An emergency training-oriented power grid fault propagation path intelligent deduction and guidance method

Abstract

The application discloses an emergency training-oriented power grid fault propagation path intelligent deduction and guidance method, relates to the technical field of power system safety and simulation, and discloses the following technical scheme: a dynamic graph network is constructed, and a topological mutation mask update is carried out on an adjacency matrix according to an operator instruction; a graph attention network is adopted to deduce a fault probability distribution and obtain a fault propagation path through physical constraint pruning; operator data of an emergency disposal is obtained through rolling and advanced deduction according to the fault propagation path; a real deduction flow and a counterfactual deduction flow are generated by adopting a causal intervention mechanism, and a node difference parameter is calculated; a teaching guidance signal is generated through visual playback and differential marking according to the node difference parameter, and a consequence severity is evaluated to perform a hierarchical intervention; a fault propagation termination node set when the hierarchical intervention is triggered is recorded and encoded as a bias vector to update the graph attention network for next deduction, and the application realizes closed-loop fusion of power grid fault deduction and emergency training guidance.

Description

Technical Field

[0001] This invention relates to the field of power system safety and simulation technology, and in particular to an intelligent simulation and guidance method for power grid fault propagation paths for emergency training. Background Technology

[0002] As a critical infrastructure vital to the national economy and people's livelihood, the safe and stable operation of the power grid is of paramount importance. Reviewing and simulating typical power grid accidents is a crucial means of improving the emergency response capabilities of operation and maintenance personnel and preventing similar accidents from recurring. However, existing power grid accident simulation training techniques still have the following shortcomings in fault propagation simulation and teaching guidance: First, most existing fault prediction methods based on graph neural networks model the power grid as a static graph structure, which cannot reflect the real-time topological changes caused by switching operations, resulting in serious deviations between the prediction results and the actual electrical state, and insufficient prediction accuracy.

[0003] Secondly, existing graph neural networks are trained entirely based on historical fault data, and their outputs lack the constraints of power grid physical laws. This can easily generate non-physical propagation paths that violate power flow limits or voltage boundaries, which can seriously mislead trainees' understanding of the evolution of accidents in training scenarios.

[0004] Third, existing power grid training simulation systems mostly rely on preset scripts or manual operation to trigger simulations, which is a passive response mode. They lack the ability to proactively warn and intervene in a graded manner based on the simulation results, and cannot take differentiated guidance measures according to the actual risk level caused by the operation.

[0005] Fourth, the existing training system only generates a single exercise report for evaluating trainees. Trainee error information is not recorded and fed back into the optimization process of the deduction model, which means that the training content cannot achieve targeted reinforcement and personalized evolution for individual cognitive weaknesses. Summary of the Invention

[0006] In view of the above-mentioned prior art, the present invention provides an intelligent simulation and guidance method for power grid fault propagation paths for emergency training, which mainly solves the technical problems existing in the background art.

[0007] To achieve the above objectives, the technical solution of this invention is implemented as follows: This invention provides an intelligent simulation and guidance method for power grid fault propagation paths for emergency training, the method comprising the following steps: S1: Obtain power grid topology data, real-time operation measurement data and student operation instructions, construct a dynamic graph network, and update the adjacency matrix of the dynamic graph network with a topology mutation mask according to the student operation instructions to obtain the updated dynamic graph network. S2: Based on the updated dynamic graph network, a graph attention network is used to deduce the fault probability distribution of each node, and the fault probability distribution is pruned according to physical constraints to obtain the fault propagation path. S3: Perform rolling forward simulation based on the fault propagation path to obtain the emergency response operation data of the trainees. Based on the emergency response operation data, use a causal intervention mechanism to generate a real simulation flow and a counterfactual simulation flow, and calculate the electrical quantity deviation of the real simulation flow and the counterfactual simulation flow at each corresponding node to obtain the node difference parameter. S4: Based on the node difference parameters, the real-world deduction stream and the counterfactual deduction stream are visually replayed. During the replay, nodes with deviations exceeding the preset range are marked differently to obtain teaching guidance signals. The severity of the consequences is assessed based on the node difference parameters, and graded interventions are implemented based on the severity of the consequences. S5: Record the set of fault propagation termination nodes when tiered intervention is triggered in each exercise, encode the set of fault propagation termination nodes into a bias vector, update the graph attention network with the bias vector, and use the updated graph attention network for the next simulation.

[0008] As a preferred embodiment of the present invention, step S1 specifically includes the following process: Based on the power grid topology data, the power grid equipment is mapped as nodes of the graph network, and the electrical connection relationships and protection logic relationships are mapped as edges of the graph network, thus constructing an initial dynamic graph network; Based on real-time operational measurement data, electrical measurement data of each node at multiple consecutive time sections are extracted using a preset time sliding window. After normalization processing, the data are spliced ​​into a time-series feature tensor for each node, and the time-series feature tensor is attached to the corresponding node of the initial dynamic graph network. Based on the student's operation instructions, detect whether the student's operation instructions include the opening and closing of a switch or disconnector. When the split-merge operation is detected, the range of the subgraph affected by the split-merge operation is determined according to the position of the operation object in the initial dynamic graph network; A topological mutation mask is generated based on the subgraph range, and the adjacency matrix of the initial dynamic graph network is updated using the topological mutation mask to obtain the updated dynamic graph network.

[0009] As a preferred embodiment of the present invention, the specific process of inferring the fault probability distribution of each node using a graph attention network based on the updated dynamic graph network in step S2 includes: The updated dynamic graph network is input into the graph attention network, which contains multiple parallel attention heads, each of which calculates the attention coefficients between adjacent nodes in the updated dynamic graph network. The features of neighboring nodes are aggregated based on the attention coefficient to update the embedding representation of each node, and the failure probability distribution of each node at subsequent time steps is output based on the updated embedding representation.

