage subgraph. When the graph shrinkage rate does not exceed the preset shrinkage threshold, the original node feature vectors of each residual node in the shrunken subgraph are retained, and the shrunken subgraph is directly output.

[0028] Specifically, the graph shrinkage rate is calculated as the ratio of the number of fault-immune nodes stripped at each time step to the total number of nodes in the initial power grid topology graph, reflecting the degree of shrinkage of the power grid topology graph relative to the initial state at the current time step. When the graph shrinkage rate does not exceed the preset shrinkage threshold, the number of fault-immune nodes stripped accounts for a small proportion of the initial total number of nodes, the real-time power flow data of the remaining nodes changes only slightly relative to the initial state, and the power flow data components in the feature vectors of each remaining node are still within the initial normalization range. Retaining the original node feature vectors will not cause significant deviations in the numerical dimensions of the feature vectors between nodes, and the shrunken subgraph is directly output.

[0029] When the graph shrinkage rate exceeds a preset shrinkage threshold, a large number of fault-immune nodes have been removed from the power grid topology. The residual nodes, due to receiving transferred power flows, experience significant changes in real-time power flow data relative to their initial state. The power flow data components in the original node feature vectors can no longer accurately reflect the actual operating status of the remaining nodes, and the numerical ranges of power flow data among different residual nodes may vary considerably. Directly inputting the original node feature vectors into the graph neural network would lead to inconsistencies in the numerical dimensions of the feature vectors of each node. Therefore, the power flow data components in the node feature vectors are re-normalized using the current real-time power flow data of each residual node. Normalization is performed using linear scaling with the maximum and minimum values ​​of the real-time power flow data of all residual nodes at the current time step as upper and lower bounds, restoring the node feature vectors of each residual node to a uniform numerical range, resulting in an updated shrunken subgraph.

[0030] In one specific embodiment, step S3 involves inputting the shrunken subgraph into a graph neural network, using the node feature vector as the initial input, and outputting the risk score for each node after multi-layer neighborhood aggregation, including: The node feature vectors of each node in the shrinking subgraph are used as the input of the 0th layer. When aggregating at the kth layer, the hidden vectors of all the neighboring nodes of each node at the (k-1)th layer are weighted and summed according to the update edge weights of the corresponding connection edges. This summation is then concatenated with the hidden vector of the corresponding node itself at the (k-1)th layer and processed by a linear transformation and activation function to obtain the hidden vector of the node at the kth layer. After L-layer neighborhood aggregation, the hidden layer vector of each node in the L-layer is input into the scoring mapping layer. After linear transformation of the hidden layer vector of each node in the L-layer, it is normalized by the Sigmoid function to obtain the risk score of each node at the current time step. Nodes with risk scores exceeding a preset risk threshold are selected as candidate frontier nodes. Nodes that have been marked as fault-immune nodes are removed from all candidate frontier nodes, and the remaining candidate frontier nodes form the active frontier node set.

[0031] Specifically, the multi-layer neighborhood aggregation of the graph neural network uses the node feature vectors of each node in the contracted subgraph as the input of layer 0. During aggregation at layer k, the hidden vector of each node is composed of two parts: the first part is the weighted sum of the hidden vectors of all its neighbors at layer k-1, calculated according to the update edge weights of their corresponding connections. Neighbors with larger update edge weights contribute more to the weighted sum, giving higher weights to electrically coupled neighbors that handle large amounts of power flow transfer. The second part is the node's own hidden vector at layer k-1, preserving its historical state information. After concatenation, the two parts are mapped to a unified dimension using a linear transformation matrix, and then a nonlinear transformation is introduced through an activation function to obtain the hidden vector of the node at layer k. As the aggregation layer k increases, the hidden vector of each node gradually integrates the state information of all nodes within its k-hop neighborhood, allowing the spatial diffusion characteristics of fault propagation in the power grid topology to be accumulated layer by layer in the hidden vector.

[0032] After L-layer neighborhood aggregation, the hidden vector of each node in the L-layer has been integrated with the topological structure and operational status information of all nodes within its L-hop neighborhood. The scoring mapping layer applies a linear transformation to the hidden vector of each node in the L-layer, compressing the high-dimensional hidden vector into a scalar output. Then, the scalar is mapped to a continuous interval of 0 to 1 by the Sigmoid function, obtaining the risk score of each node at the current time step. The closer the output value of the Sigmoid function is to 1, the closer its neighborhood state is to the feature pattern of the fault propagation front node under the current topological structure and node feature vector distribution of the shrunken subgraph; the closer the output value is to 0, the greater the difference between the current state of the node and the feature pattern of the fault propagation front.

