A power-traffic resilience evaluation method and device based on causal tracing

Abstract

This invention discloses a method and apparatus for assessing power-transportation resilience based on causal attribution, belonging to the field of smart distribution network optimization and operation technology. The method includes: generating typical fault scenario data based on acquired historical fault data by combining adversarial networks and physical mechanism models; training an agent based on the typical fault scenario data using a hypergraph convolutional network and the DQN algorithm to optimize the strategy for damaging nodes and output vulnerable node information; generating disaster impact results based on the vulnerable node information using a cascaded fault model; confirming the unique mapping relationship from fault phenomena to root causes using a causal reasoning network and a Bayesian network based on the disaster impact results; and confirming the causal chain between the disaster-causing mechanism, evolution path, and fault phenomena based on the unique mapping relationship to obtain the power-transportation resilience assessment result. This invention improves the safe operation capability of infrastructure in power-transportation coupled networks under extreme disasters.

Description

Technical Field

[0001] This invention belongs to the field of smart distribution network optimization operation technology, and in particular relates to a power-transportation resilience assessment method and device based on causal tracing. Background Technology

[0002] With the deep integration of the Global Energy Internet and intelligent transportation systems, the coupling and complexity of power and transportation networks, as key infrastructures of modern society, are growing exponentially. The large-scale expansion of smart distribution networks, the three-dimensional development of transportation networks, and the interactive penetration of diverse energy carriers, while improving system coordination efficiency, also cause the operating characteristics of power-transportation coupled networks to exhibit highly nonlinear features. The frequent occurrence of extreme natural disasters such as typhoons and rainstorms, coupled with multiple risks such as equipment aging and cross-network coupling, has transformed the failure mechanism of coupled networks from single-component failure to a complex mode of multi-hazard induction and cross-domain cascading evolution. Traditional fault scenario generation methods based on single networks are no longer sufficient to accurately characterize the dynamic mapping relationship between disasters and losses.

[0003] Traditional resilience assessment systems face multiple technical bottlenecks in dealing with complex failure scenarios in coupled power and transportation networks. Existing failure scenario generation mechanisms lack a quantitative description of the physical mechanisms causing damage from multiple disasters, making it difficult to generate scenario data that accurately reflects the actual evolution of disaster damage. Failure node identification methods based on traditional graph structures cannot effectively represent the multi-faceted relationships and high-dimensional topology of power and transportation networks, leading to reliance on experience in identifying vulnerable nodes and neglecting cross-network coupling failures. Furthermore, when facing coupled disasters, the system struggles to achieve in-depth diagnosis of failure causes and accurate formulation of resilience enhancement strategies, severely restricting the safe operation capability of critical infrastructure under extreme disasters. Summary of the Invention

[0004] The purpose of this invention is to provide a power-transportation resilience assessment method and device based on causal tracing, which dynamically guides the generation of multimodal data through a loss function to achieve a nonlinear and accurate mapping between disaster intensity and system loss.

[0005] To achieve the above objectives, the present invention is implemented using the following technical solution.

[0006] On the one hand, this invention provides a power-transportation resilience assessment method based on causal attribution, comprising:

[0007] Based on the acquired historical fault data, typical fault scenario data for classification are obtained through a fault scenario generation algorithm;

[0008] Based on the typical fault scenario data, topology awareness is performed through a hypergraph convolutional network to obtain node feature data;

[0009] Based on the node feature data, a Markov decision process model is constructed to address the problem of identifying vulnerable nodes in a power-transport coupled network.

[0010] The Markov decision process model is solved using the deep reinforcement learning (DQN) algorithm to obtain the strategy for destroying nodes and output information about vulnerable nodes.

[0011] Based on the information about the vulnerable nodes, a cascading failure model is constructed to address the evolution process of cascading failures.

[0012] Based on the cascaded failure model, disaster impact results are generated using information on vulnerable nodes;

[0013] Based on the disaster impact results, the correlation between the fault phenomena and the corresponding fault candidate intervals is determined through abductive reasoning network;

[0014] Based on the aforementioned correlation, a unique mapping relationship from the fault phenomenon to the root cause is confirmed through a Bayesian network.

[0015] Based on the unique mapping relationship, the causal chain between the disaster mechanism, evolution path and failure phenomenon is confirmed, and the power-transportation resilience assessment results are obtained.

[0016] Optionally, the step of obtaining classified typical fault scenario data based on the acquired historical fault data of the target area through a fault scenario generation algorithm includes:

[0017] Acquire historical fault data of the target area, as well as subjective decision data and objective decision data after classifying the historical fault data;

[0018] Based on the fault scenario generation algorithm, subjective decision data and objective decision data are processed in two channels, and the processed data are fused to obtain a fused feature dataset.

[0019] Based on the historical fault data, a fault mechanism model is constructed to calculate the physical correction term for the fault scenario data;

[0020] Based on the physical correction term and the fused feature dataset, typical fault scenario data for classification are obtained through data augmentation and scenario compression.

[0021] By using a fault mechanism model, the coupling relationship between the physical mechanism of disasters and the disaster resilience threshold of infrastructure is quantified into a loss function. This function guides the generation of disaster loss scenarios using a generative adversarial network (GAN). By combining the GAN with the physical mechanism model, the nonlinear correlation between disasters and losses is accurately reflected, providing scenario data with realistic disaster loss evolution patterns for disaster simulation of power and transportation coupled networks. The loss function dynamically guides the generation of multimodal data, achieving a precise nonlinear mapping between disaster intensity and system loss.

[0022] Optionally, based on the typical fault scenario data, topology awareness is performed using a hypergraph convolutional network to obtain node feature data, including:

[0023] Based on the typical fault scenario data, a convolution operation is performed on each node to aggregate the hyperedge information associated with each node to obtain a low-dimensional vector representation.

[0024] Based on the typical fault scenario data, each hyperedge is convolved to obtain the feature vector of the hyperedge;

[0025] The expression for the information transfer process between nodes and hyperedges is as follows:

[0026] (1);

[0027] In the formula, Representing node characteristics, Let the node degree matrix be... For learnable weight matrix, For trainable parameters, Let H be the hypermarginality matrix and H be the incidence matrix. It is a convolution function. To construct the characteristic matrix, T is the matrix transpose.

[0028] By using hypergraph convolutional networks to extract node features through topology perception, a high-level representation of the complex relationships between nodes in the power and transportation coupled network is achieved, which can represent the implicit relationships between nodes.

[0029] Optionally, the state space of the Markov decision process model includes node propagation characteristics. Hyperedge dynamic features and real-time load vector ;

[0030] The action space of the Markov decision process model is defined as the failure nodes of discrete actions. Let i represent the i-th node in the hypergraph.

[0031] By using node features extracted through hypergraph convolutional networks to confirm the state space and action space, the vulnerable node identification problem is transformed into a Markov decision process, thus optimizing the vulnerable node identification process.

[0032] Optionally, the reward function of the Markov decision process model is defined as:

[0033] (2);

[0034] (3);

[0035] in, Let F be the reward function, and F be a multi-objective function. The normalization coefficient is... For network stability, For fault recovery speed, For resource allocation capabilities, For system resilience, max is the maximization operator;

[0036] The state space, action space, and reward function are determined based on the node feature data.

[0037] A Markov decision process model is constructed based on the state space, action space, and reward function.

[0038] Based on the Markov decision process model, the DQN algorithm is used to solve the problem, resulting in an optimized strategy for destroying nodes and outputting information on vulnerable nodes.

[0039] By constructing a Markov decision process model using the state space, action space, and reward function, a comprehensive consideration of multiple key dimensions of network operation is achieved, enabling overall optimization of network performance. The disruptive node strategy selects disruptive node actions based on the current state of the agent in the Markov decision process model to disrupt nodes in the power-transportation coupled network, maximizing the degree of system failure. The core objective of this disruptive node strategy is to identify vulnerable nodes in the network.

[0040] Optionally, based on the cascading failure model, disaster impact results are generated using vulnerable node information, including:

[0041] Based on the information of vulnerable nodes, the fault propagation path output by the initial disaster node when the fault occurs is identified through the cascaded fault model;

[0042] Based on the fault propagation path, disaster impact results are generated by determining the correlation and coupling effects between nodes;

[0043] The propagation path includes the direction of the power line's spread and the main roads of the sprawling transportation network.