[0010] As a preferred embodiment of the present invention, the specific process of pruning the fault probability distribution according to physical constraints to obtain the fault propagation path in step S2 includes: The candidate propagation paths are obtained from the initial fault probability distribution and sorted from high to low probability values. The line thermal stability limit verification, node voltage over-limit verification and power balance verification are performed on each candidate propagation path in sequence. The probability values ​​of candidate propagation paths that do not meet any of the above checks are multiplied by a preset attenuation coefficient. The original probability values ​​of candidate propagation paths that meet all the above checks are retained. The probability values ​​of all candidate propagation paths are then normalized again. The path with the highest normalized probability value is selected as the fault propagation path.

[0011] As a preferred embodiment of the present invention, the specific process of obtaining the trainees' emergency response operation data by performing rolling forward simulation based on the fault propagation path in step S3 includes: The fault propagation path is recalculated on a rolling basis at a preset time period to predict the trend of fault probability change of each node on the fault propagation path within a preset time period in the future. When it is determined that the probability of a node failing within a preset time period exceeds a preset threshold, an early warning message is generated on the student's operation interface and a time-limited decision-making interaction is triggered; the emergency response plan input by the student within the time limit is collected as the emergency response operation data.

[0012] As a preferred embodiment of the present invention, the specific process of generating the reality deduction stream and the counterfactual deduction stream using the causal intervention mechanism based on the emergency response operation data in step S3 includes: The dynamic graph network is updated based on the emergency response data of the trainees. The graph attention network is then used to deduce the fault propagation path based on the updated dynamic graph network, serving as the actual deduction flow: The location of the operation node in the dynamic graph network is determined based on the emergency response operation data, and the subgraph range affected by the operation is determined along the propagation direction of the edge starting from the operation node. In the graph attention network, the output of the attention head corresponding to the subgraph range is masked, and the parent node dependency of the operation node in the causal graph is deactivated to simulate the graph state after the operation takes effect. Based on the graph state, the graph attention network is used to infer the fault propagation path as a counterfactual inference flow.

[0013] As a preferred embodiment of the present invention, the specific process of calculating the electrical quantity deviation between the actual deduction flow and the counterfactual deduction flow at each corresponding node in step S3 to obtain the node difference parameter is as follows: Based on the node difference parameters, the voltage amplitude, active power and reactive power of the actual deduction stream and the counterfactual deduction stream at each corresponding node are obtained respectively. Calculate the voltage amplitude deviation, active power deviation, and reactive power deviation of each corresponding node, and use them as the electrical quantity deviations. The electrical quantity deviations of each corresponding node are associated with the node identifier and the timestamp to obtain the node difference parameter.

[0014] As a preferred embodiment of the present invention, the specific process of differentially marking nodes whose deviations exceed the preset range during playback to obtain teaching guidance signals in step S4 includes: The electrical quantity deviation of each corresponding node is compared with a preset deviation threshold, and the node whose electrical quantity deviation exceeds the preset deviation threshold is determined as a deviation node; The deviation nodes are marked with a highlighting method, and causal prompt text is generated based on the electrical quantity deviation of the deviation nodes. The marked visual playback screen and the causal prompt text are output to the student operation interface as the teaching guidance signal.

[0015] As a preferred embodiment of the present invention, step S4, specifically the process of assessing the severity of consequences based on the node difference parameters and performing graded interventions based on the severity of consequences, includes: Based on the node difference parameters, the number of nodes whose electrical quantity deviation exceeds the preset deviation threshold and the electrical level of the nodes in the power grid topology are counted. The severity of the consequences caused by the student's operation is comprehensively evaluated by combining the number of nodes and the electrical level, and the severity of the consequences is divided into the first level, the second level and the third level. When the severity of the consequences is at the first level, a text prompt message is displayed on the student's operating interface in a non-blocking manner. When the severity of the consequences is at the second level, the current simulation process is forcibly paused, the dual-screen comparison is automatically replayed, and the simulation process is resumed after the replay is completed. When the severity of the consequences reaches the third level, an operation lockout command is sent to the actuator to lock the corresponding remote control operation channel, and a full-screen safety warning is triggered on the student's operation interface.

[0016] As a preferred embodiment of the present invention, step S5 specifically includes: An induced subgraph of the set is constructed based on the adjacency relationships of each node in the set of fault propagation termination nodes in the dynamic graph network. The adjacency matrix of the induced subgraph is extracted, and the adjacency matrix is ​​projected onto a fixed-dimensional space to obtain the bias vector. Calculate the cross-entropy between the bias vector and the attention weights of the nodes corresponding to the set of fault propagation termination nodes in the graph attention network. Add the cross-entropy as a regularization term to the loss function of the graph attention network. Use the loss function with the regularization term added to iteratively update the parameters of the graph attention network. Compared with the prior art, the beneficial effects of the present invention are as follows: This invention enables graph attention network deduction to adapt to topological changes caused by operations through a topological mutation mask update mechanism, reducing invalid computations in isolated regions and helping to improve the consistency between deduction results and actual electrical states.

[0017] This invention integrates physical constraint pruning with graph attention network probabilistic inference, so that the output fault propagation path is verified by the physical laws of the power grid while satisfying the statistical regularity of historical data. This helps to reduce the probability of generating non-physical paths and improve the rationality and credibility of the inference results.

[0018] This invention, through rolling forward deduction and causal intervention mechanisms, enables deduction to have the ability to proactively predict and compare, which helps to enrich the dimensions of information presentation in training scenarios.

[0019] This invention provides differentiated learning feedback to learners by using graded interventions based on the severity of consequences, ensuring that the intensity of instructional guidance is adapted to the actual level of risk.

[0020] This invention encodes trainee error data into bias vectors and feeds them back into the update process of the graph attention network, enabling the inference model to be gradually adjusted as trainees accumulate individual operational data. This helps the training content to focus on trainees' cognitive weaknesses and provides technical support for personalized training. Attached Figure Description

[0021] Figure 1 This is a schematic diagram illustrating the steps of an intelligent simulation and guidance method for power grid fault propagation paths aimed at emergency training. Detailed Implementation

[0022] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. In the following description, the expression "some embodiments" refers to a subset of all possible embodiments; however, it should be understood that "some embodiments" can be the same subset or different subsets of all possible embodiments and can be combined with each other without conflict.

