Disaster-based multi-network cascading failure propagation prediction method based on heterogeneous graph neural network

Abstract

The application discloses a method for predicting multi-network cascading failure propagation based on a heterogeneous graph neural network under disasters, which comprises the following steps: acquiring multi-source data of disaster, power, communication and traffic networks; constructing a four-layer heterogeneous graph model and defining multi-type inter-layer edges; establishing a mapping relationship between disaster physical quantities and physical node health states to predict initial failure states at a disaster impact moment; defining multi-scale and multi-mechanism propagation rules based on inter-layer edges to simulate a cascading failure process; constructing a space-time heterogeneous graph neural network model, which stacks a heterogeneous graph neural network module and a gated recurrent unit module, and adopts a double-head prediction layer to predict continuous operation states and discrete function states of nodes in a self-recurrence manner; and post-processing state time series output by the model to reconstruct a cross-domain cascading failure propagation path through a causal attribution algorithm. The application can accurately simulate the whole evolution process from disaster occurrence to multi-network cascading failure, and realize multi-scale failure prediction and causal analysis.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of power systems, and relates to a multi-network cascading failure propagation prediction method based on a heterogeneous graph neural network in a disaster. BACKGROUND

[0002] Urban power, communication and transportation networks are "lifeline" systems that maintain the normal operation of modern city functions. The power network provides energy supply for the communication network and electrified transportation facilities (such as traffic signals and rail transit); the communication network provides system support for the dispatching control of the power system and the information guidance of the transportation system; and the transportation network carries the physical flow of power, communication emergency repair teams and mobile energy storage resources. With the development of electrified transportation, the three networks are increasingly closely coupled in the physical, information and geographical space layers.

[0003] However, this close coupling characteristic exposes new vulnerabilities in extreme disasters (such as typhoons, heavy rains and ice disasters). Physical failure of a single system is prone to cross system boundaries and evolve into cross-domain cascading failure of multi-network systems. For example, a typhoon causes power towers to collapse, triggering a large-scale power outage; the outage causes the standby power supply of communication base stations to run out and become unavailable, causing the power grid failure positioning and dispatching command to fall into an information blind area; at the same time, the outage also causes traffic signal lights to fail, causing traffic paralysis; and traffic paralysis seriously hinders the progress of power and communication repair teams, further prolonging the failure recovery time. This negative feedback cycle of "power failure affecting communication, communication restricting recovery, and traffic hindering repair" is a major challenge that needs to be addressed in current urban resilience construction.

[0004] Currently, there are some studies on the resilience and recovery of power, communication and transportation networks. Some of the studies focus on the coupling analysis of two networks. The existing technology focuses on recovery rather than propagation prediction in the power-communication coupling aspect. In the power-transportation coupling aspect, the research is mostly focused on electric vehicle charging behavior or recovery resource scheduling, and it is difficult to depict the amplification effect of information failure on cascading failure. Another part of the research involves three networks, but its research goal is mainly focused on "recovery strategy" or "resilience improvement" after the disaster, rather than high-precision "cascading failure propagation prediction" during the disaster. The existing technology that considers the coupling of physical, information and transportation networks mostly focuses on the scheduling of mobile energy storage; or focuses on the "strategy" formulation of pre-disaster reinforcement, in-disaster regulation and post-disaster recovery; they solve the problem of "how to recover" or "how to reinforce", but lack the dynamic deduction ability of "how to propagate" failure, and cannot provide key prediction information about "who will be next, when, and why" for decision-makers during the disaster.

[0005] In summary, the prior art still has the following main defects in dealing with the safety problem of urban multi-network coupling under disaster: (1) lack of coupling propagation model integrating disaster physical quantity, power, communication, and transportation network into a unified framework; (2) most of them focus on post-disaster recovery strategy optimization, but pay little attention to the dynamic and detailed prediction of cascading failure propagation process; (3) the description of multi-scale and multi-physical propagation mechanism of power supply, control, load and impedance is insufficient; (4) lack of causal attribution ability of prediction results, which is difficult to support accurate emergency decision-making. SUMMARY

[0006] The purpose of the present application is to provide a multi-network cascading failure propagation prediction method based on heterogeneous graph neural network under disaster, based on the initial failure generated by disaster production, a four-layer heterogeneous graph model of disaster-power-transportation-communication is constructed, the attribute characteristics of nodes and four types of heterogeneous edges are defined using the idea of graph theory, a heterogeneous graph neural network is designed, the neighborhood message aggregation and transmission ability is used to deduce the cascading failure propagation path, the cause of failure is analyzed, and the cascading failure propagation prediction is realized.