[0033] A preset risk threshold serves as the screening boundary for risk scores. Nodes with risk scores exceeding the preset risk threshold are included in the candidate frontier node set. The candidate frontier node set may include nodes that have been marked as fault-immune nodes in previous time steps. Fault-immune nodes have completed physical isolation and exited the propagation chain. Their high risk scores are due to the residual influence of their historical neighborhood characteristics rather than actual propagation threats. Therefore, the marked fault-immune nodes need to be removed one by one from the candidate frontier node set. The remaining candidate frontier nodes constitute the active frontier node set for the current time step. The nodes in the active frontier node set are all nodes that are still under the threat of chain fault propagation at the current moment and have not yet completed physical isolation.

[0034] In one specific embodiment, step S4 involves arranging the contracted subgraphs of each time step in chronological order, using the node differences of the active leading-edge node sets at adjacent times as the propagation increment, and reconstructing the cascading failure propagation trajectory, including: The difference between the active leading edge node set at time step t and time step (t-1) is used as the propagation increment node set at time step t. The propagation increment node sets at each time step are connected end to end in time sequence to obtain the chain fault propagation trajectory. Taking the node with the highest risk score in the first time step as the fault origin point, the fault propagation path is obtained by arranging the risk scores of each node in the incremental node set of each time step along the chain fault propagation trajectory from high to low. Based on the product of the cumulative risk score of each node across all time steps and the frequency of its appearance in the set of active frontier nodes, a comprehensive rating is given to each node in the fault propagation path to obtain the equipment risk rating.

[0035] Specifically, the propagation increment node set is calculated using the difference between the active frontier node set at time step t and time step (t-1). The nodes in the difference set are those newly entering the active frontier at time step t but not yet appearing in the active frontier node set at time step (t-1), reflecting the range of equipment newly affected by the cascading fault at time step t. The propagation increment node sets of each time step are then connected sequentially, starting with the propagation increment node set of time step 1, and successively concatenating the propagation increment node sets of time steps 2, 3, up to T, to obtain the complete cascading fault propagation trajectory. The fault propagation path further incorporates node sorting based on the cascading fault propagation trajectory. The node with the highest risk score in the first time step is taken as the fault origin point. Nodes within the propagation increment node set of each time step are arranged from highest to lowest risk score along the propagation trajectory, ensuring that multiple newly added frontier nodes within the same time step are arranged sequentially according to their propagation risk level. This yields a fault propagation path starting from the fault origin point and extending along the direction of decreasing risk score in each time step.

[0036] Equipment risk rating is based on the product of the cumulative risk score of each node across all time steps and the frequency of its appearance in the set of active frontier nodes. The cumulative risk score is the summation of the risk scores for each time step from time 1 to time T; a higher cumulative score indicates that the node remains in a high-risk state throughout the cascading failure process. The frequency of appearance is the number of times the node is included in the set of active frontier nodes across all time steps; a higher frequency indicates that the node remains at the forefront of propagation for more time steps during the cascading failure propagation. The product of these two scores reflects the relationship between the degree of risk and the duration of risk for the node throughout the cascading failure process; a larger product value indicates a higher equipment risk rating, and a smaller product value indicates a lower equipment risk rating.

[0037] The above describes the power grid cascading fault propagation path identification method based on graph neural networks in the embodiments of this application. The following describes the power grid cascading fault propagation path identification system based on graph neural networks in the embodiments of this application. One embodiment of the power grid cascading fault propagation path identification system based on graph neural networks in the embodiments of this application includes: The module is used to map the busbars, lines, and transformers into graph nodes and connecting edges according to their electrical coupling relationships. The real-time power flow data, protection action signals, and switch status of each node are used as node feature vectors to construct the power grid topology graph. The calibration module is used to calibrate fault-immune nodes from the power grid topology diagram using the tripping triggered by the protection action signal and the opening of the switch status as dual criteria. The fault-immune nodes are then removed from the power grid topology diagram, and their power flow loads are redistributed to the edge weights of the residual connection edges according to the capacity margin ratio of each neighboring node, resulting in a contracted subgraph. The filtering module is used to input the shrinking subgraph into the graph neural network, take the node feature vector as the initial input, and output the risk score of each node after multi-layer neighborhood aggregation. The active frontier node set is obtained by filtering the risk scores. The output module is used to arrange the contracted subgraphs of each time step in chronological order, use the node difference of the active frontier node set at adjacent time points as the propagation increment, reconstruct the cascading fault propagation trajectory, and take the node with the highest risk score in the first time step as the fault origin point to output the fault propagation path and equipment risk rating.

