A method for identifying critical lines with cascading faults in power grids under ice storms based on heterogeneous graph neural networks

By constructing a power grid cascading fault impact map using a heterogeneous graph neural network and integrating physical topology and fault propagation relationships, the problem of accuracy and efficiency in identifying critical lines under ice storm conditions was solved, enabling efficient critical line identification and protection decision-making under ice storm conditions.

CN123115945APending Publication Date: 2026-11-17NORTHEAST DIANLI UNIVERSITY +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611308777.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-27
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing technologies are insufficient to accurately identify critical lines of power grid cascading faults under ice storm conditions, especially in reflecting the coupling effects of transmission line icing, wind turbine output degradation and cascading faults, as well as the heterogeneous properties of busbars and lines and their non-local propagation relationships.

Method used

A heterogeneous graph neural network is used to generate fault samples through ice disaster impact model and sequential Monte Carlo simulation, construct a cascading fault impact graph, integrate physical topology and fault propagation relationship, extract local electrical constraints and non-local propagation features, and use an adaptive gating mechanism for feature fusion to achieve critical line identification.

Benefits of technology

It improves the accuracy of identifying critical lines with cascading faults in the power grid under ice storms and the efficiency of online application. It can dynamically coordinate the contributions of physical constraint information and fault propagation information, and is suitable for ice storm early warning and protection decision-making.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123115945A_ABST
    Figure CN123115945A_ABST
Patent Text Reader

Abstract

This invention discloses a method for identifying critical lines in power grid cascading faults under ice storm conditions based on heterogeneous graph neural networks. It belongs to the fields of power system safety and stability analysis, disaster risk early warning, and artificial intelligence applications. The method includes acquiring information on power grid components, topology, operating status, and ice storm forecasts; generating ice storm cascading fault samples based on ice storm impact models and sequential Monte Carlo simulations; calculating line fault reproduction rates and inter-line fault propagation rates according to the intergenerational propagation relationship of the fault chain to construct a cascading fault impact graph; fusing the physical topology and impact graph into a heterogeneous graph of physical connection edges and impact propagation edges; constructing bus node features and line node features, and using an adaptive gating mechanism to complete feature fusion; and finally outputting the predicted probability that the line is a critical line. This invention can simultaneously characterize the environmental impact of ice storms, physical topology constraints, and the non-local propagation characteristics of cascading faults, improving the accuracy of critical line identification and the efficiency of online applications.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the fields of power system safety and stability analysis, disaster risk early warning and artificial intelligence application technology, and specifically relates to a method for identifying critical lines of power grid cascading faults based on heterogeneous graph neural networks under ice storms. Background Technology

[0002] In frigid regions, power grids are prone to problems such as icing of transmission lines, line breaks, and reduced power output of wind turbines under ice storm conditions such as freezing rain, icing, and strong winds. The successive shutdown of multiple lines within a short period can trigger large-scale power flow shifts and cascading faults, potentially leading to massive power outages. Therefore, in ice storm scenarios, rapidly and accurately identifying critical lines in the power grid that significantly impact the spread of cascading faults is crucial for early warning, developing protective strategies, and preventing fault propagation.

[0003] Existing critical path identification methods mainly include topology analysis methods based on complex networks, pattern search methods based on cascading failure simulation, and data-driven methods based on graph neural networks. Topology analysis methods have high computational efficiency, but they are difficult to reflect power flow constraints, ice storm environmental disturbances, and power transfer processes after a fault. Simulation methods can more fully characterize the fault evolution mechanism, but they require the generation of a large number of high-fidelity fault chain samples, resulting in a heavy online computational burden. Although traditional graph neural network methods have good inference speed, they are usually based on aggregation of isomorphic graphs or local neighborhoods of physical topology, making it difficult to simultaneously distinguish the heterogeneous attributes of buses and lines, and also difficult to capture the non-local jump propagation relationship of cascading faults between physically non-adjacent lines.

