Cascade failure early warning method and device for correlation infrastructure system

Abstract

This invention relates to the field of Internet of Things (IoT) technology, and particularly to a method and apparatus for early warning of cascading failures in interconnected infrastructure systems. The method includes: generating a disturbance scenario to be evaluated based on a multi-layered interdependent infrastructure network according to specified parameters; applying the disturbance to the network, performing a physical constraint cascading simulation using the post-disturbance state as initial information to obtain complete cascading trajectories for each scenario; extracting trajectory indicators to generate severity labels, constructing pre-cascading heterogeneous graph samples by combining network attributes and initial information, training a heterogeneous graph neural network to obtain a risk prediction model, and outputting the severe cascading probability and corresponding risk by inputting the scenario to be evaluated. This solves the problem of difficulty in early identification of severe cascading failures caused by multi-source disturbances in interdependent infrastructure, enabling risk assessment to determine the likelihood of the scenario to be evaluated developing into a severe cascading failure based solely on disturbance information, initial network topology, and static physical attributes before cascading propagation has fully unfolded.

Description

Technical Field

[0001] This invention relates to the field of Internet of Things (IoT) technology, and in particular to a method and apparatus for early warning of cascading failures in interconnected infrastructure systems. Background Technology

[0002] Currently, the risk assessment of cascading failures in interdependent infrastructure networks mainly adopts either the "cascading simulation and posterior risk statistics" or "cascading simulation and ordinary machine learning classification" approach. These approaches typically involve first establishing a network model and setting a single or limited number of disturbance scenarios. Then, they simulate the cascading propagation process using power flow, optimal power flow, or cascading failure rules. The severity of the failure is statistically analyzed based on indicators such as the final load loss and the number of failed components. Some approaches also utilize simulation results to train classifiers or regressors to predict the risks of similar scenarios.

[0003] However, this type of solution has the following drawbacks: First, the organization and representation of disturbance scenarios lack uniformity. Existing solutions typically analyze one type of failure or a small number of pre-defined scenarios, lacking a unified parameterized description of different failure mechanisms such as availability loss, capacity degradation, load surges, and cross-layer dependency weakening. The scope of the disturbance and the degree of damage to individual components are also easily mixed, making it difficult to distinguish between different scenarios such as "large-scale minor damage" and "small-scale severe damage." At the same time, random failures, damage to critical targets, and spatially clustered disturbances usually employ independent generation rules, resulting in insufficient scenario coverage and making it difficult to repeatedly generate and compare scenarios horizontally.

[0004] Second, some risk prediction models rely on posterior information. Their input features may include branch failure states, cascade propagation depth, final load loss, line utilization after cascading, or key observation results that can only be obtained during or after the cascading process. While this information can improve the fitting ability to existing accident results, it is usually unavailable when the disturbance has just occurred and the cascading has not yet fully propagated. Therefore, the relevant models are more suitable for post-event interpretation and difficult to use for true pre-event or early warning.

[0005] Third, traditional feature engineering has limited ability to express interdependent network structures. Directly constructing feature vectors from indicators such as node degree, line capacity, generation capacity, and load size makes it difficult to simultaneously preserve the power network topology, component physical operating attributes, natural gas-electric coupling relationships, and the propagation paths of disturbances between different types of components. In particular, when disturbances occur at lines, generators, loads, gas nodes, or cross-layer coupling edges, existing solutions lack an effective mechanism to uniformly map disturbances from different components to learning nodes while maintaining their network positional relationships.

[0006] Fourth, the efficiency of risk assessment needs improvement. Some solutions require re-executing a complete multi-round cascaded simulation every time a new disturbance scenario is assessed. As the network size, disturbance type, and number of candidate components increase, the number of scenarios that need to be assessed grows rapidly. Solving power flow or emergency dispatch models scenario by scenario incurs significant computational overhead, making it difficult to meet the needs of rapid screening of large-scale scenarios and online assessment of operational safety. Summary of the Invention

[0007] This invention provides a method and apparatus for early warning of cascading failures in interconnected infrastructure systems. It addresses the shortcomings of existing methods for expressing disturbance scenarios, which makes it difficult to simultaneously cover different failure mechanisms, disturbance ranges, damage levels, and spatial distribution patterns. Furthermore, some prediction models rely on posterior information such as final load loss, cascading depth, and subsequent component failures, making them unsuitable for pre-emptive or early warning. Traditional feature engineering struggles to uniformly express intra-layer network topology, component physical attributes, cross-layer dependencies, and disturbance information on different types of components. Performing complete cascading simulations for all candidate scenarios individually incurs significant computational overhead, failing to meet the need for rapid screening of large-scale scenarios.

[0008] A first aspect of the present invention provides a method for early warning of cascading failures in interconnected infrastructure systems, comprising the following steps: Based on a pre-built multi-layer interdependent infrastructure network model, various disturbance scenarios to be evaluated are generated according to the disturbance mechanism, target, disturbance range, disturbance amplitude, spatial pattern, and random seed. Various disturbance scenarios to be evaluated are applied to the multi-layer interdependent infrastructure network model. The system state after the disturbance is set as the initial disturbance information. Physical constraint cascade simulation is performed in discrete rounds to obtain the complete cascade propagation trajectory of each disturbance scenario. Extract the corresponding operational indicators from the complete cascading propagation trajectory of each disturbance scenario, and convert the operational indicators into scenario severity labels; Based on the multi-layer interdependent infrastructure network model, preset component static physical properties, preset cross-layer dependencies, and the initial disturbance information, construct a sample of multi-relationship heterogeneous graphs before cascading. The pre-constructed heterogeneous graph neural network is iteratively trained using the scenario severity labels and the cascaded pre-relationship heterogeneous graph samples to obtain a heterogeneous graph-trained risk prediction model. The scenario to be evaluated is input into the heterogeneous graph training risk prediction model to obtain the severe cascade probability and its corresponding risk level.

[0009] Optionally, the multi-layer interdependent infrastructure network model includes at least two interdependent infrastructure subsystems, and each infrastructure subsystem corresponds to an independent network topology, entity nodes, and connection edge set. The entity nodes correspond to the physical entities of their respective infrastructures, and each connection edge corresponds to the physical connection relationship of its respective infrastructure.

[0010] Optionally, the step of applying various disturbance scenarios to be evaluated to the multi-layer interdependent infrastructure network model, setting the system state after the disturbance as the initial disturbance information, and performing physical constraint cascade simulation in discrete rounds to obtain the complete cascade propagation trajectory of each disturbance scenario includes: Obtain the current system state of each infrastructure subsystem to construct an independent operation model for each infrastructure subsystem, and match corresponding physical balance constraints, capacity constraints, and operation boundary constraints for each independent operation model; With the goal of minimizing the weighted sum of the total system operating cost, total service loss penalty, and total constraint violation penalty, cross-subsystem unified emergency scheduling calculations are performed on the independent operating models of the infrastructure subsystems of all subsystems to obtain the operating traffic and load utilization of all components in this round. Components whose load utilization exceeds a preset failure threshold are selected and added to the candidate failure set; The newly added failed components in this round are determined from the candidate failure set, and the network topology and component status of each infrastructure subsystem are updated; Based on the preset cross-layer dependency relationships, the running status of all affected cross-layer dependency components is updated synchronously; Repeat the aforementioned process until the preset simulation termination condition is met to obtain the complete cascading propagation trajectory of each disturbance scenario.

[0011] Optionally, the step of extracting corresponding operational indicators from the complete cascading propagation trajectory of each disturbance scenario and converting the operational indicators into scenario severity labels includes: Extract the final service loss, total service demand, number of failed components, and total number of components for each infrastructure subsystem from the complete cascading propagation trajectory of each disturbance scenario. The comprehensive service loss ratio of each infrastructure subsystem is calculated based on the final service loss amount and the total service demand, and the comprehensive failure component ratio of each infrastructure subsystem is calculated based on the number of failed components and the total number of components. Based on the preset judgment rules, corresponding binary severity labels are generated according to the comprehensive service loss ratio and the comprehensive failure component ratio, and the binary severity labels are associated with the corresponding disturbance scenarios to obtain the scenario severity labels.

[0012] Optionally, the step of constructing a pre-cascade multi-relationship heterogeneous graph sample based on the multi-layer interdependent infrastructure network model, preset component static physical attributes, preset cross-layer dependencies, and the initial perturbation information includes: The entity nodes of the multi-layer interdependent infrastructure network model are treated as heterogeneous nodes; The internal topological connections and cross-layer dependencies of each infrastructure subsystem in the multi-layer interdependent infrastructure network model are used as heterogeneous graph edges. Extract node features from the static physical properties of the preset component; Extract the node perturbation features and scenario-level global features from the initial perturbation information; The heterograph nodes, heterograph edges, node features, node perturbation features, and scenario-level global features are standardized to generate the cascaded pre-multi-relation heterograph samples, and the pairing association between the heterograph samples and the corresponding severity labels is stored.

[0013] Optionally, the heterogeneous graph neural network is configured with a node feature encoding layer, a global feature encoding layer, multiple relationship-specific message passing layers, a relationship fusion layer, a graph-level pooling layer, a feature concatenation layer, and a classification output layer in sequence.

[0014] Optionally, the step of inputting the scenario to be evaluated into the heterogeneous graph training risk prediction model to obtain the severe cascading probability and its corresponding risk level includes: The scenario to be evaluated is preprocessed to generate a multi-relationship heterogeneous graph before cascading; The multi-relationship heterogeneous graph before cascading is input into the heterogeneous graph training risk prediction model to obtain the severe cascading probability of the scenario to be evaluated. Based on a preset threshold classification rule, the risk level corresponding to the severe cascading probability is obtained.