[0044] Optionally, the abductive reasoning network adopts a hierarchical architecture design, and the network units are composed of basic low-order polynomials;

[0045] The operational steps of the abductive reasoning network to determine the correlation between fault phenomena and corresponding fault candidate intervals include:

[0046] S1, based on the disaster impact results, uses multiple regression techniques to evaluate the parameters of the function and generate the intermediate node function of the current inference iteration layer;

[0047] S2, rank the intermediate node functions based on the squared prediction error and decide which node function will enter the next inference iteration layer;

[0048] S3. Repeat steps S1 and S2 until the preset abductive reasoning network cutoff rule is met to stop the reasoning iteration.

[0049] Optionally, the node function expression generated by the abductive reasoning network using multiple regression techniques is:

[0050] (4);

[0051] In the formula, This represents the calculation result of the function at the intermediate node of the current inference iteration layer; , , , , , , , , , , , , and For training parameters; , and The input feature variables are obtained from multi-source operation data of the power-transportation coupled network, including power grid operation parameters, traffic flow data and environmental monitoring parameters;

[0052] The formula for calculating the squared prediction error is:

[0053] (5);

[0054] In the formula, PSE is the squared prediction error of the abductive reasoning network; FSE is the fitting variance; OP is the overfitting penalty factor; C is the complexity penalty multiplier; K is the implicit value of the complexity of the abductive reasoning network; and N is the number of training samples. Let be the variance of the error in a priori estimate of the abductive reasoning network.

[0055] Optionally, the Bayesian network is a directed acyclic graph based on probabilistic reasoning, expressed as:

[0056] (6);

[0057] In the formula: Represents a directed graph; The node feature matrix, Each hyperedge contains a set of hyperedges with a number of outlier nodes greater than or equal to 2. P represents the conditional probability table on each node; B is a Bayesian network function consisting of a directed graph and a conditional probability table.

[0058] Secondly, the present invention provides a power-transportation resilience assessment device based on causal attribution, comprising:

[0059] The data acquisition and processing module is used to collect historical fault data and perform data preprocessing.

[0060] The scenario generation module is used to generate typical fault scenario data of classification based on preprocessed fault history data by combining generative adversarial networks and the physical mechanism of fault disasters.

[0061] The fault location module is used to optimize the strategy for damaging nodes and output vulnerable node information based on typical fault scenario data by combining hypergraph convolutional networks and deep reinforcement learning DQN algorithm.

[0062] The causal assessment module is used to generate disaster impact results based on vulnerable node information using a cascading failure model, and based on the disaster impact results, to confirm the causal chain between disaster-causing mechanisms, evolution paths and failure phenomena through causal reasoning networks and Bayesian networks.

[0063] The visualization module is used to generate power-transportation resilience assessment results through various visualization methods based on the causal chain between disaster-causing mechanisms, evolution paths, and failure phenomena.

[0064] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0065] This invention generates disaster impact results from vulnerable node information through a cascading failure model, and uses abductive reasoning networks and Bayesian networks to trace the causal origins of the disaster impact results. This enables a deep analysis of the entire chain of disaster node impact, cascading propagation, and fault causes, improving the accuracy of fault cause diagnosis results in disaster scenarios of power-transportation coupled networks. It also provides a systematic logical architecture for the quantitative analysis of the resilience of power-transportation coupled networks in multi-hazard scenarios.

[0066] This invention introduces a hypergraph convolutional network to characterize the diverse node types and complex relationships within a power-transportation coupled network. It combines this with the DQN algorithm to solve a Markov decision process model, enabling dynamic identification of vulnerable nodes within the hypergraph convolutional network's topological space. This optimizes the vulnerability node identification process's reliance on experience and increases the weighting of cross-network coupling failures, allowing for intelligent and precise location of critical nodes in fault propagation under complex disaster conditions. This invention improves the safe operation capability of infrastructure in power-transportation coupled networks under extreme disasters, and has significant practical implications for ensuring the stability and security of power supply. Attached Figure Description

[0067] Figure 1 This is an architecture diagram of the power-transportation resilience assessment method based on causal tracing of the present invention;

[0068] Figure 2 This is a flowchart of the power-transportation resilience assessment method based on causal tracing of the present invention;

[0069] Figure 3 This is a flowchart of the fault scenario generation method based on multi-hazard loss guidance of the present invention;

[0070] Figure 4 This is a flowchart of the intelligent fault node identification method based on the fusion of hypergraph and DQN of the present invention;

[0071] Figure 5 This is a flowchart of the integrated model for fault tracing and disaster resilience assessment of the present invention;

[0072] Figure 6 This is a structural diagram of the intelligent power-transportation resilience assessment device of the present invention. Detailed Implementation

[0073] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0074] The term "and / or" simply describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0075] Example 1: This example introduces a power-transportation resilience assessment method based on causal attribution, such as... Figure 1 As shown, the specific steps include:

[0076] Step 1: Fault Scenario Data Generation: Based on the acquired historical fault data, typical fault scenario data of classification are generated through a fault scenario generation algorithm;

[0077] The fault scenario generation algorithm uses a Generative Adversarial Network (GAN) as its basic framework. It deeply embeds knowledge of the physical mechanisms of faults and disasters into the training process of the GAN, effectively constraining the direction of scenario generation. This ensures that the generated typical fault scenario data reflects actual operating conditions, physical mechanisms, and operational patterns, providing accurate sample data for subsequent vulnerable node identification. The historical fault data includes power system data and transportation network-related data for a specified area.

[0078] Step 2, Causal Cause Assessment: The fault correlation matrix and fault feature matrix in the typical fault scenario data are used as input to the hypergraph convolutional network for topology sensing to obtain node feature data;

[0079] Based on the hyperedge dynamic features and node propagation features in the node feature data, as well as the obtained real-time load vector, the state space, action space, and reward function are determined; an intelligent agent is constructed to address the problem of identifying vulnerable nodes in power-transport coupled networks.

[0080] The agent is trained using the Deep Reinforcement Learning (DQN) algorithm. The agent iteratively updates the network by interacting with the environment, optimizes the strategy for destroying nodes, identifies vulnerable nodes, and outputs information about vulnerable nodes.

[0081] Step 3: Fault Origin Tracing and Disaster Resilience Assessment: Based on the information of vulnerable nodes, a cascading fault model is constructed to analyze the evolution process of cascading faults. According to the cascading fault model, the disaster impact results are output. Specific steps include identifying the fault originating from the initial disaster node, outputting the path of fault propagation to surrounding nodes, i.e., the propagation direction along power lines and the main arteries of the traffic network. Simultaneously, the relationships between nodes are analyzed, the coupling effects between different network nodes are examined, the impact of power supply failures on traffic signal systems is analyzed, and the reaction of traffic congestion to power load distribution is analyzed, outputting the interdependence and constraints between nodes during the cascading propagation process.

[0082] Based on the disaster impact results, the correlation between the fault phenomena and corresponding fault candidate intervals is determined through a causal reasoning network. A Bayesian network is introduced to calculate the causal contribution of each node to the disaster, resolving ambiguity in scenarios with multiple co-occurring causes. For coupling between direct and indirect causes, posterior probabilities are calculated based on a historical fault case database to filter out minor factors with low confidence. For cross-network coupled causes, the degree of indirect influence is quantified by defining a coupling strength index, establishing a hierarchical structure of causal relationships. Finally, an evidence weighting mechanism is used to determine the dominant cause and its path of action, forming a unique mapping relationship from the fault phenomenon to the root cause, eliminating analytical ambiguity under multi-factor interference.

[0083] Based on the unique mapping relationship, the causal chain between the disaster mechanism, evolution path and failure phenomenon is confirmed, and the power-transportation resilience assessment results are obtained.

[0084] This embodiment accurately characterizes the various types of nodes and complex relationships within the power-transportation coupled network using a hypergraph convolutional network. It constructs an agent to address the problem of vulnerable node identification and trains the agent using the DQN algorithm to dynamically identify vulnerable nodes in the topological space of the hypergraph convolutional network. This optimizes the dependence of the vulnerable node identification process on experience and improves the safe operation capability of infrastructure in the power-transportation coupled network under extreme disasters, which has significant practical implications for ensuring the stability and security of power supply.