[0004] Existing methods for power system fault prediction, transient stability assessment, or control based on graph neural networks typically abstract electrical equipment, buses, or lines as graph nodes, construct adjacency matrices based on physical connectivity, and then extract features using graph convolution, graph attention, GraphSAGE, or similar graph neural networks. These methods demonstrate that graph neural networks can be effective tools for processing power grid topology data. However, their core focus is usually on using existing physical topology for state prediction or control assessment. They do not construct inter-line influence graphs to address the intergenerational propagation relationships of accident chains under ice storm conditions, nor do they unify physical connectivity relationships and non-local propagation relationships obtained from accident chain statistics into a heterogeneous graph oriented towards critical line identification.

[0005] Meanwhile, existing heterogeneous graph neural network methods are mostly geared towards tasks such as general heterogeneous network representation learning, meta-path aggregation, cooperative relationship prediction, or node classification. Their innovations typically focus on the network structure, meta-path selection, or neighbor aggregation mechanism itself. For identifying critical lines in power grid cascading faults under ice storms, simply applying a general heterogeneous graph neural network model is insufficient to express the coupling mechanism between transmission line icing, wind turbine output degradation, intergenerational fault propagation, load shedding consequences, and cross-regional propagation causal relationships between lines.

[0006] Therefore, there is an urgent need for a critical line identification method that can unify the environmental impact of ice storms, the statistical results of cascading fault accident chains, and the physical topology of the power grid. This method would allow graph neural networks to be used only as inference tools on specific heterogeneous graphs, while focusing the technical emphasis on generating ice storm cascading fault samples, constructing fault propagation impact graphs, fusing physical topology and propagation topology, and fusing dual-channel features. This would improve the accuracy of critical line identification under ice storm cascading faults and the efficiency of online applications. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for identifying critical lines of power grid cascading faults based on heterogeneous graph neural networks under ice storms. This method addresses the technical problem that existing critical line identification methods based on physical topology or general graph neural networks are unable to simultaneously characterize the impact of ice storms, the heterogeneous attributes of buses and lines, local constraints of physical topology, and the non-local propagation characteristics of cascading faults.

[0008] To address the aforementioned technical problems, this invention provides a method for identifying critical lines experiencing cascading faults in power grids under ice storm conditions based on heterogeneous graph neural networks, comprising the following steps, performed sequentially: Step 1: Obtain information on power grid components, topology, operating status, and ice storm forecasts; generate ice storm cascading failure samples based on the ice storm impact model and sequential Monte Carlo simulation. Step 2: Based on the ice storm cascading fault samples obtained in Step 1, perform fault generation classification, calculate line fault reproduction rate and inter-line fault propagation rate, and construct a cascading fault impact map; Step 3: Integrate the physical topology and cascading fault impact diagram to construct a heterogeneous power grid cascading fault diagram that includes bus nodes, line nodes, physical connection edges, and impact propagation edges; Step 4: Construct bus node features and icing-state line node features, and map different types of node features to a unified implicit space; Step 5: Extract local electrical constraint features through physical channels and non-local fault propagation features through influence channels to construct a critical path identification model; Step Six: Employ an adaptive gating mechanism to fuse local electrical constraint characteristics and non-local fault propagation characteristics; Step 7: Output the predicted probability of the critical line and obtain the identification results of the critical line for power grid cascading failures under ice storm.

[0009] The ice disaster impact model described in step one includes an icing growth model, a line fault probability model under ice disaster, and a wind turbine power correction model.

[0010] The ice disaster cascading fault sample in step one includes updating the ice thickness of transmission lines, the ice thickness of wind turbine blades, the line fault probability, and the wind turbine output power in time periods. The power grid topology and system operating status are inherited and updated between adjacent time periods. New fault lines are sampled according to the updated line fault probability. The cascading fault evolution analysis is performed on the subsequent line faults triggered by the new fault lines. The new fault lines and the subsequent line faults they trigger are recorded in adjacent generations to form a parent-child correspondence for determining the directed propagation probability.

[0011] The method for constructing the cascading fault impact map in step two is as follows: based on the ice disaster cascading fault samples, the faulty lines are divided into generations, the propagation relationship between the parent faulty lines and the child faulty lines is statistically analyzed, the line fault reproduction rate and the inter-line fault propagation rate are calculated, and a cascading fault impact map is constructed based on the line fault reproduction rate and the inter-line fault propagation rate to characterize the non-local fault propagation relationship between lines.