[0007] The technical solution for achieving the purpose of the present application is:

[0008] A multi-network cascading failure propagation prediction method based on heterogeneous graph neural network under disaster, comprising the following steps:

[0009] Obtain multi-source data of disaster, power, communication, and transportation network, and construct a four-layer heterogeneous graph model including disaster layer, power layer, communication layer, and transportation layer , wherein is a heterogeneous graph node, the heterogeneous graph node includes a disaster node, a power node, a communication node, and a transportation node, is a heterogeneous graph edge, is a time step, and the heterogeneous graph model includes intra-layer edges and multi-type inter-layer edges;

[0010] Establish a mapping relationship between the disaster feature vector of the disaster node and the health status of the power, communication, and transportation nodes in the heterogeneous graph model to predict the initial failure state at the initial time triggered by the disaster;

[0011] According to the physical mechanism of the activated inter-layer edge type, a cross-domain cascading failure propagation rule is constructed, which includes energy-dependent propagation, information control propagation, state self-evolution propagation, and function-recovery-impedance propagation;

[0012] Construct and train a spatio-temporal heterogeneous graph neural network model, which stacks a heterogeneous graph neural network module and a gated recurrent unit module, and adopts a double-head prediction layer to predict the continuous operation state and discrete function state of the node in a self-recurrent manner, to simulate the cascading failure propagation and reconstruct the cross-domain cascading failure propagation path.

[0013] The application further discloses a multi-network cascading fault propagation prediction system based on a heterogeneous graph neural network under disasters.

[0014] The heterogeneous graph model construction module acquires multi-source data of disasters, electric power, communication and traffic networks, and constructs a four-layer heterogeneous graph model comprising a disaster layer, an electric power layer, a communication layer and a traffic layer The heterogeneous graph model comprises a disaster layer, an electric power layer, a communication layer and a traffic layer. The heterogeneous graph nodes comprise disaster nodes, electric power nodes, communication nodes and traffic nodes. The heterogeneous graph edges comprise intra-layer edges and multi-type inter-layer edges. The heterogeneous graph model comprises intra-layer edges and multi-type inter-layer edges.

[0015] The initial fault state prediction module establishes a mapping relationship between a disaster feature vector of a disaster node and health states of electric power nodes, communication nodes and traffic nodes in the heterogeneous graph model, so as to predict an initial fault state at an initial moment caused by disasters.

[0016] The cross-domain cascading fault propagation rule construction module constructs cross-domain cascading fault propagation rules according to physical mechanisms of inter-layer edge types, and the rules comprise energy supply dependence propagation, information control propagation, state self-evolution propagation and function-recovery-resistance propagation.

[0017] The cascading fault propagation prediction module constructs and trains a spatiotemporal heterogeneous graph neural network model, the spatiotemporal heterogeneous graph neural network model stacks a heterogeneous graph neural network module and a gated recurrent unit module, and adopts a double-head prediction layer to predict continuous operation states and discrete function states of nodes in a self-recurrent manner, so as to simulate cascading fault propagation and reconstruct a cross-domain cascading fault propagation path.

[0018] The application further discloses a computer storage medium, which stores a computer program, and a computer executes the computer program to implement the multi-network cascading fault propagation prediction method based on the heterogeneous graph neural network under disasters.

[0019] The application further discloses an electronic device comprising a memory and a processor, the memory stores a computer program, the processor runs the computer program stored on the memory, and the computer program is executed to implement the multi-network cascading fault propagation prediction method based on the heterogeneous graph neural network under disasters.

[0020] Compared with the prior art, the application has the following advantages:

[0021] 1. The application constructs a "disaster-multinet" coupling model closer to the physical reality. Most of the prior art constructs a double-layer network of power-communication or power-transportation, or inputs the disaster as an external static parameter. The application constructs a four-layer heterogeneous graph model including the disaster layer, power layer, communication layer and transportation layer, establishes the mapping relationship between the disaster physical quantity and the physical component integrity, overcomes the limitation of the traditional method relying on "preset fault set", realizes the endogenous and dynamic prediction from the disaster process to the system failure, and significantly improves the model accuracy and generalization ability.

[0022] 2. The application finely depicts the multi-scale and multi-mechanism fault chain propagation path. The description of the coupling relationship in the prior art is relatively simple, and most of them focus on energy supply dependence or do not distinguish the propagation mechanism. The application defines four different physical mechanisms and time scale propagation rules, so that the model can distinguish different types of edges through the heterogeneous graph neural network, and simulate the internal state evolution such as power depletion and the impedance effect of traffic congestion on repair and recovery through the time sequence module, and improves the physical authenticity and time accuracy of the chain failure prediction.