[0085] Example 2 is based on the same inventive concept as Example 1, such as... Figure 1 As shown in the figure, this embodiment introduces a power-transportation resilience assessment method based on causal attribution. The method includes the following steps:

[0086] Based on the acquired historical fault data, a fault scenario generation algorithm is used to obtain typical fault scenario data for classification, specifically including:

[0087] Obtain historical fault data;

[0088] The fault history data is classified into subjective decision data and objective decision data, and processed separately through dual channels to obtain a fused feature dataset.

[0089] A fault mechanism model is constructed based on the aforementioned fault history data to calculate the physical correction term;

[0090] Based on the physical correction term and the fused feature dataset, typical fault scenario data for classification are obtained through data augmentation and scenario compression.

[0091] By using a fault mechanism model, the coupling relationship between the physical mechanism of disasters and the disaster resilience threshold of infrastructure is quantified into a loss function. This function guides the generation of disaster loss scenarios using a generative adversarial network (GAN). By combining the GAN with the physical mechanism model, the nonlinear correlation between disasters and losses is accurately reflected, providing scenario data with realistic disaster loss evolution patterns for disaster simulation of power and transportation coupled networks. The loss function dynamically guides the generation of multimodal data, achieving a precise nonlinear mapping between disaster intensity and system loss.

[0092] Based on the typical fault scenario data, topology awareness is performed through a hypergraph convolutional network to obtain node feature data;

[0093] Based on the typical fault scenario data, a convolution operation is performed on each node to aggregate the hyperedge information associated with each node to obtain a low-dimensional vector representation.

[0094] Based on the typical fault scenario data, each hyperedge is convolved to obtain the feature vector of the hyperedge;

[0095] The expression for the information transfer process between nodes and hyperedges is as follows:

[0096] (1);

[0097] In the formula, Representing node characteristics, Let the node degree matrix be... For learnable weight matrix, For trainable parameters, Let H be the hypermarginality matrix and H be the incidence matrix. It is a convolution function. To construct the characteristic matrix, T is the matrix transpose.

[0098] By using hypergraph convolutional networks to extract node features through topology perception, a high-level representation of the complex relationships between nodes in the power and transportation coupled network is achieved, which can represent the implicit relationships between nodes.

[0099] Based on the node feature data, the state space, action space, and reward function are determined, and an intelligent agent is constructed to address the problem of identifying vulnerable nodes in the power-transportation coupled network.

[0100] The state space includes node propagation characteristics. Hyperedge dynamic features and real-time load vector ;

[0101] The action space is defined as the destruction nodes of discrete actions, and the destruction nodes... Let i represent the i-th node in the hypergraph.

[0102] The reward function is defined as follows:

[0103] (2);

[0104] (3);

[0105] in, Let F be the reward function, and F be a multi-objective function. The normalization coefficient is... For network stability, For fault recovery speed, For resource allocation capabilities, For system resilience, max is the maximization operator;

[0106] The state space, action space, and reward function are determined based on the node feature data.

[0107] A Markov decision process model is constructed based on the state space, action space, and reward function.

[0108] Based on the Markov decision process model, the DQN algorithm is used to solve the problem, obtaining an optimized strategy for destroying nodes and outputting information about vulnerable nodes. The strategy for destroying nodes involves selecting actions to destroy nodes based on the current state of the agent in the Markov decision process model, aiming to maximize network failure after destroying a node.

[0109] By using node features extracted through hypergraph convolutional networks to confirm the state space and action space, the vulnerable node identification problem is transformed into a Markov decision process, thus optimizing the vulnerable node identification process.

[0110] By constructing the state space, action space, and reward function, a Markov decision process model is built, which comprehensively considers multiple key dimensions of network operation and enables overall optimization of network performance.

[0111] The agent is trained using the Deep Reinforcement Learning (DQN) algorithm to optimize the strategy for destroying nodes and output information about vulnerable nodes.

[0112] Based on the information about the vulnerable nodes, a cascading failure model is constructed to address the evolution process of cascading failures.

[0113] Based on the cascading failure model, disaster impact results are generated using vulnerable node information, including:

[0114] Based on the information of vulnerable nodes, the fault propagation path output by the initial disaster node when the fault occurs is identified through the cascaded fault model;

[0115] Based on the fault propagation path, disaster impact results are generated by determining the correlation and coupling effects between nodes;

[0116] The propagation path includes the direction of the power line's spread and the main roads of the sprawling transportation network.

[0117] Based on the disaster impact results, the correlation between the fault phenomena and the corresponding fault candidate intervals is determined through abductive reasoning network;

[0118] The abductive reasoning network adopts a hierarchical architecture design, and the network units are composed of basic low-order polynomials.

[0119] The operational steps of the abductive reasoning network to determine the correlation between fault phenomena and corresponding fault candidate intervals include:

[0120] S1, based on the disaster impact results, uses multiple regression techniques to evaluate the parameters of the function and generate the intermediate node function of the current inference iteration layer;

[0121] S2, rank the intermediate node functions based on the squared prediction error and decide which node function will enter the next inference iteration layer;

[0122] S3. Repeat steps S1 and S2 until the preset abductive reasoning network cutoff rule is met to stop the reasoning iteration.

[0123] The node function expression generated by the abductive reasoning network using multiple regression techniques is as follows:

[0124] (4);

[0125] In the formula, This represents the calculation result of the function at the intermediate node of the current inference iteration layer; , , , , , , , , , , , , and For training parameters; , and The input feature variables are obtained from multi-source operation data of the power-transportation coupled network, including power grid operation parameters, traffic flow data and environmental monitoring parameters;

[0126] The formula for calculating the squared prediction error is:

[0127] (5);

[0128] In the formula, PSE is the squared prediction error of the abductive reasoning network; FSE is the fitting variance; OP is the overfitting penalty factor; C is the complexity penalty multiplier; K is the implicit value of the complexity of the abductive reasoning network; and N is the number of training samples. Let be the variance of the error in a priori estimate of the abductive reasoning network.

[0129] Based on the aforementioned correlation, a unique mapping from the fault phenomenon to the root cause is confirmed using a Bayesian network; the Bayesian network is a directed acyclic graph based on probabilistic reasoning, and its expression is:

[0130] (6);

[0131] In the formula: Represents a directed graph; The node feature matrix, Each hyperedge contains a set of hyperedges with a number of outlier nodes greater than or equal to 2. P represents the conditional probability table on each node; B is a Bayesian network function consisting of a directed graph and a conditional probability table.

[0132] Based on the unique mapping relationship, the causal chain between the disaster mechanism, evolution path and failure phenomenon is confirmed, and the power-transportation resilience assessment results are obtained.

[0133] In summary, this embodiment utilizes abductive reasoning networks and Bayesian networks to trace the causal origins of disaster impacts, achieving a comprehensive, in-depth analysis of the entire chain of disaster node impacts, cascading propagation, and fault causes. This enables accurate diagnosis of fault causes in power-transportation coupled networks under disaster scenarios. This embodiment improves the safe operation capability of infrastructure in power-transportation coupled networks under extreme disasters, which is of significant practical importance for ensuring the stability and security of power supply.

[0134] Example 3, based on the same inventive concept as other examples, introduces a power-transportation resilience assessment device based on causal attribution, specifically including:

[0135] S1: Propose a power-transportation resilience assessment framework based on causal attribution;

[0136] S2: Establish a fault scenario generation method based on multi-hazard loss guidance;

[0137] S3: Construct an intelligent fault node identification method based on the fusion of hypergraph and DQN;

[0138] S4: Establish an integrated model for fault tracing and disaster resilience assessment;

[0139] S5: Establish an intelligent power-transportation resilience assessment device.

[0140] Among them, the causal-based power-transportation resilience assessment framework in S1 is as follows: Figure 2 As shown, it specifically includes:

[0141] S1.1 A smart scene generation method integrating multi-hazard damage characteristics is proposed. By establishing differentiated loss correction models for disasters such as typhoons, rainstorms, and high temperatures, the coupling relationship between the physical mechanism of the disaster and the disaster resistance threshold of infrastructure is quantified into a loss function, thereby guiding Generative Adversarial Networks (GANs) to generate disaster damage scenarios. This mechanism can accurately reflect the nonlinear correlation between disasters and losses, providing scenario data with realistic disaster loss evolution patterns for disaster simulation of power-transportation coupled networks.