[0012] The characteristics of the bus node described in step four are as follows: , In the formula: V i and θ i bus nodes i Voltage amplitude and phase angle; and busbars i Injected active and reactive power; The characteristics of the line node in the icing state are as follows: , In the formula: and These are the active and reactive power of the line, respectively. d j This refers to the thickness of the ice layer. u j The line connection status is indicated by 0 (fault disconnection) and 1 (normal connection).

[0013] The local electrical constraint features described in step five are as follows: , In the formula: N phys (i ) for the line i A set of connected busbars The physical channel characteristics obtained through aggregation. W phys For the physical channel learnable weight matrix, The ReLU activation function is used, and MEAN is the mean aggregation operator. e k For nodes k The initial embedding vector; The nonlocal fault propagation characteristics are as follows: , In the formula: N infl ( i ) for the line i A set of neighboring nodes of a route that has a high-risk propagation association; w ji To affect the lines in the diagram j For the line i The prior propagation probability; The learnable weight matrix that influences the channel; e j Neighboring nodes j The initial embedding vector.

[0014] The fusion coefficients of the adaptive gating mechanism described in step six are: , In the formula: || represents feature splicing; z i This is the gating fusion coefficient vector, with a value range of (0,1), used to dynamically adjust the weights of the two types of features; W z and b z These are the learnable weight matrix and bias vector of the gated unit, respectively; The characteristics of line node fusion are: , In the formula: ⊙ represents the Hadamard product, i.e., element-wise multiplication.

[0015] The method for obtaining the predicted probability of the critical path in step seven is as follows: input the node fusion features into a classifier to obtain the path. i The predicted probability of the critical path: , In the formula: W cls and b clsThese are the weight matrix and bias parameters of the classifier, respectively.

[0016] The beneficial effects of this invention are: This invention generates accident chain samples through ice disaster impact model and sequential Monte Carlo simulation, enabling the critical line identification process to reflect the coupling effect of transmission line icing, wind turbine output degradation and cascading failure evolution. This invention constructs a chain failure impact diagram based on the intergenerational propagation relationship of faults, enabling physically non-adjacent lines that have a propagation causal relationship to directly exchange information, breaking through the limitation of relying solely on local aggregation of physical topology; This invention models bus nodes, line nodes, physical connection edges, and influence propagation edges into a heterogeneous graph oriented towards ice storm cascading faults, instead of directly applying a general graph neural network model, thereby enabling the differentiation of power grid component types, physical connection effects, and fault propagation effects. This invention uses the physical channel corresponding to the physical connection edge and the influence channel corresponding to the influence propagation edge to extract features in parallel, and fuses them through an adaptive gating mechanism, which can dynamically coordinate the contribution of physical constraint information and fault propagation information according to different ice disaster operation scenarios; This invention employs an offline training and online identification approach, bringing the high-overhead simulation and training processes forward. In the online stage, only feature extraction and model inference need to be completed, making it suitable for ice disaster early warning and protection decision-making scenarios. Attached Figure Description

[0017] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0018] Figure 1 This is a flowchart of the power grid cascading fault sample generation process based on sequential Monte Carlo under ice storm conditions according to the present invention.

[0019] Figure 2 This is a schematic diagram of the critical path identification model based on heterogeneous graph neural networks of the present invention.

[0020] Figure 3 This is a schematic diagram of the offline training and online identification process of the present invention.

[0021] Figure 4 This is the topology diagram for the actual calculation example.

[0022] Figure 5 This is a flowchart of the specific timing simulation.

[0023] Figure 6 The graph shows the evolution of system load margin under different protection schemes in a real-world example. Detailed Implementation

[0024] The following will be based on embodiments of the present invention. Figures 1-6The technical solutions in the embodiments of the present invention are clearly and completely described herein. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0025] This invention provides a method for identifying critical lines experiencing cascading faults in power grids under ice storm conditions based on heterogeneous graph neural networks, comprising: 101: Generate ice storm cascading failure samples: The system acquires the component parameters, bus and line topology connections, basic operating mode, load level, wind power output data, and ice storm forecast information for the power grid to be identified. The ice storm forecast information includes wind speed, atmospheric liquid water content, freezing rainfall, ice storm duration, ice storm movement direction, and spatial impact range. Based on the ice storm forecast information, the system updates the icing thickness of transmission lines and wind turbine blades periodically, and calculates the probability of line failure and the correction value for wind turbine output power.