[0023] 3. The application realizes the interpretable causal attribution. The method of deep learning has the "black box" characteristic, and it is difficult to explain the root cause of the fault. Therefore, the application designs a chain failure propagation path reconstruction and causal attribution module, which backtracks the state time sequence after the prediction is completed, and outputs a clear and readable fault causal chain to the decision maker according to the physical mechanism, and improves the credibility of the prediction result and the application value in the actual emergency repair command. BRIEF DESCRIPTION OF DRAWINGS

[0024] Figure 1 The flow chart of the disaster-based heterogeneous graph neural network multi-network chain failure propagation prediction method of the embodiment;

[0025] Figure 2 The schematic diagram of the spatiotemporal heterogeneous graph neural network model of the embodiment;

[0026] Figure 3 The training flow chart of the spatiotemporal heterogeneous graph neural network model of the embodiment;

[0027] Figure 4 The chain failure propagation path reconstruction and causal attribution schematic diagram of the embodiment. DETAILED DESCRIPTION

[0028] The principle of the application is: a four-layer heterogeneous graph model including a disaster layer, a power layer, a communication layer and a traffic layer is constructed, a mapping relationship between disaster physical quantities and physical component integrity is established, the limitation of traditional methods relying on "preset fault set" is overcome, four different physical mechanisms and time scales of propagation rules are defined, so that the model can distinguish different types of edges through the heterogeneous graph neural network, and use the time sequence module to simulate internal state evolution such as power depletion, and the impedance effect of traffic congestion on repair and recovery, and after the prediction is completed, the state time sequence is traced back, and a clear and readable fault causal chain is output for decision makers according to the physical mechanism, solving the problems of unclear cross-domain cascading disaster mechanism description, lack of high-precision dynamic prediction model and difficulty in causal attribution in the prior art.

[0029] Embodiments:

[0030] As shown in Figure 1 A multi-network cascading failure propagation prediction method based on heterogeneous graph neural network under disaster, comprising the following steps:

[0031] Obtaining multi-source data of disaster, power, communication and traffic networks, and constructing a four-layer heterogeneous graph model including a disaster layer, a power layer, a communication layer and a traffic layer , wherein is a heterogeneous graph node, and the heterogeneous graph node includes a disaster node, a power node, a communication node and a traffic node, is a heterogeneous graph edge, is a time step, and the heterogeneous graph model includes intra-layer edges and multi-type inter-layer edges;

[0032] Establishing a mapping relationship between the disaster feature vector of the disaster node and the health status of the power, communication and traffic nodes in the heterogeneous graph model to predict the initial fault state at the initial time caused by the disaster;

[0033] According to the physical mechanism of the activated inter-layer edge type, a cross-domain cascading failure propagation rule is constructed, and the rule includes energy-dependent propagation, information control propagation, state self-evolution propagation and function-recovery-impedance propagation;

[0034] A spatio-temporal heterogeneous graph neural network model is constructed and trained, the spatio-temporal heterogeneous graph neural network model stacks a heterogeneous graph neural network module and a gated recurrent unit module, and adopts a double-head prediction layer to predict the continuous operation state and discrete function state in a self-recurrent manner to simulate cascading failure propagation and reconstruct the cross-domain cascading failure propagation path.

[0035] In a preferred embodiment, the multi-source heterogeneous data required by the prediction model includes at least:

[0036] Disaster data, obtaining disaster monitoring data, including real-time wind speed, rainfall and flooding depth in the disaster affected area;

[0037] Power network data, obtain power system topology, node static characteristics and dynamic operating state, static characteristics include node longitude and latitude, design disaster resistance level and rated capacity, dynamic operating state include node voltage, load rate and real-time load;

[0038] Communication network data, obtain communication network topology, node static characteristics and dynamic operating state; static characteristics include node longitude and latitude, design disaster resistance level and backup power configuration, dynamic operating state include bandwidth, time delay, packet loss rate and backup power remaining time;

[0039] Traffic network data, obtain traffic network topology, node static characteristics and dynamic operating state, static characteristics include node longitude and latitude, road level and design traffic capacity, dynamic operating state include traffic flow, congestion index and traffic signal light state.

[0040] A preferred embodiment, construct a four-layer heterogeneous graph model node set And edge set , wherein:

[0041] Disaster node set : generated by the geographical area grid affected by disaster, its node attribute includes disaster dynamic physical characteristic vector , the characteristic vector contains disaster location longitude and latitude, real-time wind speed, rainfall.