[0142] Based on the multi-hazard failure scenario data generated in S1.1, S1.2 introduces hypergraph theory to overcome the limitations of traditional graph structures. It achieves a high-dimensional representation of the complex relationships in the power-transportation coupled network through hyperedge and hypernode modeling. Combined with the deep reinforcement learning (DQN) algorithm, a reward mechanism considering network stability and recovery efficiency is designed, enabling the agent to conduct dynamic node destruction experiments in the hypergraph topology space and accurately identify vulnerable network nodes under disaster scenarios.

[0143] Assuming a power-transportation coupled network has n nodes, and t represents time, considering multiple key dimensions of network operation, the overall network performance can be optimized by rationally constructing each sub-objective function. The multi-objective function F is constructed as follows: In the formula: This is the normalization coefficient.

[0144] Network stability : In the formula: Let i be the state weight of node i, where normal = 1, early warning = 0.5, and disaster severity = 0. This is the connectivity index of node i, which is 1 if it belongs to the maximum connected component, and 0 otherwise.

[0145] Fault recovery speed : In the formula: m is the total number of fault recovery nodes. The time taken for node j to recover from an abnormal state to normal.

[0146] Resource allocation capability : In the formula: For the degree of power grid load imbalance, The traffic network congestion index. These are the weighting coefficients.

[0147] System resilience : In the formula: D is the intensity of the disaster impact. This represents the total time it takes for the system to go from an initial failure to a stable state.

[0148] The constraints are constructed as follows:

[0149] Power grid load constraints In the formula: Let be the load of node i at time t. Let be the rated load of node i.

[0150] Traffic flow constraints In the formula: Let be the flow of node j at time t. Let be the passage capacity of node j.

[0151] Power grid load constraints In the formula: D is the intensity of the disaster impact. This represents the total time it takes for the system to go from an initial failure to a stable state.

[0152] Resource constraints In the formula: The amount of resources allocated to node i at time t. This represents the total amount of available resources in the system.

[0153] S1.3, based on the vulnerable nodes identified in S1.2, constructs a full-link tracing system covering "node impact - cascading propagation - root cause". This method uses a three-color indicator to quantify the degree of disaster impact, combined with spatiotemporal characteristics analysis of cascading propagation, specifically including scope, path, and correlation, and integrates multi-source operational data such as power grid load and traffic flow to establish a multi-dimensional analysis model encompassing both direct and indirect causes. By organically combining the disaster scenario data from S1.1, the node identification results from S1.2, and the propagation analysis of this module, a complete tracing logic from phenomenon to mechanism, from evolution to root cause, is formed, providing systematic support for network fault diagnosis and prevention decisions under coupled disasters.

[0154] like Figure 3 As shown, the multi-hazard loss-guided fault scenario generation method in step S2 includes:

[0155] S2.1 Fault Mechanism Model: To quantify the relationship between the physical processes of different natural disasters such as typhoons, rainstorms, and high temperatures and the damage mechanisms of facilities in power-transportation coupled networks, this invention establishes a fault mechanism model.

[0156] In response to typhoon disasters, this invention constructs the following typhoon failure model to calculate the comprehensive failure index of facilities under the influence of typhoons:

[0157] ;

[0158] In the formula: The overall failure index of the facility; Wind load refers to the wind pressure generated by the typhoon's wind speed; The structural vulnerability factor reflects the sensitivity of the facility's structural characteristics, such as tower height, pole type, and windward area, to wind loads. This is a function for calculating the typhoon failure index. Secondary disaster impact factors include additional damage caused by foreign object impacts due to strong winds; The overall wind resistance of a facility depends on its design standards, building material strength, and current condition.

[0159] The risk of facility failure depends not only on direct wind loads and the facility's inherent vulnerability, but also on the cumulative effects of secondary disasters. This overall disaster-causing effect is... The disaster resilience of the facility itself The game is played, and ultimately the function is determined. This is mapped to a specific failure index.

[0160] In response to rainstorm disasters, this invention constructs the following rainstorm failure model to assess the comprehensive failure index of facilities due to rainstorm flooding or related impacts:

[0161] ;

[0162] In the formula: The overall failure index of the facility; Water depth or water level is a key quantitative indicator of urban flooding caused by rainstorms. The immersion damage factor reflects the degree of performance degradation or accelerated damage to the facility caused by prolonged immersion in water. This is the function for calculating the rainstorm failure index. Insulation failure risk refers to the additional risk of failure caused by the reduced insulation performance of electrical equipment due to high humidity or water accumulation. This refers to the basic stability of the facility, the flood control capability of the substation, or the erosion resistance of the transmission tower foundation.

[0163] This model specifically describes the mechanisms by which torrential rain causes disasters through multiple pathways. The risk of failure stems from two aspects: firstly, the physical damage caused by waterlogging... Secondly, high humidity environments damage electrical insulation. The overall disaster-causing effect is related to the basic stability of the facility itself. To counteract each other, ultimately determined by the function Calculate the failure index.

[0164] To address high-temperature disasters, this invention constructs the following high-temperature failure model to evaluate the comprehensive failure index of facilities under high-temperature environments:

[0165] ;

[0166] In the formula: The overall failure index of the facility; Equipment temperature rise refers to the increase in the actual operating temperature of the equipment relative to the ambient reference temperature. This is the function for calculating the high-temperature failure index; The thermal damage coefficient reflects the effect of increased temperature on the accelerated aging or performance degradation of equipment. It is an equivalent overload factor used to quantify the effect of a decrease in the rated current carrying capacity of a line due to high temperature, thereby causing a de facto overload effect on a normal load. The heat resistance threshold of a facility is the highest temperature at which the equipment is designed to operate safely for a long period of time.

[0167] This model specifically illustrates the coupled effects of high temperatures on electrical equipment. The risk of failure stems from two parts: firstly, the aging and performance degradation directly caused by increased temperature. Secondly, the equivalent overload effect caused by the decrease in line current-carrying capacity due to high temperature. This total combined thermoelectric stress is subject to the equipment's own heat resistance threshold, i.e. The constraints are ultimately determined by the function. Provide the failure index.

[0168] S2.2 Dual-channel data processing integrates two key types of information influencing the evolution of fault scenarios: objective physical data and subjective decision-making information. Traditional methods often focus only on analyzing objective data, neglecting the crucial role of subjective factors such as human judgment and response in disaster response. The subjective-objective dual-channel architecture of this invention organically combines the two, thereby generating a more comprehensive fault scenario that closely reflects real-world decision-making processes.

[0169] The processing flow of this module can be specifically divided into the following steps:

[0170] 1) Classification and extraction of decision data.

[0171] From massive amounts of heterogeneous, multi-source raw data, we identify, extract, and categorize information into two types: subjective and objective.

[0172] Input: Historical fault data. This is a multi-source heterogeneous raw data set, including numerical data, text data, and geospatial data; the numerical data includes grids and icons, the text data includes accident reports and contingency plan documents, and the geospatial data includes GIS maps and heat maps.

[0173] Output: After classification and extraction, two types of structured information are output: Subjective decision-making data mainly consists of "judgment and response information" extracted from text and contingency plans, including expert experience, scheduling rules, and emergency response procedures. Objective decision-making data mainly consists of station and road network information data extracted from logs, GIS, and other data, including network topology, equipment physical parameters, and historical fault frequencies.

[0174] 2) Dual-channel parallel processing.

[0175] This step involves assigning the categorized decision data to the corresponding channels for in-depth processing.

[0176] Objective channel: Input: Objective decision data output from the previous step.

[0177] Processing steps: The input objective factual data is cleaned, aligned, and feature-engineered. Machine learning models such as neural networks are used to perform deep learning and mining of nonlinear relationships and complex patterns in the objective data, resulting in a precise quantitative description of the state of the objective world.

[0178] Output: Objective feature vectors. This is a structured dataset where each vector is a quantified representation of the objective state in a specific scenario.