[0026] In each simulation period, the sequential Monte Carlo method is used to sample the line fault status of the current period. When a new faulty line is added, the new faulty line is used as the initial disturbance to perform cascading fault evolution analysis, update the power grid topology and system operating status, and record the new faulty line, subsequent cascading faulty lines, system load shedding losses, and the corresponding operating status at that time. When no new faulty line is added, the next simulation period begins. The above process is repeated until the set total simulation period is reached, obtaining a complete fault chain sample.

[0027] 102: Construct a cascading failure impact diagram: Based on the ice storm cascading fault samples, the faulty lines are divided into generations, the propagation relationship between the parent faulty lines and the child faulty lines is statistically analyzed, the line fault reproduction rate and the inter-line fault propagation rate are calculated, and a cascading fault influence diagram is constructed based on the line fault reproduction rate and the inter-line fault propagation rate to characterize the non-local fault propagation relationship between lines.

[0028] The initial fault line in the accident chain is defined as the generation 0 fault, and subsequent fault lines directly caused by the previous generation fault are defined as the next generation faults. For multiple ice disaster cascading fault samples, statistical analysis is performed on the lines... i The number of child faults triggered when a parent fault occurs in each generation, and the line fault count is calculated. i Failure reproduction rate λ i :

[0029] In the formula: M The maximum number of generations in the cascading failure evolution; Pi,m For the lines in the statistical sample i In the m The total number of times the term appears; C i,m For the lines in the statistical sample i The number of offspring failures resulting from parent failures in the corresponding generation.

[0030] Statistical routes i Triggering line j The frequency of faults, and calculation of line faults. i To the line j Fault propagation rate p ij :

[0031] In the formula: N Indicates the total number of lines; and The lines in the statistical sample are respectively i Fault Triggering Circuit j and lines k Frequency of failures. Based on p ij Constructing a cascading failure impact diagram ,in, V L For the set of line nodes, E phys For the set of physically connected edges, E infl To influence the propagation edge set, w Let be the set of edge weights. The weights of edges affecting propagation and the set of edges affecting propagation are determined by the following formula:

[0032]

[0033] In the formula, ε is the threshold that affects the weight of the propagation edge. l i For line nodes i , l j For line nodes j This threshold filters out low-probability propagation edges while retaining the inter-line propagation relationships that significantly contribute to the evolution of ice storm cascading failures.

[0034] 103: Constructing a heterogeneous diagram of cascading faults in the power grid: By fusing the physical topology of the power grid to be identified with the cascading fault impact diagram, a heterogeneous diagram of power grid cascading faults is constructed. This heterogeneous graph includes two types of nodes: bus nodes and line nodes, and two types of edges: physical connection edges and influence propagation edges. Physical connection edges represent the endpoint connection relationships between buses and lines, while influence propagation edges represent the non-local fault propagation relationships between lines obtained from fault chain statistics.

[0035]

[0036]

[0037]

[0038] In the formula: V B For the set of bus nodes, V L For the set of line nodes, E phys For the set of physically connected edges, E infl To influence the propagation edge set, It is a set of heterogeneous graph nodes, consisting of bus nodes of the physical power grid and line nodes of the influence graph; It is a set of edges in a heterogeneous graph, consisting of physical connection edges and fault propagation edges.

[0039] 104: Construct node features and perform feature encoding: For bus nodes, bus node features are constructed, including bus voltage amplitude, bus voltage phase angle, bus injected active power, and bus injected reactive power. For line nodes, line node features are constructed, including line active power, line reactive power, line connectivity, and line icing thickness. The bus node features and line node features are represented as follows:

[0040]

[0041] In the formula: V i and θ i bus nodes i Voltage amplitude and phase angle; and busbars i Injected active and reactive power, and These are the active and reactive power of the line, respectively. d j This refers to the thickness of the ice layer. u jThis represents the connected state of the network. After standardizing continuous features, different types of nodes are mapped to a unified latent space using a type-specific encoder.