[0042] Power node set : including transformer substation, transformer, line tower and distributed power supply, its node attribute includes: static characteristics : contains longitude and latitude, design disaster resistance level, rated capacity; dynamic operating state : contains voltage, load rate, real-time load; health state vector : contains physical integrity And functional state ; physical integrity is a normalized scalar value , 1.0 means physical structure is intact, 0.0 means physical structure is completely destroyed.

[0043] Determined by the mapping relationship between the disaster characteristic vector of the disaster node and the health state of the power, communication and traffic nodes in the heterogeneous graph model, specifically a vulnerability function , which receives the real-time wind speed of the disaster physical quantity in and the design disaster resistance level in as input, and outputs the physical failure probability at time t, the formula is:

[0044]

[0045] where, is the standard normal cumulative distribution function; and are the mean and standard deviation of the lognormal distribution determined by the design disaster resistance level respectively, then .

[0046] function status , 1.0 means normal function, 0.0 means failure, is calculated as:

[0047]

[0048] where, means the function status is subject to the physical integrity; is the set of communication neighbor nodes sending control signals to the power node ; is the product operation, meaning that the failure of any critical control signal will lead to function failure.

[0049] communication node set : including base stations, fiber nodes, node attributes include: static characteristics : including latitude and longitude, design disaster resistance level, whether there is a backup power supply; dynamic operating status : including bandwidth, latency, packet loss rate, backup power supply time remaining ; health status vector : including physical integrity and function status ;

[0050] where, the definition and quantification of physical integrity are the same as ;

[0051] function status , the calculation formula is:

[0052]

[0053] where, means the function is subject to the physical integrity; is the set of power neighbor nodes supplying energy to the communication node ; means taking the value of the best function status among all main power sources; is the Heaviside step function, the minimum threshold for starting the backup power supply; the function of depends on the physical integrity, and its power supply is taken from the best state of the main power supply and the backup power supply.

[0054] a set of traffic nodes : including main intersections, charging stations, emergency centers, whose node attributes include: : static features : containing latitude and longitude, road level, designed traffic capacity; dynamic operating state : containing traffic volume, congestion index, traffic signal state; health state vector : containing physical integrity and functional state ;

[0055] wherein the definition and quantification of physical integrity are the same as ; the functional state , 1.0 represents complete unobstructed or normal function, 0.0 represents complete failure or paralysis, the calculation formula is:

[0056]

[0057] wherein, is the set of power neighbor nodes that supply power to the traffic node (such as traffic signal), and the formula represents that the functional state of the traffic node is degraded in proportion ) and depends on the functional state of the main power supply .

[0058] the node set of the four-layer heterogeneous graph model is the union of .

[0059] the inter-layer edge of the heterogeneous graph model clearly defines the functional interaction between nodes, and the edge set includes:

[0060] disaster-physical edge: from to , conveying disaster characteristics at .

[0061] power-dependent edge: and , conveying functional state .

[0062] information control edge: and , conveying functional state .

[0063] Running load edge: and ,renew and .

[0064] Function - Recovery - Impedance Side: and During the recovery phase, the dynamic operating status is transmitted. This is used to quantify the time delay of emergency repairs.

[0065] Specifically, disaster-physical edge: from disaster node Pointing to physical nodes within its geographic coverage area The directed edge has the function of transmitting the dynamic physical feature vector of the disaster at time T=1. This serves as the driving input for triggering the initial fault state prediction.

[0066] Power supply dependency edge: power-communication edge, from power node Point to communication node The directed edges (such as those of a base station) function to provide... Provides the power required for operation, and its functional status quilt Used to calculate its own functional state ; Power-transportation edge, from power node Point to traffic node The directed edges, whose function is to provide... Provides the power required for operation, and its functional status quilt Used to calculate its own functional state .

[0067] Information control edge: Communication-power edge, from communication node Point to power node The directed edge, whose function is to transmit... The status monitoring data and scheduling control instructions required for operation, and their functional status quilt Used to calculate its own functional state ; Communication-Traffic Edge, from Communication Node Point to traffic node The directed edge, whose function is to transmit... The control commands required for operation.

[0068] Running load edge: Traffic-Power edge, from traffic node Point to power node The directed edge, its function is to connect... The dynamic electrical load generated at the location is applied to and update Dynamic operating status ; Transportation-communication edge, from transportation nodes Point to communication node The directed edge, its function is to connect... The sudden surge in communication data traffic generated at the location is applied to and update Dynamic operating status .

[0069] Function-Recovery-Impedance Edge: From Traffic Node Point to geographically nearby power nodes or communication node The directed edge, whose function is to... or During the recovery phase of a functional failure, transmission Dynamic operating status As a measure of the traffic delay for the quantified emergency repair teams to reach the target node, this hinders... or Health status vector recover.