[0179] Subjective channel:

[0180] Input: Subjective decision data output from the previous step.

[0181] Processing steps: The input subjective knowledge is formalized and digitized by constructing a knowledge graph or rule base. Models such as neural networks can be used to learn the formalized subjective knowledge, enabling it to simulate and predict expert judgment and decision-making behaviors in different scenarios.

[0182] Output: Subjective feature vectors. This forms a structured dataset where each vector is a quantified representation of subjective factors in a specific scenario, including resource scheduling priority, activation status of response strategies, etc.

[0183] 3) Data fusion.

[0184] This step is the final stage of dual-channel processing, designed to merge the results from the two parallel channels.

[0185] Input: Objective feature vectors from the objective channel and subjective feature vectors from the subjective channel.

[0186] Processing procedure: Objective feature vectors and subjective feature vectors describing the same scene or the same time step are concatenated or fused in other ways to form a composite feature vector with higher dimensions and more comprehensive information.

[0187] Output: Feature dataset. This dataset is the final output of Module 2, containing both objective facts and subjective insights. It will serve as the core input for the subsequent "scene generation algorithm" during model training.

[0188] The S2.3 scene generation algorithm, after completing data preprocessing and dual-channel feature fusion, addresses the problems of insufficient physical realism and disconnect from domain knowledge in traditional methods when generating complex disaster scenes. This invention proposes a loss-guided fault scene generation algorithm. This algorithm uses a Generative Adversarial Network (GAN) as its basic framework, but through a unique loss-guided mechanism, it deeply embeds the knowledge of the physical mechanism model into the network's training process, thereby effectively constraining the direction of scene generation.

[0189] 1) Basic framework of generative adversarial networks:

[0190] Generative Adversarial Networks (GANs) are deep learning models that have demonstrated outstanding performance in the field of data generation in recent years. They primarily consist of a generator (G) and a discriminator (D). The generator aims to learn the distribution patterns of real data and generate pseudo-data that is indistinguishable from real data; the discriminator aims to accurately distinguish whether the input data comes from real samples or is a pseudo-sample generated by the generator. Through mutual competition and adversarial training, the two eventually reach a Nash equilibrium, at which point the generator can produce high-quality data that is highly consistent with the distribution of real data.

[0191] Loss functions of the generator and discriminator in a graph generative adversarial network model and They are respectively:

[0192] );

[0193] ;

[0194] In the formula: The expected value of the sample distribution; For real samples The probability distribution; To generate samples The probability distribution, This is a fake image generated by noise Z. To minimize the function, To maximize the function, Let G be the probability that the discriminator output is a real image. Based on the above formula, we can derive the objective function of a generative adversarial network that minimizes the generator G and maximizes the discriminator D. The specific expression is:

[0195] .

[0196] 2) Loss guidance mechanism:

[0197] To address the potential disconnect between the results generated by standard GAN models and the underlying physical mechanisms, this invention designs a loss guidance mechanism. The core of this mechanism lies in introducing an additional loss guidance term driven by the physical mechanism model, based on the original loss function of the GAN. This forms a new overall loss function. The overall loss function is designed as follows:

[0198] ;

[0199] In the formula: It is a standard loss function for generative adversarial networks, used to ensure that the generated scene is similar to the real scene in terms of data distribution; The loss guide term is a weighted sum of one or more physical correction terms, used to ensure that the generated scene conforms to the physical damage laws of the disaster; This is a hyperparameter used to balance the importance of GAN loss and guidance loss.

[0200] Loss guidance item It consists of the following physical correction terms for different disasters:

[0201] 1. Typhoon Correction Items :

[0202] This correction term is used to ensure that the generated typhoon failure scenario is consistent with the laws of wind physics, and its formula is:

[0203] ;

[0204] In the formula: It is an action on facilities in the generated scene. Wind speed above, =1, 2, ...,N1, where N1 is the set of facilities in the typhoon failure scenario. This is the wind resistance threshold of the facility. Facilities Importance weights.

[0205] 2. Rainstorm Correction Items :

[0206] This correction aims to ensure that the generated rainstorm fault scenarios adhere to the inherent physical mechanisms of rainstorm-induced disasters. Its formula is:

[0207] ;

[0208] In the formula: Facilities in the generated scene The depth of the water at the location, =1, 2, ...,M1, where M1 is the set of facilities in the rainstorm failure scenario. This is the drainage capacity threshold for that location. Facilities Importance weights.

[0209] 3. High Temperature Correction Item :

[0210] To ensure the physical plausibility of the simulated high-temperature fault scenario, this correction term is introduced to constrain the generation process. Its formula is:

[0211] ;

[0212] In the formula: Devices in the generated scene Operating temperature =1, 2, ...,K1, where K1 is the set of facilities in the high-temperature fault scenario. This is the upper limit of the safe operating temperature of the equipment. It is equipment Importance weights.

[0213] 3) Algorithm flow and output:

[0214] The complete process aims to deeply integrate multi-source historical data, physical mechanism models, and subjective and objective decision-making knowledge to generate high-fidelity coupled disaster scenarios. The overall process is as follows:

[0215] Phase 1: Data Input and Preprocessing

[0216] 1. Multi-source information collection and correlation: First, the necessary data for the research is collected from meteorological and power grid departments, mainly divided into three categories: power grid information, meteorological information, and geographic information. These various types of data are then correlated geographically to lay the foundation for subsequent processing.

[0217] 2. Data cleaning and preprocessing: To improve the quality of input data, the collected data needs to be preprocessed.

[0218] 3. Decision Information Classification: The preprocessed data is classified into two information streams based on its attributes: objective decision data and subjective decision data. These streams serve as inputs for the subsequent dual-channel modules.

[0219] The second stage involves dual-channel feature processing and fusion.

[0220] Parallel channel processing: The classified data from the previous step is fed into a parallel dual-channel input for feature learning.

[0221] Objective channels: Process objective decision-making data to deeply explore its inherent, data-driven statistical regularities and nonlinear relationships.

[0222] Subjective channel: This involves formalizing and digitizing subjective decision-making data to learn and simulate the decision-making logic of experts in specific scenarios.

[0223] Data Augmentation and Fusion: To address the issues of sparse fault samples and insufficient data volume in disaster scenarios, Generative Adversarial Networks (GANs) can be used to generate fault-related data to supplement and balance the dataset. Finally, the output features from objective and subjective channels, along with the augmented data, are fused to form a fused feature dataset.

[0224] The third stage involves GAN scene generation with loss guidance:

[0225] 1. Using the fused feature dataset output from step S2.2 as input, train the dataset with the aforementioned overall loss function. Generative adversarial networks.

[0226] 2. During the training process, The project will continuously adjust the generator's parameters to ensure that the generated scenes not only statistically approximate real data but also are physically logically more reasonable.

[0227] 3. After training, a generator is used to generate a large number of diverse original fault scenarios.

[0228] 4. Perform post-processing such as data augmentation and scene compression on the generated original scene to extract typical fault modes.

[0229] 5. Finally, a series of typical fault scenario data that have been classified are output, such as distributed disasters, mobile disasters, and point-borne disasters.

[0230] like Figure 4 As shown, step S3, which constructs a fault node intelligent identification method based on the fusion of hypergraph and DQN, includes:

[0231] S3.1 Hypergraph Representation:

[0232] 1) Hypergraph Definition: This invention intends to use hypergraphs for modeling and analysis. A hypergraph can be understood as a generalization of a graph, and the set of multiple nodes in a hypergraph is called a hyperedge, denoted as […]. This section uses To represent a hypergraph, where This represents the set of nodes in the graph.

[0233] W represents the set of hyperedges where each hyperedge contains at least two out-of-class nodes, and W is the hyperedge weight matrix derived from the incidence matrix H. This is used to represent the connection relationship between a hyperedge and a node. Furthermore, This is the node feature matrix, used to represent node features.

[0234] For each node in the hypergraph Its degree matrix elements in This represents the weight of all hyperedges connecting this node. The sum can be expressed as:

[0235] ; in, , The total number of super-edges, Elements within the correlation matrix;

[0236] Hypermarginal matrix medium elements The sum of the number of nodes on this edge can be represented as:

[0237] ;

[0238] Correlation Matrix Represents the connection relationship between hyperedges and nodes, and the elements in the incidence matrix. Defined as ; that is, a node Belongs to superedge ,but It is 1 if it is not 1, otherwise it is 0.