[0042] 105: Extract dual-channel features and perform adaptive fusion: The heterogeneous graph of cascading faults in the power grid is input into the critical line identification model. This model includes physical channels and influence channels. The physical channels aggregate the features of bus neighbor nodes connected to the target line node based on the physical connection edges to obtain local electrical constraint features; the influence channels aggregate the features of line neighbor nodes that have propagation associations with the target line node based on the influence propagation edge weights to obtain non-local fault propagation features.

[0043]

[0044]

[0045] In the formula: N phys ( i ) for the line i A set of connected busbars The physical channel characteristics obtained through aggregation. The influence channel features obtained by aggregation, N infl ( i ) for the line i A set of neighboring nodes of a route that has a high-risk propagation association; w ji To affect the lines in the diagram j For the line i The prior propagation probability; The learnable weight matrix that influences the channel; e j Neighboring nodes j The initial embedding vector.

[0046] Local electrical constraint characteristics and non-local fault propagation characteristics are input into the adaptive gating fusion unit to obtain the gating fusion coefficients. z i Features of line node integration :

[0047]

[0048] In the formula, || represents feature splicing. z i This is the gating fusion coefficient vector, with a value range of (0,1), used to dynamically adjust the weights of the two types of features;W z and b z , respectively, are the learnable weight matrix and bias vector of the gated unit; ⊙ represents the Hadamard product, and represents element-wise multiplication.

[0049] 106: Output the critical path identification results and train the model: By inputting the fused features of the line nodes into the classifier, the predicted probability that line i is a critical line is obtained:

[0050] During the offline training phase, key line indicators are constructed based on line fault probability, expected load shedding when the line is the initial fault, and causal gain factor. R i :

[0051] In the formula: P i For the line in the operating scenario i The probability of failure; D i For the line i This represents the expected load shedding amount at the initial fault. λ max This represents the line's maximum fault reproduction rate. γ i This is the causal gain factor, with values ​​in the range [1, 2]. Model training uses a weighted binary cross-entropy loss function:

[0052] In the formula: N This refers to the batch sample size. y i and The lines are respectively i The true label and the predicted probability; w pos These are the weight coefficients for positive samples. During the online identification phase, the node features and heterogeneous graphs corresponding to the new ice disaster operation scenario are input into the trained model, which outputs the critical probabilities of each route and the ranking results of the critical routes. Example

[0053] This embodiment verifies the identification accuracy and computational efficiency of the method of the present invention based on a modified IEEE 118-node system. To ensure that the simulation environment can reflect the operating conditions of a high-proportion renewable energy grid in frigid regions, some synchronous generator units in the IEEE 118-node system are replaced with wind power clusters of the same capacity, achieving a wind power penetration rate of 25%. The case study verification includes the following steps: 201: Setting ice storm simulation parameters and training parameters: Wind speed, icing, sample generation, and model training parameters are set according to Table 1.

[0054] Table 1 Parameter Settings

[0055] 202: Conduct model comparison tests: Comparative tests were conducted using GCN, GAT, GraphSAGE, and the critical path identification model described in this invention. The test results are shown in Table 2. Table 2 shows that compared to GCN and GraphSAGE, which rely on local aggregation based on physical topology, and GAT, which uses isomorphic graph attention mechanisms, the proposed HetGNN model improves critical path identification by mining the neighborhood physical associations and cross-regional influence associations of faulty paths. All indicators are superior to the comparative models. Furthermore, since non-critical paths constitute the vast majority, the accuracy of each model is generally high; therefore, the F1 score is the primary focus here.

[0056] Table 2 Comparison of Critical Path Identification Results of Different Models

[0057] 203: Perform ablation test: To verify the roles of physical channels, influence channels, and gating fusion mechanisms, ablation models were set up with influence channels removed, physical channels removed, and gating fusion mechanisms removed. The test results are shown in Table 3. Table 3 shows that the complete HetGNN model performs best. After removing influence channels, the degraded model performs similarly to the standard GCN, with the F1 score decreasing from 95.95% to 86.07%, indicating that the non-local propagation information of faults contained in the influence graph can improve the accuracy of critical path identification. After removing physical channels and gating mechanisms, the F1 scores decreased to 90.27% and 91.88%, respectively, indicating that physical topology information is also not negligible, and that gating fusion can balance the contributions of the two types of information, improving the model's identification ability.