[0015] A second aspect of the present invention provides a cascading failure early warning device for interconnected infrastructure systems, comprising: The generation module is used to generate various disturbance scenarios to be evaluated based on a pre-built multi-layer interdependent infrastructure network model, according to the disturbance mechanism, target, disturbance range, disturbance amplitude, spatial pattern and random seed. The iterative simulation module is used to apply various disturbance scenarios to the multi-layer interdependent infrastructure network model, set the system state after disturbance as the initial disturbance information, and perform physical constraint cascade simulation in discrete rounds to obtain the complete cascade propagation trajectory of each disturbance scenario. The conversion module is used to extract the corresponding operational indicators from the complete cascading propagation trajectory of each disturbance scenario and convert the operational indicators into scenario severity labels. The construction module is used to construct a pre-cascade multi-relationship heterogeneous graph sample based on the multi-layer interdependent infrastructure network model, preset component static physical properties, preset cross-layer dependencies and the initial disturbance information; The iterative training module is used to iteratively train the pre-constructed heterogeneous graph neural network using the scenario severity labels and the cascaded pre-relationship heterogeneous graph samples to obtain a heterogeneous graph training risk prediction model. The risk prediction module is used to input the scenario to be evaluated into the heterogeneous graph training risk prediction model to obtain the severe cascade probability and its corresponding risk level.

[0016] A third aspect of the present invention provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the cascading failure early warning method for associated infrastructure systems as described in the above embodiments.

[0017] A fourth aspect of the present invention provides a computer program product that, when executed by a processor, implements the above-described method for early warning of cascading failures in interconnected infrastructure systems.

[0018] A fifth aspect of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for early warning of cascading failures in interconnected infrastructure systems.

[0019] The cascading failure early warning method and device for interconnected infrastructure systems proposed in this invention can uniformly describe multiple failure mechanisms and spatial distribution patterns by setting disturbance families, disturbance ranges, disturbance amplitudes, spatial patterns, and random seeds. It can distinguish between the scale of affected components and the damage level of individual components, improving the coverage, comparability, and reproducibility of disturbance scenarios. By generating supervised labels through physical constraint cascading simulation, the severity judgment simultaneously considers network balance, component operational capabilities, emergency dispatch feasibility, and cascading propagation results, resulting in training labels with clear engineering and physical meaning. Furthermore, by projecting disturbances on different types of components to a unified learning node and setting relationship channels such as same-layer topology and cross-layer dependencies, it can achieve early warning of cascading failures in the same heterogeneous system. The diagram integrates network structure, physical attributes, perturbation locations, and cross-layer coupling information to enhance the ability to express interdependent structures and potential propagation conditions. By isolating the label generation variables from the model input, it explicitly excludes information such as service loss, propagation depth, and subsequent component failures that can only be obtained after cascading propagation. This enables the model to implement risk warnings based on information available when a perturbation occurs or in the early stages of cascading, avoiding label leakage due to posterior information entering the input. After the model completes training, it mainly performs feature construction and forward computation for new scenarios, eliminating the need for complete multi-round cascading simulations of all candidate scenarios during the initial screening stage. This improves the efficiency of risk screening for large-scale scenarios and prioritizes the allocation of refined simulation and emergency analysis resources to high-risk scenarios.

[0020] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0021] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 A flowchart illustrating an early warning method for cascading failures in a connected infrastructure system, provided according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating an early warning system for cascaded failure based on physical simulation and heterogeneous graph learning, according to an embodiment of the present invention. Figure 3 This is a schematic diagram illustrating the overall execution of an early warning method for cascading failures in a connected infrastructure system, provided by an embodiment of the present invention. Figure 4 A flowchart illustrating the hierarchical structure and connection relationships of a heterogeneous graph neural network according to an embodiment of the present invention; Figure 5 This is a block diagram of a cascading failure early warning device for associated infrastructure systems according to an embodiment of the present invention; Figure 6 This is a schematic diagram of the structure of an electronic device according to an embodiment of the present invention.

[0022] Explanation of reference numerals in the attached figures: 50- Cascaded failure early warning device for related infrastructure systems, 501- Generation module, 502- Iterative simulation module, 503- Conversion module, 504- Construction module, 505- Iterative training module, 506- Risk prediction module, 601- Memory, 602- Processor, 603- Communication interface. Detailed Implementation

[0023] Embodiments of the present invention are described in detail below. Examples of these embodiments are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0024] The following description, with reference to the accompanying drawings, illustrates an embodiment of the cascading failure early warning method and apparatus for interconnected infrastructure systems according to the present invention. Addressing the issues mentioned in the background section regarding the lack of a unified representation of existing disturbance scenarios, making it difficult to simultaneously cover different fault mechanisms, disturbance ranges, damage levels, and spatial distribution patterns; the reliance of some prediction models on posterior information such as final load loss, cascading depth, and subsequent component failures, which makes them unsuitable for pre-emptive or early warning; the difficulty of using traditional feature engineering to uniformly represent intra-layer network topology, component physical attributes, cross-layer dependencies, and disturbance information on different types of components; and the high computational cost of performing complete cascading simulations on all candidate scenarios, which is insufficient to meet the needs of rapid screening for large-scale scenarios, the present invention provides a cascading failure early warning method for interconnected infrastructure systems. In this method, disturbance scenarios are uniformly parameterized according to disturbance families, disturbance ranges, disturbance amplitudes, spatial patterns, and random seeds to generate disturbance scenarios covering fault mechanisms such as availability loss, capacity degradation, load surges, and weakened cross-layer dependencies. Secondly, disturbances are applied to the interdependent infrastructure network. Through cascaded simulations that satisfy network balance, component operational capacity, and security constraints, results such as emergency dispatch feasibility, service loss ratio, cascade propagation depth, and component failure ratio are obtained, and severity supervision labels are generated accordingly. Then, the network topology, static physical properties, and cross-layer dependencies are constructed into a multi-relationship graph structure, and disturbances on components such as lines, generators, loads, gas nodes, and cross-layer coupling edges are projected onto the corresponding learning nodes to form a heterogeneous graph containing only information available before cascade propagation. Posterior variables such as final service loss, cascade depth, subsequent component failures, and simulation termination reasons used to generate labels are not included in the model input. Finally, a risk prediction model is trained using the heterogeneous graph with severity labels. For new disturbance scenarios, heterogeneous graphs are generated according to entity mapping, feature construction, and standardization rules consistent with the training phase. The model forward computation outputs the severity cascade probability or risk level, and high-risk scenarios are prioritized for submission to fine-grained cascade simulation or emergency response analysis.

[0025] In the offline phase, firstly, interdependent infrastructure network data is acquired and parameterized disturbance scenarios are generated. Then, on one hand, severity supervision labels are generated through physical constraint cascade simulation; on the other hand, a pre-cascade heterogeneous graph is constructed using network topology, static physical properties, cross-layer dependencies, and disturbance information. Finally, a risk prediction model is trained using the heterogeneous graph and severity labels. In the risk prediction phase, according to the entity mapping, feature order, standardized parameters, and relational edge construction rules saved during the training phase, the network state and disturbance scenarios to be evaluated are converted into a pre-cascade heterogeneous graph, and the trained model is called to output the severity cascade probability and risk level. The initial risk screening phase does not require repeating complete multi-round cascade simulations for each new scenario; for high-risk scenarios, further refined cascade simulations or emergency response analyses can be conducted. Figure 2As shown, to further illustrate the module connections and data transfer relationships between offline training and risk prediction, each module transfers data through scenario identifiers, network entity identifiers, and unified data fields. The trained model parameters, risk thresholds, entity mappings, feature order, standardized parameters, and relational edge construction rules generated in the offline phase are transferred to the risk prediction phase to ensure that the scenario to be evaluated and the training samples use a consistent heterogeneous graph construction method. Posterior variables such as emergency dispatch feasibility, final service loss, cascading propagation depth, subsequent component failures, and simulation termination reasons are only used for offline severity label generation and do not enter the pre-cascading heterogeneous graph and risk prediction process. This achieves the isolation of label generation information from prediction input information, ensuring that the model makes risk judgments based on information obtainable before cascading propagation.

[0026] Specifically, Figure 1 A flowchart illustrating an early warning method for cascading failures in a connected infrastructure system, provided by an embodiment of the present invention.

[0027] like Figure 1 As shown, this method for early warning of cascading failures in interconnected infrastructure systems includes the following steps: In step S101, based on the pre-built multi-layer interdependent infrastructure network model, various disturbance scenarios to be evaluated are generated according to the disturbance mechanism, the target of the disturbance, the disturbance range, the disturbance amplitude, the spatial pattern, and the random seed.

[0028] In some embodiments, the multi-layer interdependent infrastructure network model includes at least two interdependent infrastructure subsystems, and each infrastructure subsystem corresponds to an independent network topology, entity nodes, and connection edge set. Entity nodes correspond to the physical entities of their respective infrastructures, and each connection edge corresponds to the physical connection relationship of its respective infrastructure.

[0029] In actual implementation, such as Figure 3 As shown, the network topology, component attributes, operating status, physical constraints, and cross-layer dependencies of each infrastructure subsystem within the area to be evaluated are obtained. A multi-layer interdependent infrastructure network model is then constructed based on entity type, connection relationships, and dependency directions. Assume the interdependent infrastructure system includes... Each infrastructure subsystem has a multi-layered interdependent infrastructure network model. It can be represented as:

[0030] in, For the first A set of nodes for an infrastructure subsystem For the first A set of connection edges within an infrastructure subsystem For the first Subsystem and the first The set of cross-layer dependency edges between subsystems A collection of static properties for nodes and components. This is a set of system operation constraints and component capacity constraints.

[0031] Infrastructure subsystems can be power, natural gas, water supply, transportation, communication, heating, drainage, or other infrastructure systems with physical flows, energy supply, information control, service support, or spatial relationships. Nodes and edges in different infrastructure subsystems can have different meanings. For example, in a power system, a node can be a busbar or substation, and an edge can be a transmission line; in a natural gas system, a node can be a gas source, gas station, or demand node, and an edge can be a natural gas pipeline; in a water supply system, a node can be a water plant, pumping station, water tank, or water user node, and an edge can be a water pipeline; in a transportation system, a node can be an intersection, station, or transportation hub, and an edge can be a road or rail section; in a communication system, a node can be a base station, switching equipment, or control center, and an edge can be a wired or wireless communication link.