[0239] 2) Processing procedure:

[0240] First, a fault probability matrix is ​​constructed. In this part, fault data of power grid nodes is used as input, mainly including historical fault probabilities of charging stations, charging piles, transmission networks, and distribution networks. During the construction of the fault probability matrix, this invention statistically analyzes the fault probabilities of each node under multiple disaster conditions and forms a probability distribution through normalization. The final output is the fault probability matrix. The matrix elements represent the probability of each node failing.

[0241] Secondly, the correlation matrix and feature matrix are constructed. In this part, the topology and node attribute data of the power-transportation coupled network are used as input. The correlation matrix is ​​constructed as follows. This invention defines hyperedges based on power lines or transportation networks, and constructs a feature matrix. At the same time, this invention integrates attributes such as node data, voltage level, and traffic flow.

[0242] Finally, the fault correlation matrix and fault feature matrix are constructed. In this part, the fault probability matrix... Correlation matrix and characteristic matrix As input, output fault correlation matrix. With fault feature matrix Among them, through The fault correlation matrix is ​​obtained by quantifying the strength of the fault propagation path. The fault feature matrix is ​​obtained by fusing the fault probability and node attributes.

[0243] S3.2 Feature Extraction: Hypergraph Convolutional Network (HGCN) is an extension of graph neural networks, specifically designed for processing hypergraph structured data. Unlike traditional graph structures, hyperedges in a hypergraph can connect any number of nodes simultaneously, enabling more accurate modeling of complex high-order relationships in the real world.

[0244] This section, based on a hypergraph structure, uses the fault correlation matrix and fault feature matrix as input to a hypergraph convolutional neural network. Specific features are obtained through convolution operations between nodes. Output node representation... Hyperedge representation of the fault propagation path With supplementary features.

[0245] When using HGCN for feature extraction, the first step is node representation learning. A convolution operation is performed on each node to aggregate its associated hyperedge information, resulting in a low-dimensional vector representation. The second step is hyperedge representation learning. A convolution operation is performed on each hyperedge to obtain its feature vector. The information transfer process between nodes, hyperedges, and nodes can be achieved through... express.

[0246] ;

[0247] In the formula: Representing node characteristics, Let the node degree matrix be... For learnable weights, Here are the trainable parameters, and H is the correlation matrix. It is a convolution function. To construct the characteristic matrix, T is the matrix transpose.

[0248] S3.3 Markov Decision Process (MDP) is a mathematical framework for modeling sequential decision problems in stochastic environments. Its core principle is that the future state of a system depends only on the current state and actions, and is independent of historical states.

[0249] Node representation is obtained based on feature extraction. With hyperedge representation Subsequently, this invention constructs the vulnerable node identification problem as a Markov decision process, whose inputs include node representations, hyperedge representations, and the current network topology state. The construction process of the MDP is as follows:

[0250] 1) State Space Information about the external environment in which the agent exists. The constructed state space needs to accurately describe the dynamically encoded state of catastrophe propagation. In the formula: For node propagation characteristics, For hyperedge dynamic features, This is the real-time load vector.

[0251] 2) Action Space The action an agent takes after perceiving its external environment. This action requires the agent to destroy nodes and update the network topology. In the formula: For discrete actions, the destruction node is... Let t be the agent's action at time t.

[0252] 3) Reward function : , where F is a multi-objective function, representing the feedback value obtained by the agent after perceiving the external environment and taking action.

[0253] The rewards in this section mainly include stability, fault recovery speed, resource scheduling capabilities, and resilience.

[0254] S3.4, DQN training process:

[0255] After the MDP is built, the DQN network is trained. The agent updates the network based on experience and eventually learns the strategy to destroy nodes and identify vulnerable nodes.

[0256] DQN is an improvement on the traditional deep Q-network algorithm. DQN uses a neural network to approximate the Q function, taking the state s as input and outputting the Q value of all actions.

[0257] ;

[0258] In the formula: These are the parameters of the neural network. To achieve the optimal action value function, execute action a in state s;

[0259] The goal of DQN is to make the Q-value output by the Q-network as close as possible to the ideal target Q-value. Its objective function is... for:

[0260] ;

[0261] In the formula: For the parameters of the target network, Discount rate; The parameters of the target network are The target Q value is in state Next action ; To maximize the function, Let t be the target network reward value;

[0262] Its loss function is , For neural network parameters, Let be the expected value function, and s be the current state. For the current action, For the next state, Let DQN be the objective function. The loss function represents the expected value of a sample drawn from the experience replay pool under the given distribution, and is used to measure the overall prediction error of the model under multiple samples.

[0263] like Figure 1 and Figure 5 As shown, step S4, establishing the integrated model for fault tracing and disaster resilience assessment, includes:

[0264] S4.1 Disaster Impact Assessment and Propagation Simulation:

[0265] (1) Multi-state impact assessment of disaster nodes:

[0266] Based on the hypergraph topology space proposed by S3, fault nodes are classified into three levels of decision logic by integrating multi-source operational data such as power grid load and traffic flow.

[0267] “Normal” means that the node is minimally affected by disasters and is in a normal and stable operating state. All operating indicators are within the standard threshold range, the voltage and frequency of the power grid node are stable near the rated value, and the traffic flow of the traffic node is within the normal traffic capacity range.

[0268] "Early warning" indicates that a node is affected by a certain disaster, shows abnormal signs but has not yet completely failed, some operating indicators deviate from the normal range and are in an early warning state, the load of a power grid node exceeds the rated value by a certain proportion but has not yet triggered a trip, and traffic nodes experience local congestion but can still pass slowly.

[0269] "Severe Disaster" indicates that a node has suffered a severe disaster impact, has malfunctioned or failed, its operational indicators have far exceeded safety thresholds, and it is unable to perform its functions normally. This includes power grid nodes experiencing power outages due to equipment damage, and transportation nodes being completely blocked due to bridge collapses or other reasons. Through real-time monitoring and dynamic evaluation of the status of each node, a preliminary diagnostic result of the node's status is generated.

[0270] (2) Cascade propagation analysis:

[0271] 1) Determine the scope of cascading propagation. By analyzing the connections between nodes in the hypergraph, track the nodes whose states change (from green to yellow or from yellow to red) under the impact of the disaster and their affected adjacent nodes. Statistically count the number and distribution of affected nodes to define the spatial extent of cascading propagation.

[0272] 2) Analyze the transmission path.

[0273] Fault propagation paths are divided into single-layer propagation and inter-layer propagation. Single-layer propagation paths include:

[0274] ① After a fault occurs at a node in the power grid layer, the power flow is redistributed and the various parameters of the power grid are updated. It is then determined whether a new faulty node exists after the redistribution.

[0275] ② When a charging station in the road network layer fails, some charging facilities cannot operate normally, and the charging demand of electric vehicles will concentrate at the remaining stations, resulting in a lack of coordination in the charging process. A new diversion plan should be devised for electric vehicles that need to charge during specific time periods, guiding users to other available charging stations.

[0276] Inter-layer propagation: In the positive impact mechanism, the grid node maintains the normal operation of the charging station through power transmission. Once the node fails, the connected charging station will lose power supply. In the reverse impact process, the charging facility, as an external load of the grid node, may cause the connected grid node to fail when its power demand exceeds the preset capacity limit.

[0277] 3) Cascaded Fault Model:

[0278] ① Initial impact phase:

[0279] The power-transportation coupled network initially operates in equilibrium. An initial attack is launched on a specific proportion of the nodes in the network. After removing these attacked nodes, the system power flow will be redistributed. Considering the randomness of the attack event, the cascading failure evolution process of the system can be described based on a Markov model as follows:

[0280] ;

[0281] In the formula: Let be the probability that the system stops cascading. These are the nodes that are initially removed by the system. Specifically defined as the power system in Nodes removed at any time The probability of a cascading failure.