[0070] In a preferred embodiment, the step of predicting the initial fault state includes:

[0071] Disaster information dissemination: in t At any given time, only the disaster-physical edge and disaster node are activated. The physical nodes connected to it Send containing physical quantities of the disaster Disaster characteristics message ;

[0072] Initial physical integrity calculation: all received messages physical nodes , and and its own static characteristics Design disaster resistance level As input, the defined vulnerability function is called. t was calculated Initial physical integrity at time ;

[0073] Initial functional state determination: All nodes are determined based on their newly calculated state. and its neighboring nodes The function state at any given time calls the defined function state. Calculation formula to determine t Initial functional state at time ;

[0074] initial state Generation: Initial health state vector of the node and dynamic running status Updated, forming the initial fault state. The states include: by The resulting physical failure [physical integrity ≈ 0, functional state = 0], or caused by This results in [Physical integrity = 1, Functional status = 1, Remaining time of backup power] The delay function of "start reducing" fails, or the function is downgraded if "[physical integrity = 1, functional state < 1, congestion index > 0]" is activated. This is a preset physical integrity failure threshold. When the physical integrity is lower than this threshold, the node is considered to have physically failed.

[0075] A preferred embodiment, based on the initial fault state ,At Within a time step, the physical mechanisms based on inter-layer edge type activation include:

[0076] Energy supply dependency propagation mechanism: At time t, the source power node Functional status ; in t At any given moment, the model transmits the "power outage" signal along the edge to the target communication node via a message passing mechanism. or traffic nodes ; or The aggregate function receives this signal and immediately updates its internal state to trigger the defined functional state. or Recalculation.

[0077] Information control propagation mechanism: At time t, the source communication node Functional status ;t At any given moment, the model transmits a "runaway" signal to the target power node. or traffic nodes ; or The aggregation function receives this signal and updates its internal state to "out of control." The spatiotemporal heterogeneous graph neural network will learn that the "out of control" state will lead to... Its own functional state or Functional status The probability of failure increases significantly.

[0078] State self-evolution propagation mechanism: Nodes (such as communication base stations) At time t, it enters a "power outage" state, but its backup power supply... To make its functional state ; in t At any given time, without needing to communicate across nodes, the time-series state update will execute its (learned) internal state transition function. To ensure the remaining time of its backup power supply The formula for calculating the self-evolution is:

[0079]

[0080] in, for In t The dynamic operating state at any given moment; evolution continues until... At that time, the node Functional status The mutation to 0.0 triggers the information control propagation mechanism.

[0081] Function-Restoration-Impedance Propagation Mechanism: At time t, a failed power node... or communication node In the recovery phase; or The aggregation function receives data from traffic nodes via the impedance side. Dynamic operating status information; As a penalty or gating signal, extend or Physical integrity Time step required to restore to 1.0 .

[0082] In a preferred embodiment, the computation and training of the spatiotemporal heterogeneous graph neural network model includes:

[0083] Encode node features to convert heterogeneous graph node sets any node According to static attributes Health status running status Feature fusion encoding is performed, and the calculation formula is as follows:

[0084]

[0085] in, Represents the node at time t The initial fused feature vector before inputting into the heterogeneous graph neural network model; MLP is a multilayer perceptron. node type; static attribute contains node latitude and longitude, design disaster resistance level, rated capacity, road level, backup power type and other characteristics that do not change over time; represents the feature splicing operation.

[0086] Cross-network message passing using fused features, based on a defined set of inter-layer edge types , update the fused feature vector of the node by aggregating messages from neighbor nodes with different relationship types , , is the layer index of the heterogeneous graph neural network, , , is the defined set of inter-layer edge types (disaster, energy supply dependency, information control, load action, recovery resistance). is a learnable graph neural network, whose input is the source node (also a neighbor node of v), the target node , the node static feature , and the dynamic edge attribute at time t, the message construction formula is as follows:

[0087]

[0088] Using attention mechanism aggregate messages from the same relationship , and then aggregate all relationship messages by weighted sum to generate the comprehensive spatial aggregation vector at time t , is the total number of layers of the heterogeneous graph neural network, represents the last layer of node fused feature vector output after message passing through layers, through attention mechanism, the sensitivity of power nodes, communication nodes and traffic nodes to different neighbors is different due to their inherent structural characteristics.