[0058] Table 3 Comparison of model identification performance under different ablation structures

[0059] 204: Verify physical validity and computational efficiency: Under typical ice storm conditions with an average wind speed of 12 m / s, precipitation of 2.4 mm / h, and uniform icing across the entire network, the method of this invention is compared with risk theory, FP-growth, PageRank, and GAT. Cumulative attack tests were conducted on the Top-5 critical lines identified by each method. The results of physical validity verification and computational efficiency comparisons are shown in Tables 4 and 5.

[0060] Table 4 shows that there is some overlap and significant differences in the critical paths identified by different methods. Specifically, paths 36 and 44 were identified by multiple methods, indicating strong common risk characteristics. However, the identification results of different methods still differ, mainly due to their different focuses in characterizing the criticality of the paths: Risk Theory focuses more on the probability of path failure and its direct consequences of load loss; FP-growth emphasizes the frequent co-occurrence relationships in the accident chain; PageRank emphasizes the structural importance of a path in the fault propagation network determined by propagation relationships; and GAT mainly learns features based on physical topological neighborhoods.

[0061] As shown above, the method of this invention has good performance in terms of identification accuracy, physical validity of critical paths, and online calculation speed. The non-local fault propagation information contained in the influencing propagation edges can significantly improve the identification accuracy of critical paths, and the local electrical constraint information contained in the physical connection edges is also not negligible. Adaptive gating fusion can coordinate the contributions of the two types of information.

[0062] Table 4. Verification results of the physical validity of identification results from different methods

[0063] Table 5 Comparison of Critical Path Identification Results Using Different Methods

[0064] Example 3 This embodiment verifies the engineering applicability and dynamic identification capability of the method of the present invention based on an actual provincial power grid in Northeast my country. This actual power grid example includes a 500kV backbone grid and a 220kV regional power grid, with a total of 220 bus nodes, 251 transmission lines, and 103 generator units. The topology is as follows: Figure 4 As shown. Engineering application verification includes the following steps: 301: Constructing a mobile ice storm scenario: A 48-hour moving ice storm was constructed with the following parameters: movement speed 30 km / h, influence radius 60 km, precipitation rate 1.5 mm / h, and spatial attenuation coefficient 10000. Forty-eight time-series scenarios were built with a 1-hour time step, and the total network ice thickness, system state, and topology state were input into the trained critical path identification model.

[0065] 302: Identify critical paths at different time periods: The scenarios corresponding to the 1st hour, 16th hour, and 32nd hour were selected as typical scenarios in the early, middle, and late stages of the ice storm, respectively, for critical path identification and protection scheme comparison. The results are shown in Table 6. Figure 6As shown. To verify the application effect of dynamic identification results, three protection schemes were set up for comparison: Scheme 1 is a random protection scheme, which selects the same number of random lines as Scheme 3 for protection in each stage; Scheme 2 is a delayed protection scheme, which performs random protection in the early stage of the ice storm, implements protection based on the early identification results in the middle stage, and updates the protection objects based on the middle identification results in the later stage of the ice storm; Scheme 3 is a real-time dynamic protection scheme, which implements protection based on the critical line identification results in the early, middle and late stages of the ice storm. Each scheme adopts state inheritance and rolling update, taking the system state after protection in the previous stage as the initial state of the next stage, and re-identifying and simulating. The timing simulation flowchart is shown in [link to flowchart]. Figure 5 .