[0023] In the following description, numerous specific details are set forth in order to provide a more thorough understanding of the invention. However, it will be apparent to those skilled in the art that the invention can be practiced without one or more of these details. In other instances, certain technical features well-known in the art have not been described in order to avoid obscuring the invention.

[0024] It should be understood that the present invention can be embodied in various forms and should not be construed as being limited to the embodiments set forth herein. Rather, providing these embodiments will make the disclosure thorough and complete, and will fully convey the scope of the invention to those skilled in the art. Furthermore, the terminology used herein is intended only to describe particular embodiments and is not intended to limit the invention. When used herein, the singular forms “a,” “an,” and “the” are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the terms “compose” and / or “comprising,” when used in this specification, identify the presence of the stated features, integers, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups. When used herein, the term “and / or” includes any and all combinations of the associated listed items.

[0025] It should also be noted that when an element is referred to as being "fixed to" another element, it can be directly attached to the other element or there may be an intervening element. When an element is referred to as being "connected to" another element, it can be directly connected to the other element or there may be an intervening element. The terms "vertical," "horizontal," "inner," "outer," "left," "right," and similar expressions used herein are for illustrative purposes only and do not represent the only possible implementation.

[0026] To fully understand this invention, a detailed structure will be presented in the following description to illustrate the technical solution proposed by this invention. Optional embodiments of the invention are described in detail below; however, in addition to these detailed descriptions, the invention may have other embodiments.

[0027] This application provides an intelligent simulation and guidance method for power grid fault propagation paths for emergency training. Please refer to the attached document for details. Figure 1 The method includes the following steps: S1: Obtain power grid topology data, real-time operational measurement data, and student operation instructions, construct a dynamic graph network, and update the adjacency matrix of the dynamic graph network with a topology mutation mask according to the student operation instructions to obtain the updated dynamic graph network.

[0028] In this embodiment, firstly, power grid topology data is obtained from the power grid dispatch automation system. The topology data includes the physical connection relationships of primary equipment such as buses, incoming and outgoing line bays, circuit breakers, disconnect switches, and grounding switches at each voltage level within the substation, as well as the start and end point information of inter-substation transmission lines.

[0029] Based on the acquired topology data, a graph network mapping is performed: primary equipment in the power grid (including substation nodes, buses, transformers, circuit breakers, lines, etc.) is mapped as nodes of the graph network; electrical connections between equipment (such as conductor connections, bus connections, circuit breaker series connections, etc.) and relay protection logic relationships (such as the objects protected by a protection device and its action output logic) are mapped as edges of the graph network, thereby constructing an initial dynamic graph network. ,in, For a set of nodes, Let it be the set of edges.

[0030] Secondly, real-time operational measurement data is obtained from the power grid dispatch automation system. The measurement data includes the voltage amplitude, current amplitude, active power, and reactive power of each node.

[0031] Let the length of the time sliding window be... In this embodiment, take There are 10 time sections, and the time interval between each section is... In this embodiment, For each node Take it from the past Measurement data within each cross section constitute the original measurement sequence for that node. :

[0032] in, Represents a node At any moment The electrical measurement vector includes four dimensions: voltage amplitude, current amplitude, active power, and reactive power.

[0033] For the original measurement sequence Z-score normalization is performed to ensure that the mean of each dimension is 0 and the standard deviation is 1, eliminating differences between different units of measurement. The normalized cross-sectional data are then concatenated in chronological order to form the temporal feature tensor for that node. The temporal feature tensor Attached to the initial dynamic graph network The corresponding node The above serves as the initial feature of this node.

[0034] Finally, the system acquires the switching operation commands input by the trainee through the operating terminal. These commands include the operation object (e.g., "221 switch") and the operation type ("open" or "close"). The system also monitors in real-time whether the trainee's operation commands include opening or closing operations for switches or disconnectors.

[0035] When a split / merge operation command is detected, a topology mutation mask update procedure is executed. First, based on the position of the operation object in the initial dynamic graph network, the range of the subgraph affected by the split / merge operation is determined. Specifically: Let's take a specific circuit breaker tripping operation as an example. Suppose the student executes the operation command "disconnect circuit breaker 221". Locate the node corresponding to circuit breaker 221 in the dynamic graph network. And the two edges it connects to, namely the upstream side connection. of The connection between the side and downstream side of the No. 1 busbar The high-voltage side of the main transformer.

[0036] When a circuit breaker changes from the closed position to the open position, from an electrical topology perspective, the downstream sub-diagram of the circuit breaker (i.e., The main transformer and its connected medium- and low-voltage side equipment area are isolated from the energized network, and all nodes within this area no longer participate in the current power grid fault propagation process. Therefore, using the 221 circuit breaker node as the boundary point, a downstream search is conducted along the edge to determine the downstream sub-graph affected by this tripping operation. In this embodiment, include The main transformer node and its connected medium-voltage side busbars, feeder switches, and other nodes. This includes all edges within the downstream subgraph and cross-regional connection edges with the external network.

[0037] Determine the affected downstream subplot range Then, a topological mutation mask corresponding to the downstream subgraph is generated. The mask and the adjacency matrix of the dynamic graph network. They have the same dimensions. The topological mutation mask. The generation rule is as follows: for the adjacency matrix Zhongyu The rows and columns associated with any node within the topological mutation mask The corresponding position is set to 0; the remaining positions (areas not involving the downstream subgraph) are set to 1.

[0038] Furthermore, utilizing the aforementioned topological mutation mask For the initial dynamic graph network adjacency matrix Perform element-by-element mask update:

[0039] Here, ⊙ represents element-wise multiplication. It sets the weights of all edges within the downstream subgraph region affected by the operation, as well as the propagation edges between this region and the external region, to zero, blocking the participation of nodes and associated edges within this downstream subgraph region in subsequent graph attention network propagation calculations, while preserving the complete topological connectivity of unaffected regions. This represents the updated adjacency matrix. The corresponding dynamic graph network is the updated dynamic graph network. ,in, Only edges with a weight greater than zero after masking are included.