[0032] For each type of infrastructure component, at least the component identifier, component type, subsystem it belongs to, connection endpoint, capacity or service capability, operating status, spatial location, and other parameters related to physical operation should be recorded. For cross-layer dependencies, at least the source component, the dependent component, the direction of dependency, the type of dependency, the strength of dependency, and the effective service capability should be recorded. Cross-layer dependencies can include energy supply dependencies, physical flow dependencies, information control dependencies, service support dependencies, and spatial proximity dependencies. For example, a natural gas node can provide fuel to a gas generator, an electric power node can provide electricity to a pumping station or communication base station, a communication system can provide monitoring and control services to the power system, and a transportation system can provide access for infrastructure repair.

[0033] When the service capacity of the dependent source component decreases, the availability, capacity, efficiency, or maximum service capacity of the dependent component is updated according to a pre-defined or calculated dependency transitive function. The subsystem to the first The transitive dependency relationship of each subsystem can be represented as:

[0034] in, For dependent source components The availability status or service capabilities, For dependent components The availability status or service capabilities, For strength dependence, For cross-layer dependency transfer functions, they can take the form of proportional reduction, threshold triggering, piecewise functions, logical rules, capacity mapping, distance decay, or other functions that can describe cross-layer effects.

[0035] This process allows for the preservation of connectivity relationships, component operational capabilities, and cross-layer dependencies within each infrastructure subsystem within the same data model, providing a unified and computable network foundation for subsequent simulations of the propagation of disturbances within a single subsystem and between different subsystems.

[0036] Furthermore, based on the established interdependent infrastructure network model, various disturbance scenarios to be evaluated are generated according to the disturbance mechanism, target, scope, amplitude, spatial pattern, and random seed. A disturbance scenario can be represented as: ,in, For the disruptive tribe, For the set of disturbed components or the set of candidate objects, The disturbance range, π represents the disturbance amplitude, and π represents the spatial pattern of the disturbance. The seed is random. The disturbance range describes the number, proportion, or spatial coverage of the disturbed components; the disturbance amplitude describes the degree of state change of a single disturbed component. By separating the two, different engineering scenarios such as "a small number of components suffering severe damage" and "a large number of components suffering minor damage" can be distinguished.

[0037] The perturbation family can include at least component availability loss, capacity degradation, demand shock, and cross-layer dependency weakening. Specific parameterization methods are shown in Table 1 below.

[0038] Table 1. Disturbance parameters and their meanings

[0039] The randomized model extracts disturbed components from the candidate component set according to uniform probability or a preset probability distribution to simulate dispersed, targetless failures. The critical target model ranks components based on node degree, betweenness, capacity, service scale, cross-layer dependency, number of critical users, or pre-defined importance indicators, prioritizing components with higher importance. The spatial clustering model first selects the disturbance center, then selects surrounding components based on geographical distance, topological distance, administrative region, or adjacency to form a local disturbance cluster. In addition to the above models, spatial disturbance models can also include path-dependent models, regional disaster models, path-propagation models, failure modes related to similar components, or spatial distribution models generated by external disaster models.

[0040] Each perturbation scenario saves complete perturbation parameters, perturbed component identifiers, component states before perturbation, component states after perturbation, and a random seed. For the same combination of parameters, the same set of perturbed components can be repeatedly generated using the same random seed, thus ensuring the reproducibility of the experiment and risk assessment process.

[0041] This process allows for the use of a unified parameter interface to describe disturbances of different infrastructure types, fault mechanisms, damage levels, and spatial organization forms, thus addressing the problems of limited disturbance types, mixed disturbance ranges and amplitudes, and difficulty in reproducing scenarios in existing solutions.

[0042] In step S102, various disturbance scenarios to be evaluated are applied to the multi-layer interdependent infrastructure network model, the system state after the disturbance is set as the initial disturbance information, and physical constraint cascade simulation is performed in discrete rounds to obtain the complete cascade propagation trajectory of each disturbance scenario.

[0043] In some embodiments, various disturbance scenarios to be evaluated are applied to a multi-layered interdependent infrastructure network model, the system state after the disturbance is set as the initial disturbance information, and physical constraint cascade simulation is performed in discrete rounds to obtain the complete cascade propagation trajectory of each disturbance scenario, including: Obtain the current system state of each infrastructure subsystem to construct an independent operating model for each infrastructure subsystem, and match the corresponding physical balance constraints, capacity constraints, and operating boundary constraints for each independent operating model; With the goal of minimizing the weighted sum of the total system operating cost, total service loss penalty, and total constraint violation penalty, cross-subsystem unified emergency scheduling calculations are performed on the independent operating models of the infrastructure subsystems of all subsystems to obtain the operating traffic and load utilization of all components in this round. Components whose load utilization exceeds a preset failure threshold are selected and added to the candidate failure set; Identify newly added failed components from the candidate failure set and update the network topology and component status of each infrastructure subsystem; Based on the preset cross-layer dependency relationships, the running status of all affected cross-layer dependency components is updated synchronously; Repeat the above process until the preset simulation termination condition is met to obtain the complete cascading propagation trajectory of each disturbance scenario.

[0044] In actual implementation, such as Figure 3 As shown, a perturbation is applied to the interdependent infrastructure network, and the system state after the perturbation is set to the round 0 state. Subsequently, physical constraint cascade simulations are performed according to discrete rounds or continuous time steps. For the round... Cascaded simulation first constructs operational models for each infrastructure subsystem based on currently available components, component capacity, service requirements, and cross-layer dependency states. Different subsystems can employ different physical equations or operational rules. For example, power systems can use AC power flow, DC power flow, or optimal power flow models; natural gas systems can use gas flow equations or linearized pipeline flow models; water supply systems can use hydraulic balance models; transportation systems can use traffic assignment or capacity constraint models; and communication systems can use flow rate, connectivity, or service capacity models.

[0045] For the Each infrastructure subsystem, whose physical state and constraints can be uniformly represented as:

[0046]

[0047] in, For the first The subsystem in the first The state variables of the wheel, It is an adjustable control variable. For demand or external disturbances, This is a physical equilibrium equation. For capacity, operating boundaries and safety constraints.

[0048] In one implementation, a unified emergency dispatch model is constructed with the goal of minimizing system operating costs, service loss penalties, and constraint violation penalties:

[0049] in, For the first The operating or scheduling cost of each subsystem This represents the service loss of the subsystem. For capacity overload or constraint violation, and These are the penalty coefficients for service loss and constraint violation, respectively.

[0050] For components with flow and capacity attributes , its first Wheel utilization rate It can be uniformly represented as:

[0051] in, For the component in the The flow, load, or service volume undertaken by the wheel. This represents the effective capacity of the component in this round.

[0052] When a component's utilization exceeds a failure threshold, it is added to the candidate failure set. ,in, For components The corresponding overload, failure, or service interruption thresholds, For the first The set of components that are still available at the start of the cascade simulation. Based on component utilization, constraint violation severity, protection rules, failure probability, or preset priority, the set of newly added failed components for this round is determined from the candidate set. To avoid increasing computational costs due to continuous iterations before reaching a stable state, this embodiment sets a maximum number of cascading propagation rounds. For disturbed scenarios The actual number of cascading propagation rounds satisfy .in, This constraint can be set according to the scale of the infrastructure system, simulation accuracy requirements, and computing resources; if simulation convergence can be guaranteed by other termination conditions, this constraint may not be set.

[0053] After identifying newly added failed components in this round, the network topology, component status, and service capabilities of the relevant subsystems are updated. The availability or capacity of dependent components is also updated synchronously based on the cross-layer dependency transfer function, before proceeding to the next round of cascaded computation. Cascaded simulation terminates when any of the following conditions are met: no newly added failed components are found in the current round; the emergency operation model of at least one critical subsystem has no feasible solution; the overall service loss ratio reaches a preset threshold; the proportion of failed components reaches a preset threshold; the number of cascaded propagation rounds reaches the maximum number of rounds; or the system reaches a preset stable state or a systemic failure state. The maximum number of cascaded rounds, the maximum number of failed components per round, the overload threshold, the service loss threshold, and the penalty coefficient can be set according to the specific infrastructure type, system scale, protection settings, and industry standards. It should be noted that the specific implementation parameters do not constitute a limitation on the general technical solution.

[0054] Through this process, the cascading propagation of different infrastructure subsystems is constrained by corresponding physical equations, capacity constraints, and cross-layer dependency rules. The generated cascading trajectories have clear engineering implications, avoiding the need to determine the consequences of system failure solely based on network connectivity or abstract propagation probabilities.

[0055] In step S103, the corresponding operational indicators are extracted from the complete cascading propagation trajectory of each disturbance scenario, and the operational indicators are converted into scenario severity labels.

[0056] In some embodiments, the corresponding operational metrics are extracted from the complete cascading propagation trajectory of each disturbance scenario, and the operational metrics are converted into scenario severity labels, including: Extract the final service loss, total service demand, number of failed components, and total number of components for each infrastructure subsystem from the complete cascading propagation trajectory of each disturbance scenario; The overall service loss ratio of each infrastructure subsystem is calculated based on the final service loss amount and the total service demand. The overall failure component ratio of each infrastructure subsystem is calculated based on the number of failed components and the total number of components. Based on the preset judgment rules, corresponding binary severity labels are generated according to the proportion of comprehensive service loss and the proportion of comprehensive failed components. The binary severity labels are then associated with the corresponding disturbance scenarios to obtain scenario severity labels.

[0057] In actual implementation, such as Figure 3 As shown, the operational feasibility, service loss ratio, failure component ratio, cascading propagation depth, critical component failure status, and cascading termination reason of each infrastructure subsystem are extracted from the cascading trajectory of each disturbance scenario, and a scenario-level severity label is generated accordingly. For cases involving multiple infrastructure subsystems, the service loss of each subsystem can be normalized and weighted to obtain the comprehensive service loss ratio. :

[0058] in, For the first Service loss of each subsystem For the total service requirements of this subsystem, This represents the weight of the subsystem.

[0059] Overall failure rate of components It can be represented as:

[0060] in, For the first The number of components that have failed when the cascading terminates in each subsystem. This represents the total number of components included in the evaluation in the initial state of the subsystem.