[0066] The timing simulation steps for the actual power grid example are as follows: 1) Input the power grid topology, line parameters, initial operating data of wind turbines and loads, and ice disaster information such as wind speed, air liquid water content, and precipitation rate within 48 given time periods. Initialize the line connectivity status, wind turbine output, load status, and system operating status, and set t = 1. 2) In time period t, inherit the line connectivity status, load level and icing thickness formed in the previous time period, and calculate and update the line icing status, add icing disconnected lines and wind turbine output according to the current meteorological information to form the initial system status of the current time period; 3) If no protection plan is implemented, the initial system state of the current period is used as the basis to simulate the cascading failure evolution process under the ice disaster, referring to the method in Section 1.1; then the feature vectors of the bus node and the feature vectors of the line node are constructed and input into the trained HetGNN model to obtain the critical line identification results for the t-th period. 4) If a real-time dynamic protection scheme (Scheme 3) is implemented, the evolution process of ice storm cascading failures under unprotected conditions is first simulated based on the initial system state of the current stage, and the critical path is identified based on the prediction results. This unprotected prediction is only used to determine the protection targets for the current stage and is not used as the basis for state inheritance in the next period. Subsequently, the system returns to the initial system state of the current stage, protection is implemented on the identified critical paths, and the evolution process of ice storm cascading failures after protection is simulated again under the same ice storm conditions. Then, the critical path is identified based on the simulation results after protection, and the formal identification results and system operating status of the current stage are obtained.

[0067] 5) If a delayed protection scheme (Scheme 2) is implemented, the protection targets will be determined using a random protection method in the early stage; in the middle and late stages, the protection targets for the current stage will be determined using the critical path identification results obtained in the previous stage. After implementing protection under the initial system state of the current stage, the evolution process of ice storm cascading failures will be simulated, and the critical path identification for this stage will be carried out; the system state after simulation will be inherited to the next stage, and the identification results for this stage will be used for delayed protection in the next stage.

[0068] 6) If a random protection scheme (Scheme 1) is implemented, its state inheritance and cascading failure deduction process is consistent with the delayed protection scheme. The difference is that the protection objects at each stage are not determined by the critical path identification results, but are randomly selected from the paths that are still in a connected state, and the number of selections is consistent with other protection schemes.

[0069] 7) Record the critical path identification results, cascading fault evolution results, line connectivity status, load shedding results, turbine output and load remaining rate, etc., for the current period, and inherit the system status such as line connectivity status, turbine operation status and load shedding results formed after the formal simulation of the current period to the next period.

[0070] 8) If 48 hours have not been reached, set t = t + 1 and return to step 2); otherwise, output the critical path identification results, cascading fault evolution results, and load margin change results for the 48-hour continuous period.

[0071] The identification results of the Top-10 critical routes during different ice storm stages are shown in Table 6. The results indicate that the set of critical routes dynamically shifts with the movement of ice storms, accumulation of ice on routes, and redistribution of tidal currents.

[0072] Table 6. Identification Results of Top-10 Critical Routes at Different Stages of Ice Storms

[0073] The system load margin evolution curves under different protection schemes are shown below. Figure 6 . Figure 6 The results show that as the ice storm develops and cascading failures evolve, the system load margin gradually decreases under all scenarios. Among them, Scenario 1 shows the most severe decrease; Scenario 2 is generally better than random protection, indicating that the identified critical paths still have risk persistence in adjacent stages, but its protection effect is weaker than Scenario 3, indicating that the set of critical paths will dynamically migrate as the ice storm evolves; Scenario 3 maintains the highest load margin in all stages, indicating that the dynamic identification results can better capture the migration characteristics of the set of critical paths during the evolution of the ice storm, and the identified critical paths have a higher degree of influence on the propagation of cascading failures.

[0074] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0075] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for identifying critical lines in power grid cascading faults based on heterogeneous graph neural networks under ice storm conditions, characterized in that, Includes the following steps, And the following steps are performed in sequence: Step 1: Obtain information on power grid components, topology, operating status, and ice storm forecasts; generate ice storm cascading failure samples based on the ice storm impact model and sequential Monte Carlo simulation. Step 2: Based on the ice storm cascading fault samples obtained in Step 1, perform fault generation classification, calculate line fault reproduction rate and inter-line fault propagation rate, and construct a cascading fault impact map; Step 3: Integrate the physical topology and cascading fault impact diagram to construct a heterogeneous power grid cascading fault diagram that includes bus nodes, line nodes, physical connection edges, and impact propagation edges; Step 4: Construct bus node features and icing-state line node features, and map different types of node features to a unified implicit space; Step 5: Extract local electrical constraint features through physical channels and non-local fault propagation features through influence channels to construct a critical path identification model; Step Six: Employ an adaptive gating mechanism to fuse local electrical constraint characteristics and non-local fault propagation characteristics; Step 7: Output the predicted probability of the critical line and obtain the identification results of the critical line for power grid cascading failures under ice storm.