[0038] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to 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 of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying a cascading failure propagation path of a power grid based on a graph neural network, characterized in that, The method includes: Step S1: Map the busbars, lines, and transformers to graph nodes and connecting edges according to their electrical coupling relationships. Use the real-time power flow data, protection action signals, and switch status of each node as node feature vectors to construct the power grid topology graph. Step S2: Using the tripping triggered by the protection action signal and the opening confirmed by the switch status as dual criteria, mark the fault-immune node from the power grid topology diagram, remove the fault-immune node from the power grid topology diagram, and redistribute the power flow load it carries to the edge weight of the residual connection edge according to the capacity margin ratio of each neighboring node to obtain the shrinking subgraph. Step S3: Input the contracted subgraph into the graph neural network, using the node feature vector as the initial input, and output the risk score of each node after multi-layer neighborhood aggregation. The active frontier node set is obtained by filtering the risk scores. Step S4: Arrange the contracted subgraphs of each time step in chronological order, use the node difference of the active frontier node set at adjacent times as the propagation increment, reconstruct the cascading fault propagation trajectory, take the node with the highest risk score in the first time step as the fault origin point, and output the fault propagation path and equipment risk rating.

2. The method of claim 1, wherein, Step S1 includes: Using the busbar as the voltage aggregation node, the line as the power transmission node, and the transformer as the voltage transformation node, a set of nodes and a set of connecting edges are established according to the electrical coupling relationship between each device to obtain the initial topology framework; Based on the impedance values ​​between the nodes at both ends of each connecting edge in the initial topology framework, each connecting edge is assigned an initial edge weight to obtain a weighted topology graph. By concatenating the real-time power flow data, protection action signals, and switch status of each node in the current sampling period according to the node number sequence, a node feature vector is constructed for each node in the weighted topology diagram to obtain the power grid topology diagram.

3. The method of claim 1, wherein, In step S2, the tripping triggered by the protection action signal and the opening confirmed by the switch status are used as dual criteria, including: The first criterion is whether the time difference between the triggering time of the protection action signal of each node and the opening confirmation time of the switch state falls within a preset time window. Nodes whose time difference falls within the preset time window are marked as candidate immune nodes. Based on the consistency verification result of the circuit breaker position signal and the remote signaling confirmation signal corresponding to each candidate immune node as the second criterion, the candidate immune node that passes the consistency verification is finally confirmed as the fault immune node.

4. The method of claim 3, wherein, In step S2, removing the fault-immune node from the power grid topology includes: Remove the connection edges between each fault-immune node and all its neighboring nodes from the power grid topology graph to obtain a stripped edge set; Based on the power flow load value carried by each fault-immune node before stripping, the power flow load value is allocated to the edge weight of the corresponding residual connection edge in the stripping edge set according to the proportion of the capacity margin of each neighboring node connected to it to the total capacity margin of all neighboring nodes, so as to obtain the updated edge weight of each residual connection edge. The original edge weights of the corresponding connecting edges in the power grid topology graph are replaced with the updated edge weights to obtain a contracted subgraph.

5. The method of claim 4, wherein, In step S2, obtaining the contracted subgraph further includes: The ratio of the number of fault-immune nodes stripped at each time step to the total number of nodes in the power grid topology graph is used as the graph shrinkage rate. When the graph shrinkage rate exceeds a preset shrinkage threshold, the node feature vectors of each residual node in the shrinkage subgraph are renormalized according to the current real-time power flow data of each node to obtain the updated shrinkage subgraph. When the graph shrinkage rate does not exceed the preset shrinkage threshold, the original node feature vectors of each residual node in the shrunken subgraph are retained, and the shrunken subgraph is directly output.