[0040] Through the aforementioned topology mutation mask update mechanism, each inference of the graph attention network is based on a dynamic graph network that accurately reflects the actual electrical connection status of the current power grid, avoiding invalid calculations on the isolated downstream subgraph, thereby ensuring the effectiveness and accuracy of the inference results.

[0041] S2: Based on the updated dynamic graph network, the failure probability distribution of each node is deduced using a graph attention network, and the failure probability distribution is pruned according to physical constraints to obtain the failure propagation path.

[0042] In this embodiment, the obtained updated dynamic graph network The input graph attention network comprises an input layer, multiple parallel attention heads, a feature aggregation layer, and an output layer. In this embodiment, the number of attention heads is set. Each attention head independently calculates the attention weights between adjacent nodes.

[0043] For the updated dynamic graph network any node in and its neighboring nodes (i.e., there are edges) (Connecting the two), attention head Calculate any node and its neighboring nodes Original attention coefficient :

[0044] in, and Each node and its neighboring nodes eigenvectors, For attention head The linear transformation weight matrix, For attention head Attention parameter vector, This represents a vector concatenation operation. This is the activation function.

[0045] Then, the original attention coefficients Perform Softmax normalization to obtain arbitrary nodes For all its neighboring nodes attention weights :

[0046] in, For the summation variable, For any node The set of all neighboring nodes, Indicate attention head Next node For any of its neighbors The original attention coefficient, This indicates that during the fault propagation process, any node The fault status of the node's neighboring nodes The degree of importance of the impact. A higher weight indicates that the fault originated from any node. propagate to neighboring nodes The greater the likelihood, the higher the probability.

[0047] After obtaining the attention weights of each attention head, each attention head aggregates the features of its neighboring nodes to update the embedding representation of the current node. Specifically, the attention heads... For any node The embedded update is :

[0048] in, For non-linear activation functions, this embodiment uses the ELU function.

[0049] Will The outputs of each attention head are concatenated to obtain an arbitrary node. Comprehensive Embedded Representation :

[0050] The comprehensive embedded representation of all nodes Input / output layer, outputting the results of each node at subsequent time steps. Failure probability distribution In this embodiment, the output is simultaneous. The probability distributions at the three time points correspond to the evolution trends within three power frequency cycles (60ms) after the fault occurs. This indicates a vector concatenation operation.

[0051] The fault probability distribution output by a graph attention network may contain illogical paths that violate the laws of power grid physics. For example, it might deduce that the active power flowing through a line after a fault exceeds its thermal stability limit, or that the voltage at a node drops below the limit, yet the model still determines that the fault will continue to propagate along that path. Therefore, it is necessary to prune the initial deduction results with constraints based on the laws of power grid physics.

[0052] First, from the fault probability distribution, the candidate propagation paths are sorted from high to low probability values, and the top-ranked paths are selected. In this embodiment, the top-ranked paths are selected... Candidate paths Each candidate path is represented as an ordered sequence of nodes that start from the fault originating node and reach a certain termination node.

[0053] Then, three physical checks are performed on each candidate propagation path: (1) Line thermal stability limit verification. For each transmission line included in the candidate propagation path, calculate the active power flowing through the transmission line after the fault based on the node voltage amplitude and phase angle under the current operating mode. .like Exceeding the thermal stability limit of the transmission line (Determined by the transmission line model and operating environment temperature), then the candidate propagation path is determined not to meet the thermal stability limit constraint.

[0054] The active power The calculation is based on the voltage amplitude at both ends of the node. and phase angle difference :

[0055] in, For power transmission lines - Transmission lines The reactance value. If If the candidate propagation path fails the thermal stability limit check, then the candidate propagation path fails the thermal stability limit check.

[0056] (2) Node voltage over-limit verification. For each node included in the candidate propagation path, calculate the voltage amplitude of that node after the fault. .like Below the lower voltage limit (Typically 0.9 times the rated voltage) or higher than the upper voltage limit. If the voltage exceeds the limit (usually 1.1 times the rated voltage), then the candidate propagation path is determined to have failed the voltage limit check.

[0057] (3) Power balance verification. For each bus node in the candidate propagation path, verify whether the injected power and outflow power of the bus node satisfy the power balance equation:

[0058] in, This is the sum of the active power injected into this bus node. It is the sum of active power flowing out of this bus node. This refers to the power loss of the equipment connected to the bus node. If the absolute value of the difference between the injected power and the outflow power plus the losses exceeds the preset tolerance... (This embodiment takes) If the power balance check fails, then the candidate propagation path is deemed to have failed the power balance check.

[0059] After performing the above three checks on each candidate propagation path in sequence, pruning is performed based on the check results: the probability value of the candidate propagation path that does not meet any of the checks is multiplied by a preset attenuation coefficient. (This embodiment takes) Candidate propagation paths that simultaneously satisfy all three checks retain their original probability values.

[0060] After performing the above processing on all candidate propagation paths, the probability values ​​are re-normalized, and the candidate propagation path with the highest normalized probability value is selected as the fault propagation path. .

[0061] S3: Perform rolling forward simulation based on the fault propagation path to obtain the emergency response operation data of the trainees. Based on the emergency response operation data, use a causal intervention mechanism to generate a real-world simulation flow and a counterfactual simulation flow, and calculate the electrical quantity deviation of the real-world simulation flow and the counterfactual simulation flow at each corresponding node to obtain the node difference parameter.

[0062] In this embodiment, a preset time period is used. For the fault propagation path Rolling recalculation is performed. In this embodiment... Each time the calculation is performed, the real-time running data at the current moment is used as the initial condition, and the graph attention network inference method described in step S2 is used to predict the future preset duration. The trend of fault probability changes at each node along the internal fault propagation path. In this embodiment, .

[0063] Specifically, in each rolling simulation, the predicted failure probability curve for each node within the next 3 seconds is output, and the system monitors in real time whether the failure probability of any node exceeds a preset threshold at any future moment. In this embodiment When the probability value reaches 0.75 or higher, it indicates that the propagation trend of the fault is highly deterministic and sufficient to trigger an early warning.