[0061] In the binary classification implementation, a scenario is marked as severe cascading when at least one of the following conditions is met: the emergency operation model of the critical subsystem has no feasible solution; the overall service loss ratio reaches a preset threshold; the overall failure component ratio reaches a preset threshold; the cascading propagation depth reaches a preset threshold; or a critical infrastructure component or critical cross-layer dependency link fails. Otherwise, the scenario is marked as non-severe cascading. It can be represented as:

[0062] in, To determine whether the emergency operation model is feasible, For the service loss threshold, The threshold for the proportion of failed components. For the propagation depth of cascading failure, This is the propagation threshold for the depth of cascading failure.

[0063] Severity labels can be expanded to include multi-category labels such as non-severe, moderate, and severe, or continuous risk scores can be generated based on service loss, failure rate, propagation depth, and critical component failure status. The operational infeasibility, final service loss, number of ultimately failed components, and cascading propagation depth used to generate labels are only used in the offline physical simulation phase and are not input to the early warning model. By separating label generation information from model input information, information leakage can be avoided by preventing the model from using the final cascading consequences to infer labels.

[0064] This process allows physical simulation results from different infrastructure subsystems to be transformed into a unified supervised learning objective, while preserving the correspondence between severity labels and engineering consequences such as service disruption, operational infeasibility, and deep propagation.

[0065] In step S104, a sample of heterogeneous graphs of multiple relationships before cascading is constructed based on the multi-layer interdependent infrastructure network model, preset static physical properties of components, preset cross-layer dependencies, and initial disturbance information.

[0066] In some embodiments, a pre-cascade multi-relationship heterogeneous graph sample is constructed based on a multi-layer interdependent infrastructure network model, preset component static physical properties, preset cross-layer dependencies, and initial perturbation information, including: The entity nodes of the multi-layer interdependent infrastructure network model are treated as heterogeneous nodes. The internal topological connections and cross-layer dependencies of each infrastructure subsystem in the multi-layer interdependent infrastructure network model are used as heterogeneous graph edges. Extract node features from the static physical properties of preset components; Extract node perturbation features and scenario-level global features from the initial perturbation information; Standardization processing is performed on heterogeneous graph nodes, heterogeneous graph edges, node features, node perturbation features, and scenario-level global features to generate cascaded multi-relation heterogeneous graph samples, and the pairing association between heterogeneous graph samples and corresponding severity labels is stored.

[0067] In actual implementation, such as Figure 3 As shown, for each disturbance scenario Construct a pre-cascade multi-relationship heterogeneous graph sample based on interdependent infrastructure networks, component static properties, cross-layer dependencies, and initial perturbation information. , ,in, For a heterogeneous graph node set, For a set of edges in a heterogeneous graph, For a set of relation types, The node feature matrix, This represents a scenario-level global perturbation feature.

[0068] Heterogeneous graph nodes can directly correspond to the original nodes in each infrastructure subsystem, or they can be aggregated based on geographical region, functional unit, site, or other uniform scale. For situations where the data granularity of different infrastructure subsystems is inconsistent, one type of infrastructure node or uniform spatial unit can be selected as the anchor node, projecting equipment, edge, and disturbance information from other subsystems to the corresponding anchor node. Relationship types include at least the topological relationships within each infrastructure subsystem, cross-layer dependencies between different subsystems, and node state relationships. For example, power topology relationships, natural gas pipeline relationships, water supply network relationships, communication link relationships, transportation connection relationships, energy supply dependencies, control dependencies, and service support dependencies can be set separately. Node static characteristics can include node degree, number of adjacent components, capacity of adjacent components, average impedance or equivalent resistance, number of supply equipment, service capacity, demand scale, number of cross-layer associated components, and number of cross-layer dependent edges. Different infrastructure types can be configured with different physical characteristics and mapped to a unified node feature space through feature encoding or missing value masks.

[0069] Continuous features are standardized using only training set statistics:

[0070] in, For nodes The Dimensional standardization features For nodes The Dimensional original features, and They represent the first The mean and standard deviation of the dimensional feature in the training set. To prevent positive numbers from being divided by zero.

[0071] Project disturbances occurring on nodes, connecting edges, devices, demand units, or cross-layer dependency edges onto the associated graph nodes. For nodes... Associated set of disturbed components The node perturbation effect can be expressed as ,in, For components The amplitude of the disturbance For components To the node The weights of the projection.

[0072] Projection weights can employ endpoint equalization, capacity weighting, service scale weighting, geographical distance attenuation, topological distance attenuation, or learnable attention weights. This projection method can uniformly convert component perturbations of different types and scales into node-level perturbation masks and perturbation intensity features. Scenario-level global features include at least perturbation family encoding, perturbation range, perturbation amplitude, spatial pattern encoding, number of perturbed components, number of affected infrastructure layers, and other features that describe the overall state of the scenario. Furthermore, the node feature dimension and global feature dimension are determined based on the specific infrastructure type and available data, and are not limited to fixed values ​​in the general steps. It is worth noting that the early warning model input explicitly excludes subsequent cascading failure events, final service loss, the number of ultimately failed components, cascading propagation depth, post-cascading utilization, and the severity label itself. Therefore, the heterogeneous graph only represents information available immediately after the perturbation is applied and before cascading propagation has unfolded.

[0073] This process can transform multi-layer network topology, heterogeneous component attributes, cross-layer dependencies, and different types of perturbations into a unified graph learning input, solving the problem that traditional feature engineering struggles to simultaneously represent multiple infrastructure structures and perturbation locations.

[0074] In step S105, the pre-constructed heterogeneous graph neural network is iteratively trained using situation severity labels and cascaded pre-relationship heterogeneous graph samples to obtain a heterogeneous graph-trained risk prediction model.

[0075] In actual implementation, such as Figure 3 As shown, the heterogeneous graph neural network includes a node feature encoding layer, a global feature encoding layer, at least one relation-specific message passing layer, a relation fusion layer, a graph-level pooling layer, a feature concatenation layer, and a classification output layer.

[0076] The node feature encoding layer will convert the original node Standardized node feature vectors Mapped to the initial hidden state , ,in, It is a non-linear activation function. The trainable weight matrix for the node feature encoding layer. This is the trainable bias vector for the node feature encoding layer.

[0077] The global feature encoding layer will incorporate context-level global features. Mapped to global hidden representation , ,in, This is the trainable weight matrix of the global feature encoding layer. This is the corresponding trainable bias vector.

[0078] For the Layer Relationship ,node From the relationship Receive messages in the corresponding neighbor set :

[0079] in, For summation, mean, maximum, or attention aggregation, For relationship In the The trainable parameter matrix corresponding to the layer, For nodes In the Hidden feature representation before layer message passing, To pass through relationships With nodes The set of connected neighbors.

[0080] Connections within different infrastructure subsystems and cross-layer dependencies are communicated using corresponding parameter matrices. For example, power connections, natural gas connections, water supply connections, communication connections, and energy supply dependencies can each be configured with independent message passing parameters. The relationship fusion layer updates node representations as follows: :

[0081] in, For the first A layer is a trainable weight matrix used to transform the state of a node itself. For the first Layer Relationship The corresponding fixed weights or learnable weights, For nodes In the Hidden feature representation before layer message passing.

[0082] go through After layer message passing, graph-level pooling is performed on the node representation:

[0083] in, For graph-level representation, The pooling method can employ mean pooling, max pooling, attention pooling, hierarchical pooling, or a combination thereof. For nodes go through The final hidden feature representation obtained after layer message passing and relationship fusion.

[0084] Representing the graph level With global feature representation spliced ​​as Then, the severe cascading probability is output through one or more fully connected layers and a sigmoid function:

[0085] in, This represents the predicted probability that a perturbation scenario will develop into a severe cascading failure. The scenario output by the model The corresponding predicted score or log odds, It is a natural constant.

[0086] The number of message passing layers, the dimension of hidden features, the activation function, the Dropout ratio, and the pooling method can be configured according to the scale of the infrastructure network, the number of relation types, and the size of the training samples. Specific implementation parameters are not uniquely limited in the steps described.

[0087] Through this process, the model can learn the propagation relationships within each infrastructure subsystem and the cross-layer dependencies between different subsystems, and integrate the location of local perturbations, network structure and overall scenario features in the graph-level representation.

[0088] Furthermore, the heterogeneous graph samples before cascading are used as model input, and the aforementioned severity labels are used as supervision signals to train the heterogeneous graph neural network. The samples are divided into training, validation, and test sets according to scenario groups, perturbation configurations, random seeds, spatial regions, or other grouping rules. During partitioning, highly similar or repeatedly generated scenarios should be avoided from being uncontrollably included in both the training and test sets simultaneously. The mean, standard deviation, and other data preprocessing parameters of nodes and global continuous features are calculated only using the training set. The validation and test sets use the standardized parameters obtained from the training set to prevent statistical information from the validation or test sets from entering the model training process.

[0089]

[0090] in, Train the loss function for the model. The number of training samples. For the true labels of the samples, To predict probabilities for the model, and Indicates category weight, For model parameters, is the regularization coefficient.

[0091] The model can update its parameters using Adam, AdamW, or other gradient optimization algorithms, and select the optimal model based on validation set loss, macro-average F1, balanced accuracy, severe class recall, or other evaluation metrics. Training can be terminated early to reduce the risk of overfitting when the validation metrics no longer improve within a preset number of rounds.

[0092] The learning rate, weight decay coefficient, batch size, maximum number of training epochs, and number of early stopping epochs are determined based on the data size and computing resources.

[0093] During the validation and testing phases, at least the accuracy, balanced accuracy, macro-average F1 score, severity category precision, and severity category recall should be calculated. For security alert tasks, model selection metrics and risk thresholds can be determined primarily based on severity category recall and false negative cost.

[0094] This process allows the use of engineering labels generated by physical cascade simulation to train early warning models, enabling the models to learn the mapping relationship between initial disturbances, network structure, physical properties, cross-layer dependencies, and final severe consequences without directly performing full cascade simulation.

[0095] In step S106, the scenario to be evaluated is input into the heterogeneous graph training risk prediction model to obtain the severe cascade probability and its corresponding risk level.