6. The method for identifying the propagation path of cascading faults in a power grid based on a graph neural network according to claim 5, characterized in that, In step S3, the shrunken subgraph is input into a graph neural network, using the node feature vector as the initial input. After multi-layer neighborhood aggregation, the risk score of each node is output, including: The node feature vector of each node in the contracted subgraph is used as the input of the 0th layer. When aggregating at the kth layer, the hidden layer vectors of all the neighboring nodes of each node at the (k-1)th layer are weighted and summed according to the update edge weight of the corresponding connection edge. After being concatenated with the hidden layer vector of the corresponding node at the (k-1)th layer, the hidden layer vector of the node at the kth layer is obtained by linear transformation and activation function processing. After L-layer neighborhood aggregation, the hidden layer vector of each node in the L-layer is input into the scoring mapping layer. After linear transformation of the hidden layer vector of each node in the L-layer, it is normalized by the Sigmoid function to obtain the risk score of each node at the current time step. Nodes whose risk scores exceed a preset risk threshold are selected as candidate frontier nodes. Nodes that have been identified as fault-immune nodes are removed from all candidate frontier nodes, and the remaining candidate frontier nodes form an active frontier node set.

7. The method for identifying the propagation path of cascading faults in a power grid based on a graph neural network according to claim 1, characterized in that, In step S4, the contracted subgraphs of each time step are arranged chronologically, and the propagation increment is used to reconstruct the cascading failure propagation trajectory, including: The difference between the active frontier node set at time step t and time step (t-1) is used as the propagation increment node set at time step t. The propagation increment node sets at each time step are connected end to end in time sequence to obtain the chain fault propagation trajectory. Taking the node with the highest risk score in the first time step as the fault origin point, the risk scores of each node in the incremental node set at each time step in the chain fault propagation trajectory are arranged from high to low to obtain the fault propagation path; Based on the product of the cumulative risk score of each node across all time steps and the frequency of its appearance in the set of active frontier nodes, a comprehensive rating is given to each node in the fault propagation path to obtain the equipment risk rating.

8. A power grid cascading fault propagation path identification system based on graph neural networks, characterized in that, For implementing the power grid cascading fault propagation path identification method based on graph neural networks as described in any one of claims 1-7, the power grid cascading fault propagation path identification system based on graph neural networks comprises: The module is used to map the busbars, lines, and transformers into graph nodes and connecting edges according to their electrical coupling relationships. The real-time power flow data, protection action signals, and switch status of each node are used as node feature vectors to construct the power grid topology graph. The calibration module is used to use the tripping triggered by the protection action signal and the opening confirmed by the switch status as dual criteria to calibrate the fault-immune node from the power grid topology diagram, remove the fault-immune node from the power grid topology diagram, and redistribute the power flow load it carries to the edge weight of the residual connection edge according to the capacity margin ratio of each neighboring node to obtain a contracted subgraph. The filtering module is used to input the shrinking subgraph into the graph neural network, take the node feature vector as the initial input, and output the risk score of each node after multi-layer neighborhood aggregation. The active frontier node set is obtained by filtering the risk scores. The output module is used to arrange the contracted subgraphs of each time step in chronological order, use the node difference of the active frontier node set at adjacent time points as the propagation increment, reconstruct the cascading fault propagation trajectory, and take the node with the highest risk score in the first time step as the fault origin point to output the fault propagation path and equipment risk rating.

9. The system according to claim 8, characterized in that, The power grid topology is constructed by mapping busbars, lines, and transformers to graph nodes and connecting edges according to their electrical coupling relationships. Real-time power flow data, protection action signals, and switch status of each node are used as node feature vectors. This topology includes: Using the busbar as the voltage aggregation node, the line as the power transmission node, and the transformer as the voltage transformation node, a set of nodes and a set of connecting edges are established according to the electrical coupling relationship between each device to obtain the initial topology framework; Based on the impedance values ​​between the nodes at both ends of each connecting edge in the initial topology framework, each connecting edge is assigned an initial edge weight to obtain a weighted topology graph. By concatenating the real-time power flow data, protection action signals, and switch status of each node in the current sampling period according to the node number sequence, a node feature vector is constructed for each node in the weighted topology diagram to obtain the power grid topology diagram.

10. The system according to claim 8, characterized in that, The tripping is triggered by the protection action signal, and the opening is confirmed by the switch status as a dual criterion, including: The first criterion is whether the time difference between the triggering time of the protection action signal of each node and the opening confirmation time of the switch state falls within a preset time window. Nodes whose time difference falls within the preset time window are marked as candidate immune nodes. Based on the consistency verification result of the circuit breaker position signal and the remote signaling confirmation signal corresponding to each candidate immune node as the second criterion, the candidate immune node that passes the consistency verification is finally confirmed as the fault immune node.

Images