[0089] Use gated recurrent unit GRU to model the time dimension of the node, maintain a hidden state for each node , the gated recurrent unit GRU receives the comprehensive spatial aggregation vector output by the graph neural network as the input at the current time, and receives the hidden state at the last time as historical memory, and calculates the final hidden state at time t :

[0090] ​​​

[0091] go through After the spatial and temporal updates of the layer, the final hidden state output by the GRU is received using a dual-head prediction layer. Predicting t Node state at time step For continuous states, a regression layer is used. Predict the dynamic operating state at the next moment. Including the remaining time of backup power and dynamic running status The formula is:

[0092]

[0093] For discrete states, a classification layer is used. Used to predict health status in the next moment. Including physical integrity and functional status Failure probability:

[0094]

[0095] Among them, the regression layer and classification layer All are multilayer perceptron networks. The regression layer is used to process the hidden states output by the GRU. The mapping is to continuous numerical values, and its output layer uses a linear activation function to predict the dynamic running state at the next time step; the classification layer hides the state. The probability values ​​are mapped to the interval [0, 1], and the output layer uses the Sigmoid activation function to predict the health status at the next time step.

[0096] Heterogeneous graph neural network models such as Figure 2 As shown.

[0097] The training process for heterogeneous graph neural network models is as follows: Figure 3 As shown, the steps are as follows:

[0098] Prepare a dataset of cascading failures generated from historical disasters or simulations. The dataset contains paired input sequences. and target sequence ( ); For node status, To characterize The adjacency matrix of the heterogeneous graph model's topology at time step, which corresponds to the heterogeneous graph model. dynamic edge set in .

[0099] Constructing a weighted combination loss function Used at each time step The loss function minimizes the error between the predicted value and the true target value. Including regression loss of continuous state heads The classification loss of the discrete event header. The calculation formula is:

[0100]

[0101] in, and They are respectively Predicted values ​​of dynamic operating status and health status in real time; and This is the true target value; Let the mean squared error loss function be used. The cross-entropy loss function; and These are the trainable loss weight coefficients;

[0102] The model optimization employs the backpropagation algorithm to minimize the cumulative loss across all time steps. To achieve the goal, optimize all learnable parameters of the model (including ).

[0103] In a preferred embodiment, the spatiotemporal heterogeneous graph neural network model receives the node state at time t. And graph topology By aggregating spatial neighbor information and updating temporal states, t is predicted. Node state at time step To simulate the propagation of cascading failures; the state-time series output by the model. Post-processing is performed to reconstruct the cascading failure propagation path and perform causal attribution.

[0104] Specifically, such as Figure 4 As shown, it includes:

[0105] Identify failure events: Traverse the health status time series output by the dual-head prediction layer Identify nodes that undergo functional state mutations. and its failure time Mutation is defined as a node Satisfy its functional state The predicted value is intact or the probability is higher than the threshold, while The predicted value is invalid or the probability is higher than the threshold.

[0106] Causal attribution identification: for in Every node that fails at any time Check them in order of priority. The time point, the state of the heterogeneous neighbor node and the state of itself. The rules are: (1) disaster impact attribution: if , the failure of node is attributed to the initial impact transmitted by the "disaster-physical edge"; (2) energy supply dependence attribution: if , check the source power node connected by the "energy supply dependence edge"; if satisfies , the failure of node is attributed to "energy supply interruption"; (3) information control attribution: if (2) is not satisfied, check the source communication node connected by the "information control edge"; if satisfies , the failure of node is attributed to "information control failure"; (4) state self-evolution attribution: if (2) and (3) are not satisfied, check the continuous state of node at the time point ; if is a communication node and the remaining time of its backup power supply , the failure is attributed to "backup power depletion"; (5) running load attribution: if (2), (3) and (4) are not satisfied, check the source traffic node connected by the "running load edge"; if the dynamic running state of at the time point exceeds the preset load threshold , causing the dynamic running state of to exceed the limit, it is attributed to "load impact".

[0107] Path reconstruction and output: aggregate all identified failure events (nodes , ) and their attribution results (reasons ), construct and output a time sequence cascading failure directed graph , The directed edge in the directed graph represents the causal relationship from the reason to the failure node .