[0096] In some embodiments, the scenario to be evaluated is input into a heterogeneous graph training risk prediction model to obtain the severity cascade probability and its corresponding risk level, including: The evaluation scenario is preprocessed to generate a multi-relationship heterogeneous graph before cascading; The multi-relationship heterogeneous graph before cascading is input into the heterogeneous graph training risk prediction model to predict the probability of severe cascading of the scenario to be evaluated. Based on preset threshold classification rules, the risk level corresponding to the probability of severe cascading is obtained.

[0097] In actual implementation, after model training is completed, the model parameters, entity mapping relationships, feature order, relational edge construction rules, perturbation projection rules, and standardized parameters of the training set are saved. For new scenarios to be evaluated, the current interdependent infrastructure network state and initial perturbation information are first obtained. Then, a pre-cascade heterogeneity graph is generated according to the same entity mapping, feature order, standardized parameters, and relational edge construction rules as in the training phase. The pre-cascade heterogeneity graph is then input into the trained model to calculate the severe cascading probability. .when When the preset warning threshold is reached, a severe cascading warning is output, for example,

[0098] Among them, it should meet the following requirements. , This is the first risk classification threshold between low and medium risk. This is the second risk classification threshold between medium and high risk. At that time, a severe cascading warning will be issued.

[0099] Risk thresholds can be determined based on the recall rate, false positive rate, and cost of missed detections for severe categories on the validation set, operational safety level, and available computing resources. Early warning results should include at least scenario identification, severity cascading probability, and risk level, and may also include the affected infrastructure layer, critical nodes, critical components, critical cross-layer dependencies, risk heatmap, and high-risk area ranking. Critical nodes and critical relationships can be obtained through node attention, gradient attribution, feature masking, channel ablation, or other interpretation methods. Interpretation results are used to assist operators in locating risk sources and screening areas requiring further analysis. For high-risk scenarios, they are pushed to refined cascading simulation, emergency dispatch, or intervention analysis modules; for low-risk scenarios, the priority of full physical simulation can be reduced; for medium-risk scenarios near the risk threshold, supplementary simulation, sensitivity analysis, or manual review can be performed.

[0100] The prediction phase primarily involves feature construction and forward model computation, eliminating the need to repeatedly perform complete multi-round cascaded simulations for each scenario to be evaluated. Therefore, this process improves the efficiency of large-scale scenario screening and prioritizes the allocation of limited computational resources to high-risk and critical-risk scenarios.

[0101] It should be noted that interdependent infrastructure is not limited to power-natural gas systems, but can also be two- or multi-layer interdependent systems such as power-communication, power-transportation, power-water, and gas-heat-power. Cross-layer relationships can be energy supply dependence, communication dependence, control dependence, transportation dependence, or resource dependence; physical cascading simulation can employ AC optimal power flow, DC optimal power flow, linearized power flow, probabilistic power flow, dynamic power flow, or other simulation models that satisfy the corresponding infrastructure operation constraints; the projection of component disturbances onto learning nodes can employ methods such as endpoint equalization, direct mapping to belonging nodes, nearest node mapping, capacity weighting, distance decay, or attention weighting; risk prediction models can employ graph convolutional networks, graph attention networks, heterogeneous graph Transformers, message-passing neural networks, or other graph learning models capable of handling multiple relationship types; labels can take the form of binary classification, multi-class classification, continuous risk scores, or the time to reach a severe state; outputs can be risk probability, risk level, risk ranking, critical nodes, critical coupling paths, or risk heatmaps. Risk thresholds can be adjusted based on industry standards, historical incidents, system size, and the costs of false negatives and false positives.

[0102] The following detailed description of the cascading failure early warning method for associated infrastructure systems proposed in this invention is based on a specific embodiment. The purpose of this embodiment is to verify whether, without using posterior information such as branch failures, cascading depth, and final load loss generated during cascading propagation, it is possible to predict in advance whether a current disturbance may evolve into a severe cascading failure, solely based on the network topology, equipment operating capabilities, gas-electric coupling relationships, and disturbance parameters available at the time of the disturbance.

[0103] Taking a regionally interdependent power-natural gas system as the object, the power system is considered the main layer for cascading propagation and risk assessment, while the natural gas system is considered the related layer with a power supply dependency relationship with gas generators. The power layer includes substations or busbars, transmission lines, generators, and loads; the natural gas layer includes gas sources, gas nodes, gas pipelines, and gas demand; natural gas nodes and gas generators are connected via gas-electric coupling edges. When a natural gas node, gas transmission channel, or gas-electric coupling relationship is disturbed, the upper limit of the available output of the corresponding gas generator decreases with the gas supply capacity, thereby transmitting the disturbance from the natural gas side to the power side, and potentially further causing line overload, load reduction, and power cascading failure.

[0104] In the data preprocessing stage, the original power equipment is grouped according to its substation or busbar, forming 3389 substation-level learning nodes. Each learning node aggregates the transmission lines, generators, loads, gas generators, and natural gas supply relationships associated with that substation. This process preserves the topological connections and main physical attributes of the power network while uniformly mapping different types of components to a node space that can be processed by heterogeneous graph neural networks. It should be noted that the 3389 learning nodes are the number of nodes in the machine learning model and are not equivalent to the number of original line, generator, load, or natural gas nodes. The original components still retain a unique number and their associated node, rated capacity, operating status, and coupled objects to establish the correspondence between learning nodes and actual equipment during disturbance generation, physical simulation, and result traceability.

[0105] To ensure a consistent data foundation for physical simulation, sample generation, and model prediction, this embodiment first establishes a unified component index and entity mapping relationship. For each transmission line, at least the starting node, ending node, line reactance, thermal stability capacity, and initial commissioning status are recorded; for each generator, at least the power node it is located at, maximum active power output, minimum active power output, generation cost coefficient, and fuel type are recorded; for each load, at least the node it is located at and the baseline active power demand are recorded; for natural gas nodes and gas transmission components, at least their gas supply capacity, connected objects, and initial availability status are recorded; for gas-electric coupling edges, at least the gas supply node, the gas generator receiving the power, and the corresponding power supply capacity or conversion factor are recorded.

[0106] In this embodiment, the mapping between original components and learning nodes follows deterministic rules. Generators and loads are mapped to their respective power nodes; disturbance information of transmission lines is mapped to nodes at both ends of the line, which can be allocated using an end-point equal distribution or capacity weighting method; disturbances of natural gas nodes are mapped along the gas-electric coupling relationship to the power node where the gas generator they power is located; weakening of the gas-electric coupling edge is directly mapped to the node where the affected gas generator is located. The above mapping relationships remain unchanged during the training, validation, testing, and actual prediction phases to avoid inconsistencies in model input due to changes in entity order or node number.

[0107] Through the above data organization method, different types of disturbances, such as reduced transmission line capacity, generator shutdown, increased load, insufficient natural gas supply, and weakened gas-electric coupling, can be converted into disturbance masks and disturbance intensity characteristics on the corresponding learning nodes, while still being able to trace back to the actual disturbed components.

[0108] This embodiment constructs 7200 disturbance scenarios for interdependent power-natural gas systems. All model training, validation, testing, label statistics, and ablation analysis are based on these 7200 scenarios as a unified statistical object. Each scenario is jointly determined by the disturbance family, disturbance range, disturbance amplitude, spatial pattern, set of disturbed components, and random seed.

[0109] This embodiment sets up four types of disturbances: availability loss, capacity degradation, load surge, and cross-layer dependency weakening. Availability loss simulates the outage of lines, generators, gas nodes, or other critical components due to equipment failure, natural disasters, or attacks; capacity degradation simulates a decrease in the available capacity of transmission lines, generators, or gas supply components; load surge simulates increased electricity load or natural gas demand caused by high temperatures, sudden demand spikes, or emergency energy use; and cross-layer dependency weakening simulates a decrease in available output of gas generators due to insufficient gas supply capacity or damage to the gas-electric coupling link.

[0110] The specific parameterization methods for various disturbances are shown in Table 2.

[0111] Table 2 Experimental Disturbance Scenario Settings

[0112] in, The disturbance range and disturbance magnitude are defined separately in the scenario design: the disturbance range describes the number or proportion of affected components, while the disturbance magnitude describes the degree of change in the capacity of a single affected component. For example, in "30% of line capacity decreases by 20%", 30% represents the disturbance range, and 20% represents the disturbance magnitude. This design can distinguish between two different scenarios: a small number of components severely damaged and a large number of components slightly damaged.

[0113] In a feasible scenario parameter configuration, the perturbation range can be selected from 0.10, 0.30, 0.50, 0.70, and 0.90; for capacity degradation, load surges, and cross-layer dependency weakening, the perturbation amplitude can be selected from 0.10, 0.20, 0.30, 0.40, and 0.50; availability loss adopts a binary state of component availability or unavailability. Each basic configuration can use 20 random seeds to repeatedly generate component combinations. To increase the sample density near the transition between severe and non-severe states, transitional scenarios can be supplemented in addition to the basic scenarios. After basic scenario generation, transitional scenario supplementation, repeated scenario checking, and validity screening, this embodiment ultimately forms 7200 valid scenarios.

[0114] This embodiment sets up three spatial modes: random, critical target, and spatial clustering. The random mode samples without replacement from eligible candidate components; the critical target mode selects top-ranked components based on node degree, betweenness, transmission capacity, power generation capacity, or preset importance indicators; the spatial clustering mode first identifies the disturbance center and then selects surrounding components to form a local disturbance cluster based on geographical distance, network topology distance, or adjacency. Through these settings, 7200 scenarios can simultaneously cover dispersed random faults, critical component damage, and regional concentrated disturbances.

[0115] Each scenario stores a unique scenario ID, perturbation type, perturbation range, perturbation amplitude, spatial pattern, random seed, perturbed component ID, and the parameters of the component after perturbation. This allows the same scenario to be repeatedly constructed and physical simulations to be re-executed, facilitating the checking of the label generation process and model prediction results.

[0116] For each disturbance scenario, the corresponding disturbance is first applied to the interdependent power-natural gas system, and the disturbed network is set as the round 0 state. The natural gas-side disturbance is converted into the available output constraint of the relevant gas generators through the gas-electric coupling relationship, while the power-side disturbance directly changes the line status, line capacity, generator available output, or load demand.