[0064] Furthermore, when it is determined that the probability of a node failing within the next 3 seconds exceeds 0.75, a warning message is generated on the student's operating interface. The warning message displays the expected propagation direction and speed of the fault along the power grid topology in the first display area of ​​the interface using a dynamic arrow animation; the second display area of ​​the interface displays the fault prediction description text in natural language (e.g., "It is expected that the fault will propagate to the high-voltage side of the No. 3 main transformer in 3 seconds, and it is recommended to immediately perform a switching operation").

[0065] After the warning information is displayed, the system displays a time-limited decision button on the user interface. In this embodiment, the time limit is... The interface provides trainees with multiple emergency response options (such as "Execute switching operation", "Activate automatic transfer switch", "Request dispatch instruction", etc.). Trainees must select and submit an emergency response plan within 3 seconds, and the plan will be collected as the emergency response operation data. If the trainee does not make a selection within the time limit, it will be recorded as "no operation within time limit" by default, and the emergency response plan will be automatically executed according to the preset safety policy. At the same time, "timeout" will be recorded as the trainee's operation characteristic data.

[0066] Furthermore, using the trainees' emergency response operation data as input, a topological mutation mask update is performed on the adjacency matrix of the current dynamic graph network. Specifically, the dynamic graph network is masked according to the switching / disconnector states involved in the trainees' actual operations. Then, using the aforementioned graph attention network and physical pruning method, the fault propagation path is inferred based on the updated dynamic graph network. This fault propagation path reflects the actual evolution of the power grid fault after the trainees' actual operations, i.e., the real-world inference flow.

[0067] Furthermore, the operation nodes (i.e., the equipment nodes directly affected by the trainee's operation, such as the "221 circuit breaker" node) are extracted from the trainee's emergency response operation data, and the position of the operation node is located in the dynamic graph network. Starting from the operation node, the subgraph range affected by the operation is determined along the propagation direction of the edge. The method for determining the subgraph range is the same as the method for determining the affected subgraph in step S1, that is, the downstream electrical connection area disconnected by the switch operation.

[0068] The graph state of each node within the subgraph is changed from being affected by upstream fault propagation to an independent state based solely on the state of the node itself. Based on this, using the intervened graph state as input, the fault propagation path is re-inferred using the graph attention network and physical pruning method described in step S2. This fault propagation path simulates the evolution of the fault under the condition of "assuming the operation was executed correctly (or not executed)," which is the counterfactual inference flow.

[0069] Furthermore, at the same time point, the electrical quantity data of the actual deduction stream and the counterfactual deduction stream at each corresponding node are acquired. The electrical quantity data includes the voltage amplitude, active power, and reactive power of the node.

[0070] For each node in the power grid Calculate the voltage amplitude deviation under the actual deduced current and the counterfactual deduced current. Active power deviation reactive power deviation The specific calculation formula is as follows:

[0071] in, For nodes Voltage amplitude under realistic current conditions For nodes Voltage amplitude under counterfactual current.

[0072] Active power deviation :

[0073] in, For nodes Active power under realistic simulation For nodes Active power under counterfactual reasoning flow.

[0074] reactive power deviation :

[0075] in, For nodes Reactive power under realistic simulation For nodes Reactive power under counterfactual reasoning.

[0076] The voltage amplitude deviation Active power deviation reactive power deviation The combination forms the electrical quantity deviation vector of this node. .

[0077] The electrical quantity deviation vectors of all nodes are structured and associated with the node identifier (such as device ID) and the current timestamp to form the node difference parameters. .in, This represents the total number of nodes in the set.

[0078] S4: Based on the node difference parameters, the real-world deduction stream and the counterfactual deduction stream are visually replayed. During the replay, nodes with deviations exceeding a preset range are marked differently to obtain teaching guidance signals. The severity of the consequences is assessed based on the node difference parameters, and graded interventions are implemented based on the severity of the consequences.

[0079] In this embodiment, based on the node difference parameters The real-world deduction stream and the counterfactual deduction stream are visualized and replayed in a synchronous dual-screen comparison format.

[0080] The synchronous dual-screen comparison is as follows: the first display area (left half-screen) of the student's operating interface displays the dynamic topology rendering of the real-world deduction flow, i.e., the actual propagation process of the power grid fault after the student's actual operation. Dynamic arrow animations show the direction of fault propagation along the power grid topology, and red / orange / yellow / green risk levels indicate the fault risk status of each node. The second display area (right half-screen) synchronously displays the dynamic topology rendering of the counterfactual deduction flow, i.e., the comparative fault propagation process under the assumption of correct operation. The timelines of the two display areas are completely synchronized.

[0081] During playback, based on the node difference parameters Electrical quantity deviations at each corresponding node Real-time monitoring is performed. The electrical quantity deviation of each corresponding node is compared with a preset deviation threshold to determine whether it is a deviation node, as follows: For voltage amplitude deviation, a preset voltage deviation threshold is provided. .like Determine the node There is a significant deviation in the voltage dimension.

[0082] For active power deviation, a preset active power deviation threshold is set. .like Determine the node There is a significant deviation in the power dimension.

[0083] For reactive power deviation: a preset reactive power deviation threshold is used. .like Determine the node There is a significant deviation in the reactive power dimension.

[0084] If node If the deviation of any item exceeds the corresponding preset threshold, the node is determined to be a deviation node.

[0085] Each node identified as a deviation node and its connected branches are marked with a pulsed flashing highlight using an independent color distinct from the basic risk color scale (red / orange / yellow / green). This allows trainees to quickly locate areas where there are significant differences between the two simulation flows. These areas are precisely where the different consequences of "trainee's actual operation" and "correct operation" are concentrated.