[0108] Another embodiment, a multi-network cascading failure propagation prediction system based on heterogeneous graph neural network under disaster, for realizing the above-mentioned multi-network cascading failure propagation prediction method based on heterogeneous graph neural network under disaster, comprising:

[0109] The heterogeneous graph model construction module acquires multi-source data of disasters, power, communication, and traffic networks, and constructs a four-layer heterogeneous graph model including a disaster layer, a power layer, a communication layer, and a traffic layer , wherein is a heterogeneous graph node, and the heterogeneous graph node includes a disaster node, a power node, a communication node, and a traffic node, is a heterogeneous graph edge, is a time step, and the heterogeneous graph model includes intra-layer edges and multi-type inter-layer edges;

[0110] The initial fault state prediction module establishes a mapping relationship between a disaster feature vector of the disaster node and health states of power, communication, and traffic nodes in the heterogeneous graph model, to predict an initial fault state at an initial time caused by the disaster;

[0111] The cross-domain cascading fault propagation rule construction module constructs a cross-domain cascading fault propagation rule according to a physical mechanism of inter-layer edge type activation, and the rule includes energy supply dependence propagation, information control propagation, state self-evolution propagation, and function-recovery-resistance propagation;

[0112] The cascading fault propagation prediction module constructs and trains a spatio-temporal heterogeneous graph neural network model, the spatio-temporal heterogeneous graph neural network model stacks a heterogeneous graph neural network module and a gated recurrent unit module, and adopts a double-head prediction layer to predict node continuous operation states and discrete function states in a self-recurrent manner, to simulate cascading fault propagation and reconstruct a cross-domain cascading fault propagation path.

[0113] The specific implementation method adopts the above method, which will not be described here.

[0114] The working process of the multi-network cascading fault propagation prediction system based on the heterogeneous graph neural network under disasters will be described below by taking a specific example.

[0115] Taking a typhoon invasion of a city as an example, it is assumed that the heterogeneous graph model includes the following five key nodes:

[0116] Disaster node : typhoon center and grid area covered thereby.

[0117] Power node P1 : supplies power to the region, and has a high anti-disaster level.

[0118] Power node P2 : connected to P1, supplies power downstream, and has a low anti-disaster level (easily blown down by the wind).

[0119] Communication node C1: responsible for transmitting dispatching instructions of the P1 substation, relies on P2 for power supply, and has a battery (backup power supply).

[0120] Traffic node T1: Traffic signal, located at a key intersection, relies on P2 for power.

[0121] The system has collected data from external sensors at time T=0:

[0122] Real-time wind speed in the region (strong typhoon).

[0123] Design disaster resistance level of the tower (weak wind resistance), P1 substation is operating normally.

[0124] The base station's main power supply is currently normal, and the backup battery has hours of power.

[0125] The traffic light is working normally, and the intersection congestion index is 0.2 (smooth).

[0126] Define heterogeneous edges:

[0127] Intra-layer edges: (current flows from).

[0128] Disaster-physical edges: (typhoon directly blows towards the tower).

[0129] Energy-dependent edges: (tower provides power to base station), (tower provides power to traffic light).

[0130] Information control edges: (base station sends control signals back to substation).

[0131] Typhoon lands, activates "disaster-physical edge", calls function, result (95% probability of failure), then physical integrity Since 0.05 < threshold, determine tower physically broken.

[0132] Update initial state, P2 physically damaged, function failed ), other nodes (P1, C1, T1) are still physically intact at T=1.

[0133] Use heterogeneous graph neural networks for cascading failure propagation and deduction:

[0134] Aggregate and pass messages from each node's neighbor nodes. For example: T1's power comes from P2, because , T1 red light extinguished (traffic paralysis); C1 power from P2, but C1 has a backup battery, C1 base station although the main power is broken, but relying on the battery to continue to work, the function state The attention heterogeneous aggregation is performed on the neighbor message received for each node to obtain a comprehensive spatial aggregation vector aggregated with power, traffic, and communication information The GRU module updates the comprehensive vector according to the time step The time step cycle deduction is performed.

[0135] After the model deduction is completed, backtracking is performed, and causality is attributed:

[0136] Why does P1 substation (T=7) lose control? Check the neighbors of T=6, and find that the signal from C1 is broken, and the attribution is that information control fails; why does C1 base station (T=6) fail? Check the state of T=5, which is physically intact, but the backup power is zeroed because T=1 does not supply power to it, and the attribution is that backup power is exhausted due to power interruption; why does T1 red light (T=2) interrupt? Check the neighbors of T=1, P2 power failure, and the attribution is power interruption; why does P2 tower (T=1) damage? Check the input of T=1, the wind speed > design level, and the attribution is disaster physical impact.

[0137] Another embodiment is a computer storage medium having a computer program stored thereon, and the computer program is executed by a computer to implement the multi-network cascading failure propagation prediction method based on a heterogeneous graph neural network in a disaster described in any one of the above embodiments.

[0138] The specific implementation method adopts the above method, and details are not described herein.

[0139] Another embodiment is an electronic device including a memory and a processor, the memory has a computer program stored therein, and the processor runs the computer program stored on the memory, and the computer program is executed to implement the multi-network cascading failure propagation prediction method based on a heterogeneous graph neural network in a disaster described in any one of the above embodiments.