[0117] In each round of cascaded simulation, an emergency dispatch model is solved, incorporating nodal power balance, generator output limits, line power flow constraints, load shedding variables, and line overload relaxation variables. The objective function simultaneously considers generation costs, load shedding penalties, and line overload penalties. After solving, the line utilization rate is calculated based on line power flow and current available capacity, and it is determined whether to trigger protection actions.

[0118] The main simulation parameters used in this embodiment are shown in Table 3.

[0119] Table 3. Main simulation parameter settings in the embodiments

[0120] When the line utilization reaches the immediate tripping threshold, the corresponding line is added to the candidate tripping set for this round and sorted from highest to lowest overload level. Candidate lines are removed within a constraint of no more than 5 lines per round, the network topology and available component status are updated, and the next round of scheduling and power flow calculation begins. The cascading simulation of the current scenario terminates when no new failures occur, the emergency scheduling model has no feasible solution, the load loss ratio reaches 10%, or the simulation reaches 8 rounds.

[0121] The parameters described above are used to illustrate the possible implementation methods of this embodiment and do not constitute the only limitation on the protection threshold. In actual engineering deployments, the immediate tripping threshold, delayed tripping threshold, maximum number of trips per round, and maximum number of cascading rounds can be adjusted according to the relay protection settings, line short-time overload capacity, dispatching procedures, and network scale.

[0122] After completing the cascading simulation for each scenario, the feasibility of emergency dispatch, the maximum load loss ratio, the proportion of ultimately failed lines, the cascading propagation depth, and the reason for simulation termination are extracted from the simulation trajectory. This embodiment specifies that a scenario is marked as a severe cascading scenario when it meets any of the following conditions: the emergency dispatch model has no feasible solution; the load loss ratio is not less than 10%; the proportion of ultimately failed lines is not less than 1%; or the cascading propagation depth is not less than 5. All other scenarios are marked as non-severe cascading scenarios.

[0123] The above metrics are only used for generating supervision labels during the offline phase and are not used as input to the heterogeneous graph early warning model. Specifically, the model input does not include subsequent line failures occurring during the cascading process, the number of cascading propagation rounds, the final load loss, the line utilization rate after cascading, the reason for simulation termination, and the severity label. This setting ensures that the early warning results are derived from information available when the disturbance occurs, rather than using the consequences of the incident to infer the severity of the scenario.

[0124] Of the 7200 scenarios, 4361 were non-severe (60.57%), and 2839 were severe (39.43%). This sample distribution contains a sufficient number of both severe and non-severe scenarios, making it suitable for training a binary classification early warning model. To address the issue of differing class numbers, a class-weighted loss is employed during training to reduce the likelihood of the model biased towards the majority class.

[0125] For each scenario, a pre-cascade heterogeneous graph is constructed using 3389 substation-level learning nodes. The relationship channels in the graph include node self-state relationships, power topology relationships, and gas-electric coupling relationships. Power topology relationships are determined by transmission line connections; gas-electric coupling relationships are determined by the power supply dependency between natural gas nodes and gas generators; node self-state relationships are used to preserve the original characteristics and current perturbation states of the nodes.

[0126] This embodiment constructs a 21-dimensional node feature for each learning node. This 21-dimensional feature is composed of static structural features, static physical features, and scene perturbation features, covering at least node degree, number of adjacent lines, capacity of adjacent lines, average line reactance, number of generators, generating capacity, node load, number of gas generators, number of associated natural gas nodes, number of gas-electric coupling edges, and perturbation masks and aggregated perturbation intensity values ​​corresponding to lines, generators, loads, natural gas nodes, and gas-electric coupling relationships. Perturbations on different types of components are projected onto the corresponding power nodes through pre-defined entity mapping relationships.

[0127] Each scenario also constructs a 10-dimensional global perturbation feature to encode the perturbation family, perturbation range, perturbation amplitude, spatial pattern, and scale of the perturbed component. Node features reflect the specific location and local effect of the perturbation in the network, while global features reflect the overall perturbation category and parameters of the scenario. Both serve as inputs to the early warning model.

[0128] For continuous features, the mean and standard deviation are calculated using only the training set samples. The validation set, test set, and actual prediction scenarios all use standardized parameters obtained from the training set; statistics for the corresponding datasets are not recalculated. Categorical features and perturbation masks are generated according to a pre-determined encoding order. All feature fields, dimensions, units, and order are saved along with the model file to ensure that the input during model deployment is consistent with that during training.

[0129] The heterogeneous graph neural network in this embodiment sequentially includes a node feature encoding layer, a global feature encoding layer, a heterogeneous relation message passing layer, a relation fusion layer, a graph-level pooling layer, a feature concatenation layer, and a classification output layer. Different relation channels employ different transformation parameters to learn power topology propagation, gas-electric coupling effects, and node self-states, respectively. After relation message passing and graph-level pooling, the obtained graph-level representation is concatenated with the encoded results of the 10-dimensional global perturbation features, and a severely cascaded probability is output through a fully connected classifier.

[0130] The training parameters, such as the number of model structure layers, hidden dimensions, learning rate, and batch size, should be based on the configuration saved in the actual training program and should be consistent with the code and experimental records in the official application documents. As an optional implementation, a two-layer heterogeneous message passing layer, 64-dimensional node hidden states, 32-dimensional global feature hidden states, ReLU activation function, and Dropout with a value of 0.2 can be used. It should be emphasized that these values ​​are adjustable model parameters and do not limit the scope of protection of this invention.

[0131] The 7200 scenes were divided into training, validation, and test sets according to scene groups and random seeds. During the partitioning, duplicate scenes with the same perturbation configuration were grouped to prevent highly similar scenes with identical parameters but different random seeds from being uncontrollably distributed across different subsets. The specific partitioning results are shown in Table 4.

[0132] Table 4 Data partitioning

[0133] During training, the pre-cascade heterogeneous graph of each scenario is used as input, and the severity labels generated by the physical constraint cascade simulation are used as supervision signals. Weighted binary cross-entropy loss is employed to compensate for the difference in the number of severe and non-severe samples. Model parameters are selected based on the macro-average F1 score of the validation set during training, and the model with the best validation metrics is retained. The test set is not involved in model parameter updates, feature standardization, early stopping decisions, or risk threshold selection.

[0134] After training and validating the model on the aforementioned 7200 scenarios, the model performance was evaluated using an independent test set containing 1096 scenarios. The results of the model on the test set are shown in Table 5.

[0135] Table 5 Model Performance

[0136] The test set contains 1096 scenarios, including 695 non-severe scenarios and 401 severe scenarios. The model correctly identified 648 non-severe scenarios and 381 severe scenarios, misidentified 47 non-severe scenarios as severe scenarios, and missed identifying 20 severe scenarios as non-severe scenarios. The corresponding confusion matrix is ​​shown in Table 6.

[0137] Table 6. Cascade Failure Prediction Results for the Test Set

[0138] Using only perturbation parameters, static network topology, static physical properties, and gas-electric coupling information, the model achieved a recall rate of 0.9501 for severe scenarios, meaning that 381 out of 401 severe scenarios in the test set were correctly identified. This result demonstrates that network structure, physical properties, cross-layer dependencies, and perturbation locations prior to cascading can provide effective information for identifying severe cascading risks.

[0139] The severity category accuracy was 0.8902, indicating that most of the severity warnings output by the model correspond to real severe scenarios. Compared to risk screening methods that rely entirely on scenario-by-scenario cascaded simulation, this model can first quickly screen a large number of candidate scenarios, and then submit scenarios predicted to be high-risk to the fine-grained cascaded simulation and emergency scheduling module, thereby reducing the computational resources occupied by low-risk scenarios.

[0140] To further verify the impact of different input information on the early warning results, while maintaining the consistency of the training set, validation set, and test set divisions and the main training conditions, the perturbation input, power topology features, static physical features, and gas-electric coupling channels were removed, and the model was retrained and tested. The ablation results are shown in Table 7.

[0141] Table 7 Ablation Experiment Results

[0142] After removing the perturbation input, the accuracy dropped to 0.6341, the balanced accuracy dropped to 0.5000, and the severe category recall dropped to 0. These results indicate that, under the data distribution of this embodiment, perturbation type, perturbation range, perturbation amplitude, spatial pattern, and perturbed location are direct sources of information for identifying severe cascading risks. Without this information, the model tends to classify the scenario as the majority category, failing to achieve effective severe cascading warnings.

[0143] After removing power topology or static physical inputs, the overall metrics decreased slightly, particularly the recall rates for severe categories, which dropped to 0.9352 and 0.9401 respectively. This indicates that network connectivity and operational capability parameters can provide supplementary information for discriminating the propagation of disturbance effects. After removing gas-electric coupling inputs, the overall metrics in this test did not decrease, suggesting that under the current conditions of 7200 scenarios and a shared basic network, disturbance inputs contribute more significantly to the overall classification results, and the overall metrics under random partitioning may not be sufficient to reflect the role of cross-layer relationships alone. Therefore, this result cannot be interpreted as the gas-electric coupling relationship being unimportant. In specific tests such as natural gas-side disturbances, no disturbance combinations observed, cross-regional migration, or network structure changes, the gas-electric coupling channel is still used to describe the physical path of gas supply constraints propagating to gas generators and the power network.

[0144] In a specific operational scenario, dispatchers need to assess the cascading risks when a region experiences simultaneous reductions in transmission capacity, increased load, and insufficient natural gas supply. Operators first input the current equipment status, line capacity, available generator output, node load, gas supply capacity, and gas-electric coupling relationship of the power and natural gas networks. Then, candidate disturbances are set, such as a 30% reduction in the capacity of some transmission lines within the target area, a 20% increase in the load of relevant nodes, and a 40% reduction in the available gas supply capacity of a certain gas supply node.

[0145] The system locates the substation-level learning node corresponding to each disturbed device based on its component number. Line capacity reduction is converted into a line disturbance mask and a capacity degradation intensity of 0.30; load increase is converted into a load surge mask and a load growth intensity of 0.20; gas supply capacity reduction is mapped along the gas-electric coupling relationship to the node where the relevant gas generator is located, and converted into a natural gas-side disturbance mask, a cross-layer impact marker, and a gas supply capacity reduction intensity of 0.40. The system further generates scene-level global features based on the number, location, and spatial distribution of the disturbed components.