[0086] Simultaneously, causal prompt text is automatically generated based on the maximum deviation among the deviation nodes. This causal prompt text employs a template-based natural language generation algorithm. Specifically, it iterates through the set of deviation nodes to locate the voltage amplitude deviation. Largest node and active power deviation Largest node Extract the device name and deviation, and fill them into a preset text template: such as "You [did not execute / incorrectly executed] the operation of [operation object], resulting in the voltage amplitude of [device name] being higher than the standard operation." [Value] kV, active power is too high [ Value] MW. For example, in a certain exercise, the node with the largest deviation was "Main Transformer No. 3". The generated cause-effect prompt text was: "You did not disconnect the 221 circuit breaker in time, which caused the voltage amplitude on the high-voltage side of Main Transformer No. 3 to be 1.2kV higher than the standard operating value and the active power to be 8.5MW higher. This was caused by the voltage drop of Bus No. 3 due to the fault crossing the main transformer protection." Furthermore, the marked dual-screen visual playback and the aforementioned causal prompt text are used together as the teaching guidance signal and output to the student's operating interface. This teaching guidance signal allows the student to intuitively see the differences between their own operation and the correct operation in terms of the evolution of the power grid state and its physical consequences.

[0087] Furthermore, based on the node difference parameters The severity of the consequences of the trainee's actions is assessed. The severity assessment uses the following two core indicators: Indicator 1: Number of Deviation Nodes The electrical quantity deviations were statistically analyzed. The number of nodes exceeding the preset deviation threshold. The larger the value, the wider the range of power grid state deviations caused by the student's operation.

[0088] Indicator 2: Maximum Electrical Layer The busbars and incoming / outgoing lines directly connected to the power source are defined as Level 1 (highest level, most important); nodes connected to Level 1 equipment via a transformer or line are defined as Level 2; and so on. The smaller the electrical level value, the more critical the node's topological position in the power grid, and the greater the impact of its state deviation on the overall security of the power grid. Traverse all deviation nodes, find the node with the smallest electrical level, and record it as the node with the largest electrical level. .

[0089] Based on the above two indicators, the severity of the consequences is divided into three levels: Level 1 (low risk), Level 2 (medium risk), and Level 3 (high risk), with the severity increasing sequentially. The classification rules are as follows: like and It was determined to be at the first level.

[0090] like and ,or and It was determined to be at the second level.

[0091] like ,or It was determined to be at level three.

[0092] Specifically, the above classification rules are based on the following: the state deviation of key equipment directly connected to the power supply point has the greatest impact on the overall voltage stability and power balance of the power grid, thus directly triggering the highest risk level; the more deviation nodes there are, the larger the scope of the fault impact, and the higher the risk level after comprehensive evaluation.

[0093] Furthermore, based on the severity of the determined consequences, the following tiered intervention measures will be implemented: (1) When the severity of the consequence is Level 1, it means that although the trainee's operation deviates from the standard operation, it has not caused equipment overload or voltage exceeding the limit, and the power grid is still within the safe operating range. At this time, a text prompt box is displayed on the trainee's operation interface in a non-interrupting manner. The prompt content is a description of the deviation between the current operation and the standard operation and suggested corrective measures. This prompt is displayed in the upper right corner of the interface in the form of a floating window, without interrupting the ongoing simulation process.

[0094] (2) When the severity of the consequence is Level 2, it indicates that the trainee's operation has resulted in a large number of significantly deviated nodes or affected Level 2 equipment. If the current operation path continues, there is a tendency for the fault to escalate. At this time, the current simulation process is forcibly paused for 1.5 seconds, and the complete dual-screen comparison screen from the operation time to the current time is automatically replayed in the dual-screen comparison area, allowing the trainee to observe the evolution of the difference between the correct operation and their own operation during the pause. After the replay ends, the simulation process is automatically resumed, and the trainee can continue operating.

[0095] (3) When the severity of the consequences is Level 3, it means that the trainee's operation will directly cause a state deviation of the Level 1 critical equipment (i.e., may cause busbar undervoltage or generator disconnection), or cause a large-scale (more than 11 nodes) electrical quantity deviation. At this time, immediately send an operation interlock command to the on-site execution mechanism to lock the remote control operation channel of the trainee's current operation, physically preventing the execution of the operation. At the same time, trigger a full-screen red flashing safety warning on the trainee's operation interface, displaying the warning message "This operation will cause [critical equipment name] undervoltage / disconnection, and has been intercepted", and link to play back video clips of similar operations that caused accidents in historical accident cases to enhance the warning effect.

[0096] In summary, by using a tiered intervention mechanism, the intensity of teaching intervention measures is matched with the actual risk level of the trainees' operations. This avoids excessive interference with low-risk operations while implementing physical lock-in for high-risk operations, thus ensuring the safety of the training process.

[0097] S5: Record the set of fault propagation termination nodes when tiered intervention is triggered in each exercise, encode the set of fault propagation termination nodes into a bias vector, update the graph attention network with the bias vector, and use the updated graph attention network for the next simulation.

[0098] In this embodiment, during each exercise, the time points when the second-level and third-level interventions are triggered, as well as the corresponding fault propagation status, are recorded in real time. Specifically, at the instant the second-level or third-level intervention is triggered, the fault probability at the termination node of the fault propagation path—that is, on the propagation path with the highest calculated fault probability—is recorded to have dropped to a preset termination threshold. The last node below, in this embodiment, is taken as... All termination nodes that meet the conditions are combined into a fault propagation termination node set for this exercise.

[0099] Furthermore, through multiple drills, a historical database containing multiple sets of fault propagation termination nodes was accumulated. The union of all sets was extracted from this historical database to obtain a comprehensive termination node set. .

[0100] Based on the comprehensive set of termination nodes Based on the adjacency relationships of each node in the dynamic graph network, a comprehensive set of termination nodes is constructed. Induced subgraph The induced subgraph The covered topological regions, which are the regions where trainees frequently make mistakes in historical exercises and trigger high-level interventions, represent the trainees' cognitive weaknesses.

[0101] Furthermore, extract the induced subgraph. adjacency matrix ),in The adjacency matrix It is a sparse matrix, where the positions of non-zero elements correspond to the edges present in the induced subgraph.