2. The method for identifying critical lines of cascading faults in power grids based on heterogeneous graph neural networks under ice storm conditions, as described in claim 1, is characterized in that... The ice disaster impact model described in step one includes an icing growth model, a line fault probability model under ice disaster, and a wind turbine power correction model.

3. The method for identifying critical lines of power grid cascading faults based on heterogeneous graph neural networks under ice storm conditions, as described in claim 1, is characterized in that... The ice disaster cascading fault sample in step one includes updating the ice thickness of transmission lines, the ice thickness of wind turbine blades, the line fault probability, and the wind turbine output power in time periods. The power grid topology and system operating status are inherited and updated between adjacent time periods. New fault lines are sampled according to the updated line fault probability. The cascading fault evolution analysis is performed on the subsequent line faults triggered by the new fault lines. The new fault lines and the subsequent line faults they trigger are recorded in adjacent generations to form a parent-child correspondence for determining the directed propagation probability.

4. The method for identifying critical lines of cascading faults in power grids based on heterogeneous graph neural networks under ice storm conditions, as described in claim 1, is characterized in that... The method for constructing the cascading fault impact map in step two is as follows: based on the ice disaster cascading fault samples, the faulty lines are divided into generations, the propagation relationship between the parent faulty lines and the child faulty lines is statistically analyzed, the line fault reproduction rate and the inter-line fault propagation rate are calculated, and a cascading fault impact map is constructed based on the line fault reproduction rate and the inter-line fault propagation rate to characterize the non-local fault propagation relationship between lines.

5. The method for identifying critical lines of power grid cascading faults based on heterogeneous graph neural networks under ice storms, as described in claim 1, is characterized in that... The characteristics of the bus node described in step four are as follows: , In the formula: V i and θ i bus nodes i Voltage amplitude and phase angle; and busbars i Injected active and reactive power; The characteristics of the line node in the icing state are as follows: , In the formula: and These are the active and reactive power of the line, respectively. d j This refers to the thickness of the ice layer. u j The line connection status is indicated by 0 (fault disconnection) and 1 (normal connection).

6. The method for identifying critical lines of power grid cascading faults based on heterogeneous graph neural networks under ice storm conditions according to claim 1, characterized in that, The local electrical constraint features described in step five are as follows: , In the formula: N phys ( i ) for the line i A set of connected busbars The physical channel characteristics obtained through aggregation. W phys For the physical channel learnable weight matrix, The ReLU activation function is used, and MEAN is the mean aggregation operator. e k For nodes k The initial embedding vector; The nonlocal fault propagation characteristics are as follows: , In the formula: N infl ( i ) for the line i A set of neighboring nodes of a route that has a high-risk propagation association; w ji To affect the lines in the diagram j For the line i The prior propagation probability; The learnable weight matrix that influences the channel; e j Neighboring nodes j The initial embedding vector.

7. A method for identifying critical lines of cascading faults in power grids based on heterogeneous graph neural networks under ice storm conditions, as described in claim 1 or 6, characterized in that: The fusion coefficients of the adaptive gating mechanism described in step six are: , In the formula: || represents feature splicing; z i This is the gating fusion coefficient vector, with a value range of (0,1), used to dynamically adjust the weights of the two types of features; W z and b z These are the learnable weight matrix and bias vector of the gated unit, respectively; The characteristics of line node fusion are: , In the formula: ⊙ represents the Hadamard product, i.e., element-wise multiplication.

8. The method for identifying critical lines of power grid cascading faults based on heterogeneous graph neural networks under ice storms, as described in claim 1, is characterized in that... The method for obtaining the predicted probability of the critical path in step seven is as follows: input the node fusion features into a classifier to obtain the path. i The predicted probability of the critical path: , In the formula: W cls and b cls These are the weight matrix and bias parameters of the classifier, respectively.