[0146] After feature construction is complete, the system generates a pre-cascade heterogeneous graph containing 3389 learning nodes according to the node order, relation edge order, feature field order, and standardized parameters saved during the training phase. It then calls the trained heterogeneous graph neural network to calculate the severe cascading probability. If the predicted probability is below the low-risk threshold, the scenario is marked as low-risk; if the probability is between the two thresholds, it is marked as medium-risk; if the probability reaches or exceeds the high-risk threshold, a severe cascading warning is output, and the scenario is added to the fine-grained physical simulation queue. The risk threshold is determined by the validation set and can be adjusted based on the different costs of false negatives and false positives in engineering practice.

[0147] For scenarios predicted to be high-risk, the system simultaneously outputs the affected area, the main substations affected, the relevant transmission lines, and gas generators that may be affected by gas supply constraints. Dispatchers can then invoke the aforementioned physical constraint cascade simulation to perform multiple rounds of emergency dispatching and line tripping analysis for the scenario, and compare risk changes under different control measures. For example, measures such as activating non-gas standby units, reducing local load, adjusting network topology, increasing external power purchases, restoring natural gas supply, or prioritizing fuel supply to critical gas generator units can be evaluated.

[0148] By employing this approach of "rapid screening via machine learning—refined simulation of high-risk scenarios—comparison of control schemes," the system avoids performing complete multi-round cascade simulations on every candidate scenario. Instead, it prioritizes allocating computational resources to scenarios with a higher probability of severe cascading. This shortens the risk screening time for large-scale disturbance scenarios while retaining the role of physical simulation in confirming severe consequences and verifying emergency measures.

[0149] This embodiment unifies the description of power equipment, natural gas equipment, and gas-electric coupling relationships, enabling a decrease in natural gas supply capacity to be transmitted to the power layer in the form of changes in the available output of gas generators. This allows for the computational representation of the cascading impact across infrastructure. By constructing 7200 scenarios covering four types of disturbances, different ranges, different amplitudes, and three spatial modes, it can simultaneously examine random faults, damage to critical components, regional clustered disturbances, and weakening of cross-layer power supply dependence, overcoming the problem of insufficient coverage of single fault types or a small number of manually set scenarios. Furthermore, by using the posterior results of cascading simulations only for label generation and excluding final load loss, cascading depth, subsequent line failures, and simulation termination reasons from the model input, the model, in actual use, only relies on information available before or in the early stages of a disturbance, avoiding data leakage due to the use of accident consequence information.

[0150] Test results show that, based on a dataset of 7200 scenarios, the model achieved an accuracy of 0.9389 and a recall rate of 0.9501 for severe categories. This method first ranks a large number of candidate scenarios by risk, and then conducts detailed simulations and control scheme verification for high-risk scenarios. Therefore, it can improve the efficiency of risk screening for interdependent power-natural gas systems and provide auxiliary basis for dispatchers to formulate load reduction, standby startup, topology adjustment, and fuel security measures.

[0151] In summary, the cascading failure early warning method for related infrastructure systems proposed according to embodiments of the present invention has the following beneficial results: (1) By setting disturbance family, disturbance range, disturbance amplitude, spatial pattern and random seed respectively, it is possible to uniformly describe multiple fault mechanisms and spatial distribution methods, distinguish the scale of affected components from the damage level of individual components, and improve the coverage, comparability and reproducibility of disturbance scenarios.

[0152] (2) Supervision labels are generated through physical constraint cascade simulation, so that the severity judgment takes into account network balance, component operation capability, emergency dispatch feasibility and cascade propagation results at the same time, and the resulting training labels have clear engineering physical meaning.

[0153] (3) By projecting perturbations on different types of components to a unified learning node and setting up relationship channels such as same-layer topology and cross-layer dependency, it is possible to integrate network structure, physical attributes, perturbation location and cross-layer coupling information in the same heterogeneous graph, thereby enhancing the ability to express interdependent structures and potential propagation conditions.

[0154] (4) By isolating the label generation variables from the model input, the information such as service loss, propagation depth and subsequent component failure that can only be obtained after cascading propagation is explicitly excluded, so that the model can implement risk warning based on the information that can be obtained when the disturbance occurs or in the early stage of cascading, and avoid label leakage caused by the entry of posterior information into the input.

[0155] (5) After the model completes training, it mainly performs feature construction and model forward computation for new scenarios, without having to perform complete multi-round cascaded simulations for all candidate scenarios one by one in the initial screening stage. This can improve the efficiency of risk screening for large-scale scenarios and prioritize the allocation of fine simulation and emergency analysis resources to high-risk scenarios.

[0156] Next, referring to the accompanying drawings, a cascading failure early warning device for associated infrastructure systems proposed according to an embodiment of the present invention is described.

[0157] Figure 5 This is a block diagram of an early warning device for cascading failures in an associated infrastructure system, provided according to an embodiment of the present invention.

[0158] like Figure 5 As shown, the cascading failure early warning device 50 for related infrastructure systems includes: a generation module 501, an iterative simulation module 502, a conversion module 503, a construction module 504, an iterative training module 505, and a risk prediction module 506.

[0159] The generation module 501 generates various disturbance scenarios to be evaluated based on a pre-built multi-layer interdependent infrastructure network model, according to the disturbance mechanism, target, range, amplitude, spatial pattern, and random seed. The iterative simulation module 502 applies these disturbance scenarios to the multi-layer interdependent infrastructure network model, sets the post-disturbance system state as the initial disturbance information, and performs physical constraint cascade simulation in discrete rounds to obtain the complete cascade propagation trajectory of each disturbance scenario. The conversion module 503 extracts corresponding operational indicators from the complete cascade propagation trajectory of each disturbance scenario and converts these indicators into scenario severity labels. The construction module 504 constructs pre-cascade multi-relationship heterogeneous graph samples based on the multi-layer interdependent infrastructure network model, preset component static physical properties, preset cross-layer dependencies, and initial disturbance information. The iterative training module 505 uses scenario severity labels and pre-cascade multi-relationship heterogeneous graph samples to iteratively train the pre-built heterogeneous graph neural network to obtain a heterogeneous graph training risk prediction model. The risk prediction module 506 is used to input the scenario to be evaluated into the heterogeneous graph training risk prediction model to obtain the severe cascade probability and its corresponding risk level.

[0160] In some embodiments, the multi-layer interdependent infrastructure network model includes at least two interdependent infrastructure subsystems, and each infrastructure subsystem corresponds to an independent network topology, entity nodes, and connection edge set. Entity nodes correspond to the physical entities of their respective infrastructures, and each connection edge corresponds to the physical connection relationship of its respective infrastructure.

[0161] In some embodiments, the iterative simulation module 502 includes: The building unit is used to obtain the current system state of each infrastructure subsystem in order to build an independent operating model for each infrastructure subsystem, and to match the corresponding physical balance constraints, capacity constraints, and operating boundary constraints for each independent operating model; The first calculation unit is used to perform cross-subsystem unified emergency scheduling calculations on the independent operation models of the infrastructure subsystems of all subsystems with the goal of minimizing the weighted sum of the total system operating cost, total service loss penalty, and total constraint violation penalty, so as to obtain the operating traffic and load utilization of all components in this round. The filtering unit is used to filter out components whose load utilization exceeds a preset failure threshold and add them to the candidate failure set; The first update unit is used to determine the newly added failed components in this round from the candidate failure set and update the network topology and component status of each infrastructure subsystem; The second update unit is used to synchronously update the running status of all affected cross-layer dependent components according to the preset cross-layer dependency relationship; The iterative unit is used to repeatedly execute the aforementioned process until the preset simulation termination condition is met, thereby obtaining the complete cascading propagation trajectory of each disturbance scenario.

[0162] In some embodiments, the conversion module 503 includes: The first extraction unit is used to extract the final service loss, total service demand, number of failed components, and total number of components of each infrastructure subsystem from the complete cascading propagation trajectory of each disturbance scenario. The second calculation unit is used to calculate the comprehensive service loss ratio of each infrastructure subsystem based on the final service loss amount and the total service demand, and to calculate the comprehensive failure component ratio of each infrastructure subsystem based on the number of failed components and the total number of components. The association unit is used to generate corresponding binary severity labels based on the comprehensive service loss ratio and the comprehensive failure component ratio according to the preset judgment rules, and associate the binary severity labels with the corresponding disturbance scenarios to obtain scenario severity labels.

[0163] In some embodiments, the construction module 504 includes: The first setting unit is used to treat the entity nodes of the multi-layer interdependent infrastructure network model as heterogeneous nodes. The second setting unit is used to use the internal topological connections and cross-layer dependencies of each infrastructure subsystem in the multi-layer interdependent infrastructure network model as heterogeneous graph edges. The second extraction unit is used to extract node features from the static physical properties of preset components; The third extraction unit is used to extract node perturbation features and scenario-level global features from the initial perturbation information; The standardization processing unit is used to perform standardization processing on heterogeneous graph nodes, heterogeneous graph edges, node features, node perturbation features, and scenario-level global features to generate cascaded multi-relation heterogeneous graph samples and store the pairing association between heterogeneous graph samples and corresponding severity labels.

[0164] In some embodiments, the heterogeneous graph neural network is configured with a node feature encoding layer, a global feature encoding layer, multiple relation-specific message passing layers, a relation fusion layer, a graph-level pooling layer, a feature concatenation layer, and a classification output layer in sequence.

[0165] In some embodiments, the risk prediction module 506 includes: The preprocessing unit is used to preprocess the scenario to be evaluated in order to generate a pre-cascaded multi-relationship heterogeneous graph; The prediction unit is used to input the multi-relationship heterogeneous graph before cascading into the heterogeneous graph training risk prediction model to predict the severe cascading probability of the scenario to be evaluated. The acquisition unit is used to obtain the risk level corresponding to the severe cascading probability based on a preset threshold division rule.

[0166] It should be noted that the foregoing explanation of the embodiment of the cascading failure early warning method for related infrastructure systems also applies to the cascading failure early warning device for related infrastructure systems in this embodiment, and will not be repeated here.