[0102] The adjacency matrix Projecting onto a fixed-dimensional space yields a vector of fixed length, which serves as the bias vector. Specifically, this embodiment employs the following projection method: using the total number of nodes in the graph attention network as the output dimension. Construct a vector of all zeros For the comprehensive set of termination nodes For each node in the projection, its corresponding dimension position is assigned a normalized value to the degree of that node in the induced subgraph (i.e., the number of edges connected to that node). This projection method results in a larger dimension in the bias vector corresponding to the weak region, thus having a greater impact in subsequent calculations.

[0103] Further, the bias vector is calculated. The integrated set of termination nodes in the graph attention network The cross-entropy between the attention weights of the corresponding nodes is used as a regularization term.

[0104] Specifically, suppose a graph attention network integrates the set of termination nodes under a certain attention head. Each node in the middle The average attention weight is This indicates the current level of attention the graph attention network is giving to nodes in weak regions. For nodes Normalized attention weights.

[0105] Specifically, the cross-entropy The calculation formula is:

[0106] Let the regularization term The regularization term is added to the original loss function of the graph attention network. This indicates that the original loss function is the node classification cross-entropy loss, used to accurately predict the direction of fault propagation, and the updated loss function is obtained. :

[0107] in, For regularization coefficients, this embodiment takes... , is used to control the weight strength of the regularization term.

[0108] Furthermore, using the updated loss function The parameters of the graph attention network are iteratively updated using gradient descent backpropagation. Specifically, in each parameter iteration, when the graph attention network synthesizes the set of termination nodes... When the attention weights of nodes in the graph attention network are lower than the attention level expected by the bias vector, the loss function will increase, causing the graph attention network to gradually improve its ability to synthesize the set of termination nodes in subsequent training iterations. Attention weights for nodes. In other words, topological regions where trainees frequently make mistakes will receive higher attention allocations in the graph attention network, making subsequent deductions focus more on the state changes and fault propagation trends of these regions.

[0109] Furthermore, the updated graph attention network is applied to the fault propagation path simulation in the next exercise. After each exercise, the attention weights of the graph attention network are updated based on the newly accumulated error data. The updated network drives the simulation and guidance of the next exercise. As the number of exercises increases, the attention weights of the graph attention network on the weak areas of individual trainees continue to increase, and the simulation and guidance gradually focus on the trainees' individual cognitive shortcomings, thereby achieving personalized training goals.

[0110] In some embodiments, the above-described technical solution of the present invention can be deployed in a multi-position collaborative exercise scenario. In one application scenario, position role tags are pre-configured for each trainee, including but not limited to dispatcher, maintenance, field operation, and commander positions. After generating the realistic simulation stream and counterfactual simulation stream in step S3, the simulation results are adapted and distributed according to each trainee's position role: a fault propagation situation map and regional risk level distribution are pushed to dispatcher trainees; the fault probability distribution and suggested isolation operation sequences for the equipment areas under their jurisdiction are pushed to maintenance trainees; specific switching operation instruction sequences and safety verification results are pushed to field operation trainees; and a global handling progress overview and a summary of the operation status of each position are pushed to commander trainees. Each trainee receives simulation information matching their position responsibilities on their respective operating terminal and independently completes the exercise operation.

[0111] Specifically, in an optional multi-role collaborative assessment scenario, after a drill, the decision timestamps and operation sequences of trainees in each role are collected. Combined with the standard operation path generated in step S3, the operation accuracy and response time of each role are calculated. The delay in cross-role operation connections and the integrity of information transmission are also evaluated. Reference data for multi-role collaborative effectiveness assessment is output, allowing instructors to optimize role configuration and collaboration processes in subsequent drills. The operation data of trainees in each role can be used to update the graph attention network of the corresponding instance in step S5, enabling the inference model for each role to be adapted to the cognitive weaknesses of the trainees in that role.

[0112] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. The scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for intelligent simulation and guidance of power grid fault propagation paths for emergency training, characterized in that, The method includes the following steps: S1: Obtain power grid topology data, real-time operation measurement data and student operation instructions, construct a dynamic graph network, and update the adjacency matrix of the dynamic graph network with a topology mutation mask according to the student operation instructions to obtain the updated dynamic graph network. S2: Based on the updated dynamic graph network, a graph attention network is used to deduce the fault probability distribution of each node, and the fault probability distribution is pruned according to physical constraints to obtain the fault propagation path. S3: Perform rolling forward simulation based on the fault propagation path to obtain the emergency response operation data of the trainees. Based on the emergency response operation data, use a causal intervention mechanism to generate a real simulation flow and a counterfactual simulation flow, and calculate the electrical quantity deviation of the real simulation flow and the counterfactual simulation flow at each corresponding node to obtain the node difference parameter. S4: Based on the node difference parameters, the real-world deduction stream and the counterfactual deduction stream are visually replayed. During the replay, nodes with deviations exceeding the preset range are marked differently to obtain teaching guidance signals. The severity of the consequences is assessed based on the node difference parameters, and graded interventions are implemented based on the severity of the consequences. S5: Record the set of fault propagation termination nodes when tiered intervention is triggered in each exercise, encode the set of fault propagation termination nodes into a bias vector, update the graph attention network with the bias vector, and use the updated graph attention network for the next simulation.

2. The intelligent simulation and guidance method for power grid fault propagation paths for emergency training as described in claim 1, characterized in that, The specific process of step S1 includes: Based on the power grid topology data, the power grid equipment is mapped as nodes of the graph network, and the electrical connection relationships and protection logic relationships are mapped as edges of the graph network, thus constructing an initial dynamic graph network; Based on real-time operational measurement data, electrical measurement data of each node at multiple consecutive time sections are extracted using a preset time sliding window. After normalization processing, the data are spliced ​​into a time-series feature tensor for each node, and the time-series feature tensor is attached to the corresponding node of the initial dynamic graph network. Based on the student's operation instructions, detect whether the student's operation instructions include the opening and closing of a switch or disconnector. When the split-merge operation is detected, the range of the subgraph affected by the split-merge operation is determined according to the position of the operation object in the initial dynamic graph network; A topological mutation mask is generated based on the subgraph range, and the adjacency matrix of the initial dynamic graph network is updated using the topological mutation mask to obtain the updated dynamic graph network.