[0282] ② Cascaded transfer stage:

[0283] When a failed node in a power system loses its ability to supply power, assuming the initial proportion of nodes retained by the transportation network according to seepage theory is a specific value, the final maximum scale of the remaining connected components after propagation through cascading failures within the communication network can be described as follows:

[0284] ;

[0285] In the formula: For communication networks in The probability of stopping the cascade at any given time. This represents the failure probability coefficient of the communication network.

[0286] ③ Cascading Interaction Stage:

[0287] After the first two stages of interaction have been completed, the fault generated in the communication network will be transmitted back to the power network. This bidirectional cascading fault will continue to propagate between the two networks, forming a positive feedback mechanism through multiple iterations. This dynamic evolution will continue until the entire interconnected system reaches a new equilibrium state or eventually collapses completely. The specific evolutionary path can be described as follows:

[0288] ;

[0289] ;

[0290] ;

[0291] ;

[0292] In the formula: The percentage of nodes retained after the initial attack on the system. This represents the state of the power system after the initial attack. This represents the state of the communication network after the initial attack on the system. This represents the number of interactions in the cascading process. This represents the remaining size of the power network after t-1 interactions. This represents the remaining size of the transportation network after t-1 interactions. Representing communication networks in The probability of stopping a cascading failure at any time. Let t be the state of the power system after t interactions. This represents the state of the communication network after t interactions.

[0293] (3) Output of results:

[0294] Based on the cascading failure model, the system identifies the initial disaster node as the source of the fault and outputs the path of its propagation to surrounding nodes, including the direction of propagation along power lines and the main arteries of the traffic network. Simultaneously, it analyzes the relationships between nodes, the coupling effects between different network nodes, the impact of power supply failures on traffic signal systems, and the counter-effects of traffic congestion on power load distribution, outputting the interdependencies and constraints between nodes during the cascading propagation process.

[0295] S4.2 Collaborative Root Cause Analysis of Faults:

[0296] (1) The process of abduction:

[0297] 1) Based on the fault range and propagation path output by S4.1, integrate the cascading propagation chain data generated during the assessment phase. Fusion of multi-source operational data covers nodal load power, line power flow distribution, equipment status monitoring parameters and historical operating records of the power system; intersection traffic flow, road segment efficiency, traffic facility operating status and signal system power supply data of the traffic system; and environmental monitoring parameters during the disaster period. Spatiotemporal registration technology is used to align the time dimension of the multi-source data, and standardization methods are employed to eliminate dimensional differences, constructing a multidimensional dataset containing node identifiers, time series, operational indicators, and environmental variables.

[0298] 2) Based on the preprocessed multidimensional dataset, the causes of failures are identified by pre-set equipment safety operation thresholds and domain rules, and they are classified and labeled as direct causes and failure causes.

[0299] ① The results of direct cause identification are stored in a structured format, including cause category, affected node, trigger time, key parameter characteristics, and degree of impact, forming an initial set of disaster-causing factors. In a power-transportation coupled network, if the proportion of the rated capacity exceeds the limit for several consecutive monitoring periods, and the key operating parameters of the associated lines deviate from the safety threshold, it is determined to be a direct cause.

[0300] ② The determination of indirect causes combines a disaster type knowledge base with a network coupling effect model, and uses spatiotemporal correlation analysis algorithms to deduce indirect disaster-causing factors. By mining the nonlinear mapping relationship between disaster propagation patterns and node failures, indirect disaster-causing factors that do not act directly are identified.

[0301] 3) Integrate the analysis results into a standardized, multi-dimensional report, including:

[0302] ① A list of direct causes, covering cause categories, affected nodes, trigger times, and abnormal characteristics of key parameters;

[0303] ② A list of indirect causes, presented in the form of a directed graph, depicting the causal transmission path and describing the logical relationship from the initial cause to the intermediate state and finally to the failure;

[0304] ③ Causal interaction analysis identifies synergistic combinations of factors contributing to disasters and reveals the coupling mechanisms of different types of causes. The output results also indicate the observational evidence and confidence level for each cause.

[0305] (2) Fault cause analysis model:

[0306] A three-dimensional correlation model of "fault source - propagation node - cause set" is constructed, and a causal reasoning network is used to establish a diagnostic system to handle the complex relationship between fault states and corresponding fault candidate regions. A Bayesian network is introduced to calculate the causal contribution of each cause, resolving ambiguity issues in scenarios with multiple co-occurring causes. For coupling between direct and indirect causes, posterior probabilities are calculated based on a historical fault case database to filter out minor factors with low confidence. For causes coupled across networks, the degree of indirect influence is quantified by defining a coupling strength index, establishing a hierarchical structure of causal relationships. Finally, an evidence weighting mechanism is used to determine the dominant cause and its path of action, forming a unique mapping relationship from fault phenomena to root causes, eliminating analytical ambiguities under the interference of multiple factors.

[0307] 1) Abductive reasoning network:

[0308] The abductive reasoning network employs a hierarchical architecture, with its network units composed of basic low-order polynomials. This multi-layered structure effectively simulates fault diagnosis scenarios for complex nonlinear systems. Notably, the network utilizes a feedforward mechanism, transmitting information through multi-layered cascaded function nodes. Various parameters within the network, including connection weights, number of nodes, and feature types, are automatically learned from training data. The algebraic form of the nodes in the abductive reasoning network is as follows:

[0309] ;

[0310] Among them, the normality factor of the abductive reasoning network is The output of the abductive reasoning network is The unitization of abductive reasoning networks The line elements of the abductive reasoning network are: .

[0311] The steps of the abductive reasoning network are as follows:

[0312] ① Node generation: Generate the intermediate node function of the j-th layer, including 1-3 input variables, and evaluate the parameters of the function through multiple regression techniques.

[0313] ② Node sorting and selection: Evaluate the intermediate node function based on the PSE value.

[0314] ③ Determine the next level: Go back to step ① and repeat the ranking and selection of intermediate node functions.

[0315] ④ Stop running: Repeat steps ① to ④, and stop repeating when the stop rule is met.

[0316] 2) Bayesian networks:

[0317] ① A Bayesian network is a directed acyclic graph based on probabilistic reasoning. It consists of a directed acyclic graph and several conditional probabilities. Mathematically, it can be expressed as: .

[0318] In a network, nodes represent variables in the universe of discourse, directed arcs represent relationships between variables, and conditional probabilities between nodes characterize the dependencies between variables; the strength of these dependencies is determined by the conditional probability values. Relationships between random variables can be expressed using Bayesian networks. Bayesian networks solve problems by satisfying conditional independence between variables, transforming complex joint probability distributions into relatively simple local distributions, thereby reducing computational complexity.

[0319] (3) Output results:

[0320] Through the causal investigation process and the operation of the fault cause analysis model, the root causes of the fault are ultimately output. Specifically, these include direct causes leading to the initial disaster node's failure, such as aging failure due to long-term equipment operation or short circuits caused by sudden heavy rainfall; and indirect causes resulting from changes in the state of each node during the cascading propagation process, such as load transfer overload caused by adjacent node failures or traffic signal system failures caused by power supply interruptions. Simultaneously, the interactions between these causes are clarified, revealing the complete causal chain from the disaster-causing mechanism to the evolution path and then to the fault phenomenon, providing accurate and comprehensive root cause information for fault diagnosis and prevention strategy optimization in power grids under coupled disasters.

[0321] like Figure 6 As shown, the power-transportation resilience assessment device in step S5 includes:

[0322] The data acquisition and processing module is used to collect historical fault data and perform data preprocessing.

[0323] The scenario generation module is used to generate typical fault scenario data of classification based on preprocessed fault history data by combining generative adversarial networks and the physical mechanism of fault disasters.

[0324] The fault location module is used to optimize the strategy for damaging nodes and output vulnerable node information based on typical fault scenario data by combining hypergraph convolutional networks and deep reinforcement learning DQN algorithm.

[0325] The causal assessment module is used to generate disaster impact results based on vulnerable node information using a cascading failure model, and based on the disaster impact results, to confirm the causal chain between disaster-causing mechanisms, evolution paths and failure phenomena through causal reasoning networks and Bayesian networks.

[0326] The visualization module is used to generate power-transportation resilience assessment results through various visualization methods based on the causal chain between disaster-causing mechanisms, evolution paths, and failure phenomena.