[0167] The cascading failure early warning device for interconnected infrastructure systems proposed according to embodiments of the present invention has the following beneficial results: (1) By setting disturbance family, disturbance range, disturbance amplitude, spatial pattern and random seed respectively, it is possible to uniformly describe multiple fault mechanisms and spatial distribution methods, distinguish the scale of affected components from the damage level of individual components, and improve the coverage, comparability and reproducibility of disturbance scenarios.

[0168] (2) Supervision labels are generated through physical constraint cascade simulation, so that the severity judgment takes into account network balance, component operation capability, emergency dispatch feasibility and cascade propagation results at the same time, and the resulting training labels have clear engineering physical meaning.

[0169] (3) By projecting perturbations on different types of components to a unified learning node and setting up relationship channels such as same-layer topology and cross-layer dependency, it is possible to integrate network structure, physical attributes, perturbation location and cross-layer coupling information in the same heterogeneous graph, thereby enhancing the ability to express interdependent structures and potential propagation conditions.

[0170] (5) By isolating the label generation variables from the model input, the information such as service loss, propagation depth and subsequent component failure that can only be obtained after cascading propagation is explicitly excluded, so that the model can implement risk warning based on the information that can be obtained when the disturbance occurs or in the early stage of cascading, and avoid label leakage caused by the entry of posterior information into the input.

[0171] (5) After the model completes training, it mainly performs feature construction and model forward computation for new scenarios, without having to perform complete multi-round cascaded simulations for all candidate scenarios one by one in the initial screening stage. This can improve the efficiency of risk screening for large-scale scenarios and prioritize the allocation of fine simulation and emergency analysis resources to high-risk scenarios.

[0172] Figure 6 This is a schematic diagram of the structure of an electronic device according to an embodiment of the present invention.

[0173] The electronic device may include a memory 601, a processor 602, and a computer program stored in the memory 601 and executable on the processor 602. When the processor 602 executes the program, it implements the cascading failure early warning method for associated infrastructure systems provided in the above embodiments.

[0174] Furthermore, electronic devices also include: Communication interface 603 is used for communication between memory 601 and processor 602.

[0175] The memory 601 is used to store computer programs that can run on the processor 602.

[0176] The memory 601 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0177] If the memory 601, processor 602, and communication interface 603 are implemented independently, then the communication interface 603, memory 601, and processor 602 can be interconnected via a bus to complete communication between them. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of representation, Figure 6 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0178] Optionally, in a specific implementation, if the memory 601, processor 602, and communication interface 603 are integrated on a single chip, then the memory 601, processor 602, and communication interface 603 can communicate with each other through an internal interface.

[0179] Processor 602 may be a central processing unit (CPU), an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention.

[0180] This invention also provides a computer program product that, when executed by a processor, implements the above-described method for early warning of cascading failures in interconnected infrastructure systems.

[0181] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for early warning of cascading failures in interconnected infrastructure systems.

[0182] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0183] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "N" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0184] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or N executable instructions for implementing custom logic functions or processes, and the scope of preferred embodiments of the invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as will be understood by those skilled in the art to which embodiments of the invention pertain.

[0185] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0186] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, the N steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. If implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0187] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0188] Furthermore, the functional units in the various embodiments of the present invention can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0189] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for early warning of cascading failures in interconnected infrastructure systems, characterized in that, Includes the following steps: Based on a pre-built multi-layer interdependent infrastructure network model, various disturbance scenarios to be evaluated are generated according to the disturbance mechanism, target, disturbance range, disturbance amplitude, spatial pattern, and random seed. Various disturbance scenarios to be evaluated are applied to the multi-layer interdependent infrastructure network model. The system state after the disturbance is set as the initial disturbance information. Physical constraint cascade simulation is performed in discrete rounds to obtain the complete cascade propagation trajectory of each disturbance scenario. Extract the corresponding operational indicators from the complete cascading propagation trajectory of each disturbance scenario, and convert the operational indicators into scenario severity labels; Based on the multi-layer interdependent infrastructure network model, preset component static physical properties, preset cross-layer dependencies, and the initial disturbance information, construct a sample of multi-relationship heterogeneous graphs before cascading. The pre-constructed heterogeneous graph neural network is iteratively trained using the scenario severity labels and the cascaded pre-relationship heterogeneous graph samples to obtain a heterogeneous graph-trained risk prediction model. The scenario to be evaluated is input into the heterogeneous graph training risk prediction model to obtain the severe cascade probability and its corresponding risk level.

2. The method for early warning of cascading failures in interconnected infrastructure systems according to claim 1, characterized in that, The multi-layer interdependent infrastructure network model includes at least two interdependent infrastructure subsystems, and each infrastructure subsystem corresponds to an independent network topology, entity nodes, and connection edge set. The entity nodes correspond to the physical entities of their respective infrastructures, and each connection edge corresponds to the physical connection relationship of its respective infrastructure.

3. The method of claim 1, wherein, The process involves applying various disturbance scenarios to the multi-layered interdependent infrastructure network model, setting the system state after the disturbance as the initial disturbance information, and performing physical constraint cascade simulations in discrete rounds to obtain the complete cascade propagation trajectory of each disturbance scenario, including: Obtain the current system state of each infrastructure subsystem to construct an independent operation model for each infrastructure subsystem, and match corresponding physical balance constraints, capacity constraints, and operation boundary constraints for each independent operation model; With the goal of minimizing the weighted sum of the total system operating cost, total service loss penalty, and total constraint violation penalty, cross-subsystem unified emergency scheduling calculations are performed on the independent operating models of the infrastructure subsystems of all subsystems to obtain the operating traffic and load utilization of all components in this round. Components whose load utilization exceeds a preset failure threshold are selected and added to the candidate failure set; The newly added failed components in this round are determined from the candidate failure set, and the network topology and component status of each infrastructure subsystem are updated; Based on the preset cross-layer dependency relationships, the running status of all affected cross-layer dependency components is updated synchronously; Repeat the aforementioned process until the preset simulation termination condition is met to obtain the complete cascading propagation trajectory of each disturbance scenario.

4. The cascading failure early warning method for a correlation-oriented infrastructure system according to claim 1, wherein, The step of extracting corresponding operational indicators from the complete cascading propagation trajectory of each disturbance scenario and converting the operational indicators into scenario severity labels includes: Extract the final service loss, total service demand, number of failed components, and total number of components for each infrastructure subsystem from the complete cascading propagation trajectory of each disturbance scenario. The comprehensive service loss ratio of each infrastructure subsystem is calculated based on the final service loss amount and the total service demand, and the comprehensive failure component ratio of each infrastructure subsystem is calculated based on the number of failed components and the total number of components. Based on the preset judgment rules, corresponding binary severity labels are generated according to the comprehensive service loss ratio and the comprehensive failure component ratio, and the binary severity labels are associated with the corresponding disturbance scenarios to obtain the scenario severity labels.

5. The cascading failure early warning method for a correlation-oriented infrastructure system according to claim 1, wherein, The step of constructing a pre-cascade multi-relationship heterogeneous graph sample based on the multi-layer interdependent infrastructure network model, preset component static physical attributes, preset cross-layer dependencies, and the initial perturbation information includes: The entity nodes of the multi-layer interdependent infrastructure network model are treated as heterogeneous nodes; The internal topological connections and cross-layer dependencies of each infrastructure subsystem in the multi-layer interdependent infrastructure network model are used as heterogeneous graph edges. Extract node features from the static physical properties of the preset component; Extract the node perturbation features and scenario-level global features from the initial perturbation information; The heterograph nodes, heterograph edges, node features, node perturbation features, and scenario-level global features are standardized to generate the cascaded pre-multi-relation heterograph samples, and the pairing association between the heterograph samples and the corresponding severity labels is stored.

6. The cascading failure early warning method for a correlation-oriented infrastructure system according to claim 1, wherein, The heterogeneous graph neural network is configured with a node feature encoding layer, a global feature encoding layer, multiple relationship-specific message passing layers, a relationship fusion layer, a graph-level pooling layer, a feature concatenation layer, and a classification output layer in sequence.

7. The cascading failure early warning method for a correlation-oriented infrastructure system according to claim 1, wherein, The step of inputting the scenario to be evaluated into the heterogeneous graph training risk prediction model to obtain the severe cascading probability and its corresponding risk level includes: The scenario to be evaluated is preprocessed to generate a multi-relationship heterogeneous graph before cascading; The pre-cascade multi-relationship heterogeneous graph is input into the heterogeneous graph training risk prediction model to predict the severe cascading probability of the scenario to be evaluated. Based on a preset threshold classification rule, the risk level corresponding to the severe cascading probability is obtained.

8. A cascading failure early warning device for a correlation-oriented infrastructure system, characterized in that, include: The generation module is used to generate various disturbance scenarios to be evaluated based on a pre-built multi-layer interdependent infrastructure network model, according to the disturbance mechanism, target, disturbance range, disturbance amplitude, spatial pattern and random seed. The iterative simulation module is used to apply various disturbance scenarios to the multi-layer interdependent infrastructure network model, set the system state after disturbance as the initial disturbance information, and perform physical constraint cascade simulation in discrete rounds to obtain the complete cascade propagation trajectory of each disturbance scenario. The conversion module is used to extract the corresponding operational indicators from the complete cascading propagation trajectory of each disturbance scenario and convert the operational indicators into scenario severity labels. The construction module is used to construct a pre-cascade multi-relationship heterogeneous graph sample based on the multi-layer interdependent infrastructure network model, preset component static physical properties, preset cross-layer dependencies and the initial disturbance information; The iterative training module is used to iteratively train the pre-constructed heterogeneous graph neural network using the scenario severity labels and the cascaded pre-relationship heterogeneous graph samples to obtain a heterogeneous graph training risk prediction model. The risk prediction module is used to input the scenario to be evaluated into the heterogeneous graph training risk prediction model to obtain the severe cascade probability and its corresponding risk level.

9. An electronic device, comprising: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, the processor executing the program to implement the cascading failure early warning method for associated infrastructure systems as described in any one of claims 1-7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to implement the cascading failure early warning method for associated infrastructure systems as described in any one of claims 1-7.

Images