3. The intelligent simulation and guidance method for power grid fault propagation paths for emergency training as described in claim 1, characterized in that, In step S2, the specific process of using a graph attention network to deduce the failure probability distribution of each node based on the updated dynamic graph network includes: The updated dynamic graph network is input into the graph attention network, which contains multiple parallel attention heads, each of which calculates the attention coefficients between adjacent nodes in the updated dynamic graph network. The features of neighboring nodes are aggregated based on the attention coefficient to update the embedding representation of each node, and the failure probability distribution of each node at subsequent time steps is output based on the updated embedding representation.

4. The intelligent simulation and guidance method for power grid fault propagation paths for emergency training according to claim 3, characterized in that, In step S2, the specific process of pruning the fault probability distribution according to physical constraints to obtain the fault propagation path includes: The candidate propagation paths are obtained from the initial fault probability distribution and sorted from high to low probability values. The line thermal stability limit verification, node voltage over-limit verification and power balance verification are performed on each candidate propagation path in sequence. The probability values ​​of candidate propagation paths that do not meet any of the above checks are multiplied by a preset attenuation coefficient. The original probability values ​​of candidate propagation paths that meet all the above checks are retained. The probability values ​​of all candidate propagation paths are then normalized again. The path with the highest normalized probability value is selected as the fault propagation path.

5. The intelligent simulation and guidance method for power grid fault propagation paths for emergency training according to claim 1, characterized in that, In step S3, the specific process of performing rolling forward simulations based on the fault propagation path to obtain the trainees' emergency response operation data includes: The fault propagation path is recalculated on a rolling basis at a preset time period to predict the trend of fault probability change of each node on the fault propagation path within a preset time period in the future. When it is determined that the probability of a node failing within a preset time period exceeds a preset threshold, an early warning message is generated on the student's operation interface and a time-limited decision-making interaction is triggered; the emergency response plan input by the student within the time limit is collected as the emergency response operation data.

6. The intelligent simulation and guidance method for power grid fault propagation paths for emergency training according to claim 5, characterized in that, In step S3, the specific process of generating the reality simulation stream and the counterfactual simulation stream using the causal intervention mechanism based on the emergency response operation data includes: The dynamic graph network is updated based on the emergency response data of the trainees. The graph attention network is then used to deduce the fault propagation path based on the updated dynamic graph network, serving as the actual deduction flow: The location of the operation node in the dynamic graph network is determined based on the emergency response operation data, and the subgraph range affected by the operation is determined along the propagation direction of the edge starting from the operation node. In the graph attention network, the output of the attention head corresponding to the subgraph range is masked, and the parent node dependency of the operation node in the causal graph is deactivated to simulate the graph state after the operation takes effect. Based on the graph state, the graph attention network is used to infer the fault propagation path as a counterfactual inference flow.

7. The intelligent simulation and guidance method for power grid fault propagation paths for emergency training according to claim 6, characterized in that, In step S3, the specific process of calculating the electrical quantity deviation between the actual deduction stream and the counterfactual deduction stream at each corresponding node to obtain the node difference parameter is as follows: Based on the node difference parameters, the voltage amplitude, active power and reactive power of the actual deduction stream and the counterfactual deduction stream at each corresponding node are obtained respectively. Calculate the voltage amplitude deviation, active power deviation, and reactive power deviation of each corresponding node, and use them as the electrical quantity deviations. The electrical quantity deviations of each corresponding node are associated with the node identifier and the timestamp to obtain the node difference parameter.

8. The intelligent simulation and guidance method for power grid fault propagation paths for emergency training according to claim 1, characterized in that, In step S4, the specific process of differentially marking nodes whose deviations exceed the preset range during playback to obtain teaching guidance signals includes: The electrical quantity deviation of each corresponding node is compared with a preset deviation threshold, and the node whose electrical quantity deviation exceeds the preset deviation threshold is determined as a deviation node; The deviation nodes are marked with a highlighting method, and causal prompt text is generated based on the electrical quantity deviation of the deviation nodes. The marked visual playback screen and the causal prompt text are output to the student operation interface as the teaching guidance signal.

9. The intelligent simulation and guidance method for power grid fault propagation paths for emergency training according to claim 8, characterized in that, In step S4, the specific process of assessing the severity of the consequences based on the node difference parameters and performing graded interventions based on the severity of the consequences includes: Based on the node difference parameters, the number of nodes whose electrical quantity deviation exceeds the preset deviation threshold and the electrical level of the nodes in the power grid topology are counted. The severity of the consequences caused by the student's operation is comprehensively evaluated by combining the number of nodes and the electrical level, and the severity of the consequences is divided into the first level, the second level and the third level. When the severity of the consequences is at the first level, a text prompt message is displayed on the student's operating interface in a non-blocking manner. When the severity of the consequences is at the second level, the current simulation process is forcibly paused, the dual-screen comparison is automatically replayed, and the simulation process is resumed after the replay is completed. When the severity of the consequences reaches the third level, an operation lockout command is sent to the actuator to lock the corresponding remote control operation channel, and a full-screen safety warning is triggered on the student's operation interface.

10. The intelligent simulation and guidance method for power grid fault propagation paths for emergency training according to claim 1, characterized in that, Step S5 specifically includes: An induced subgraph of the set is constructed based on the adjacency relationships of each node in the set of fault propagation termination nodes in the dynamic graph network. The adjacency matrix of the induced subgraph is extracted, and the adjacency matrix is ​​projected onto a fixed-dimensional space to obtain the bias vector. Calculate the cross-entropy between the bias vector and the attention weights of the nodes corresponding to the set of fault propagation termination nodes in the graph attention network. Add the cross-entropy as a regularization term to the loss function of the graph attention network. Use the loss function with the regularization term added to iteratively update the parameters of the graph attention network.

Images