[0327] In summary, this invention proposes a power-transportation resilience assessment architecture based on causal tracing, constructing a full-chain analysis framework of "disaster node impact - cascading propagation - fault cause", breaking through the traditional single network assessment mode and providing a systematic logical architecture for the quantitative analysis of the resilience of coupled networks in multi-hazard scenarios.

[0328] This invention constructs a multi-hazard differentiated loss correction model, which integrates the coupling relationship between the physical mechanism of disaster damage and the disaster resistance threshold of facilities into the GAN scene generation process. By dynamically guiding the generation of multimodal data through the loss function, it achieves a nonlinear and accurate mapping between disaster intensity and system loss.

[0329] Hypergraph theory is introduced to characterize the various types of nodes and complex relationships in the power-transportation network. A multi-objective reward mechanism is constructed by combining DQN to drive the agent to dynamically identify vulnerable nodes in the hypergraph topology space. This breaks through the reliance on experience and neglect of cross-network coupling failures in traditional methods, and achieves intelligent and accurate positioning of key nodes for fault propagation under complex disasters.

[0330] An integrated model for fault tracing and resilience assessment is established. A three-mode quantitative system for node status is constructed based on hypergraph topology. Multi-source operational data such as power grid load and traffic flow are integrated to construct a fault cause analysis framework that includes direct and indirect causes, enabling cross-scale in-depth tracing from fault phenomena to disaster-causing mechanisms and evolution paths.

[0331] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0332] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0333] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0334] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0335] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A power-transportation resilience assessment method based on causal attribution, characterized in that, include: Based on the acquired historical fault data of the target area, typical fault scenario data for classification are obtained through a fault scenario generation algorithm; Based on the typical fault scenario data, topology awareness is performed through a hypergraph convolutional network to obtain node feature data; Based on the node feature data, a Markov decision process model is constructed to address the problem of identifying vulnerable nodes in a power-transport coupled network. The Markov decision process model is solved using the deep reinforcement learning (DQN) algorithm to obtain the strategy for destroying nodes and output information about vulnerable nodes. Based on the information about the vulnerable nodes, a cascading failure model is constructed to address the evolution process of cascading failures. Based on the cascaded failure model, disaster impact results are generated using information on vulnerable nodes; Based on the disaster impact results, the correlation between the fault phenomena and the corresponding fault candidate intervals is determined through abductive reasoning network; Based on the aforementioned correlation, a unique mapping relationship from the fault phenomenon to the root cause is confirmed through a Bayesian network. Based on the unique mapping relationship, the causal chain between the disaster mechanism, evolution path and failure phenomenon is confirmed, and the power-transportation resilience assessment results are obtained.

2. The power-transportation resilience assessment method based on causal attribution as described in claim 1, characterized in that, The typical fault scenario data, classified by the fault scenario generation algorithm, is obtained based on the acquired historical fault data of the target area, including: Acquire historical fault data of the target area, as well as subjective decision data and objective decision data after classifying the historical fault data; Based on the fault scenario generation algorithm, subjective decision data and objective decision data are processed in two channels, and the processed data are fused to obtain a fused feature dataset. Based on the historical fault data, a fault mechanism model is constructed to calculate the physical correction term for the fault scenario data; Based on the physical correction term and the fused feature dataset, typical fault scenario data for classification are obtained through data augmentation and scenario compression.

3. The power-transportation resilience assessment method based on causal attribution as described in claim 1, characterized in that, Based on the typical fault scenario data, topology awareness is performed using a hypergraph convolutional network to obtain node feature data, including: Based on the typical fault scenario data, a convolution operation is performed on each node to aggregate the hyperedge information associated with each node to obtain a low-dimensional vector representation. Based on the typical fault scenario data, each hyperedge is convolved to obtain the feature vector of the hyperedge; The expression for the information transfer process between nodes and hyperedges is as follows: （1）； In the formula, Representing node characteristics, Let the node degree matrix be... For learnable weight matrix, For trainable parameters, Let H be the hypermarginality matrix and H be the incidence matrix. It is a convolution function. To construct the characteristic matrix, T is the matrix transpose.

4. The power-transportation resilience assessment method based on causal attribution as described in claim 1, characterized in that, The state space of the Markov decision process model includes node propagation characteristics. Hyperedge dynamic features and real-time load vector ; The action space of the Markov decision process model is defined as the failure nodes of discrete actions. Let i represent the i-th node in the hypergraph.

5. The power-transportation resilience assessment method based on causal attribution as described in claim 4, characterized in that, The reward function of the Markov decision process model is defined as follows: （2）； （3）； in, Let F be the reward function, and F be a multi-objective function. The normalization coefficient is... For network stability, For fault recovery speed, For resource allocation capabilities, For system resilience, max is the maximization operator; The state space, action space, and reward function are determined based on the node feature data. A Markov decision process model is constructed based on the state space, action space, and reward function. Based on the Markov decision process model, the DQN algorithm is used to solve the problem, resulting in an optimized strategy for destroying nodes and outputting information on vulnerable nodes.

6. The power-transportation resilience assessment method based on causal attribution as described in claim 1, characterized in that, Based on the cascading failure model, disaster impact results are generated using vulnerable node information, including: Based on the information of vulnerable nodes, the fault propagation path output by the initial disaster node when the fault occurs is identified through the cascaded fault model; Based on the fault propagation path, disaster impact results are generated by determining the correlation and coupling effects between nodes; The propagation path includes the direction of the power line's spread and the main roads of the sprawling transportation network.

7. The power-transportation resilience assessment method based on causal attribution as described in claim 1, characterized in that, The abductive reasoning network adopts a hierarchical architecture design, and the network units are composed of basic low-order polynomials. The operational steps of the abductive reasoning network to determine the correlation between fault phenomena and corresponding fault candidate intervals include: S1, based on the disaster impact results, uses multiple regression techniques to evaluate the parameters of the function and generate the intermediate node function of the current inference iteration layer; S2, rank the intermediate node functions based on the squared prediction error and decide which node function will enter the next inference iteration layer; S3. Repeat steps S1 and S2 until the preset abductive reasoning network cutoff rule is met to stop the reasoning iteration.

8. The power-transportation resilience assessment method based on causal attribution as described in claim 7, characterized in that, The node function expression generated by the abductive reasoning network using multiple regression techniques is as follows: （4）； In the formula, This represents the calculation result of the function at the intermediate node of the current inference iteration layer; , , , , , , , , , , , , and For training parameters; , and The input feature variables are obtained from multi-source operation data of the power-transportation coupled network. The input feature variables include power grid operation parameters, traffic flow data and environmental monitoring parameters. The formula for calculating the squared prediction error is: （5）； In the formula, PSE is the squared prediction error of the abductive reasoning network; FSE is the fitting variance; OP is the overfitting penalty factor; C is the complexity penalty multiplier; K is the implicit value of the complexity of the abductive reasoning network; and N is the number of training samples. Let be the variance of the error in a priori estimate of the abductive reasoning network.

9. The power-transportation resilience assessment method based on causal attribution as described in claim 1, characterized in that, The Bayesian network is a directed acyclic graph based on probabilistic reasoning, and its expression is: （6）； In the formula, denoted as a directed graph; X is the node feature matrix; E is the set of hyperedges for each hyperedge containing at least two out-of-class nodes; P represents the conditional probability table for each node; and B is the Bayesian network function composed of the directed graph and the conditional probability table.

10. A power-transportation resilience assessment device based on causal attribution, characterized in that, include: The data acquisition and processing module is used to collect historical fault data and perform data preprocessing. The scenario generation module is used to generate typical fault scenario data of classification based on preprocessed fault history data by combining generative adversarial networks and the physical mechanism of fault disasters. The fault location module is used to optimize the strategy for damaging nodes and output vulnerable node information based on typical fault scenario data by combining hypergraph convolutional networks and deep reinforcement learning DQN algorithm. The causal assessment module is used to generate disaster impact results based on vulnerable node information using a cascading failure model, and based on the disaster impact results, to confirm the causal chain between disaster-causing mechanisms, evolution paths and failure phenomena through causal reasoning networks and Bayesian networks. The visualization module is used to generate power-transportation resilience assessment results through various visualization methods based on the causal chain between disaster-causing mechanisms, evolution paths, and failure phenomena.

Images