[0140] The specific implementation method adopts the above method, and details are not described herein.

[0141] The above embodiments are preferred embodiments of the present application, but the embodiments of the present application are not limited by the above embodiments, and any changes, modifications, substitutions, combinations, simplifications made without departing from the spirit and principles of the present application should be equivalent replacement methods, and all are included in the protection scope of the present application.

Claims

1. A method for predicting the propagation of multi-network cascading faults under disasters based on heterogeneous graph neural networks, characterized in that, Includes the following steps: Acquire multi-source data on disasters, power, communication, and transportation networks, and construct a four-layer heterogeneous graph model comprising disaster, power, communication, and transportation layers. ,in These are nodes in a heterogeneous graph, which include disaster nodes, power nodes, communication nodes, and transportation nodes. For edges of a heterogeneous graph, For time steps, the heterogeneous graph model includes intra-layer edges and multiple types of inter-layer edges; the multiple types of inter-layer edges include disaster-physical edges: from disaster nodes... Pointing to physical nodes within its geographic coverage area The directed edge, whose function is to transmit the dynamic physical feature vector of the disaster at time t=1. , as the driving input that triggers the initial fault state prediction; Establish a mapping relationship between the disaster feature vector of disaster nodes and the health status of power, communication and transportation nodes in the heterogeneous graph model, so as to predict the initial fault state at the initial moment caused by the disaster; Predicting the initial fault state at the initial moment triggered by the disaster includes: exist At any given time, only the disaster-physical edge and disaster node are activated. The physical nodes connected to it Sending information including real-time wind speed Disaster characteristics message ; All received disaster characteristic messages physical nodes , and and its own static characteristics Design disaster resistance level As input, the calculation is obtained Initial physical integrity at time ; All nodes are based on their newly calculated... and its neighboring nodes The functional state at any given time is determined according to the respective functional state calculation formula. Initial functional state at time ; Update the initial health state vector of the node and dynamic running status This forms the initial fault state. The initial fault state includes: by The resulting physical failure [physical integrity ≈ 0, functional state = 0], or caused by This results in [Physical integrity = 1, Functional status = 1, Remaining time of backup power] The delay function of "start reducing" fails, or the function is downgraded if "[physical integrity = 1, functional state < 1, congestion index > 0]" is downgraded. The physical integrity failure threshold is set as the preset threshold. When the physical integrity is lower than this threshold, the node is determined to have physically failed. Based on the physical mechanism of inter-layer edge type activation, cross-domain cascading fault propagation rules are constructed, including power supply dependency propagation, information control propagation, state self-evolution propagation, and function-recovery-impedance propagation. A spatiotemporal heterogeneous graph neural network model is constructed and trained. The spatiotemporal heterogeneous graph neural network model stacks heterogeneous graph neural network modules and gated recurrent unit modules, and uses a dual-head prediction layer to predict the continuous operating state and discrete functional state of nodes in an autoregressive manner to simulate the propagation of cascading faults and reconstruct the cross-domain cascading fault propagation path.

2. The method for predicting multi-network cascading fault propagation under disasters based on heterogeneous graph neural networks according to claim 1, characterized in that, The acquired multi-source data includes: Disaster data, including acquired disaster monitoring data, which includes real-time wind speed, rainfall, and inundation depth in the disaster-affected areas; Power network data includes the acquired power system topology, node static characteristics, and dynamic operating status. The static characteristics include node latitude and longitude, design disaster resistance level, and rated capacity. The dynamic operating status includes node voltage, load rate, and real-time load. Communication network data includes the acquired communication network topology, node static characteristics, and dynamic operating status; the static characteristics include node latitude and longitude, design disaster resistance level, and backup power configuration, and the dynamic operating status includes bandwidth, latency, packet loss rate, and remaining backup power time; Traffic network data includes the acquired traffic network topology, static node characteristics, and dynamic operating status. The static characteristics include node latitude and longitude, road grade, and design capacity. The dynamic operating status includes traffic flow, congestion index, and traffic light status.

3. The method for predicting multi-network cascading fault propagation under disasters based on heterogeneous graph neural networks according to claim 2, characterized in that, disaster node set Generated by a gridded representation of the geographical area affected by the disaster, its node attributes include dynamic physical feature vectors of the disaster. The feature vector includes the latitude and longitude of the disaster location, real-time wind speed, and rainfall. Power node set Including substations, transformers, transmission line towers, and distributed power sources, their node attributes include: static characteristics. Dynamic operating status and health status vector ; The static features Includes latitude and longitude, design disaster resistance level, and rated capacity; The dynamic operating state Includes voltage, load rate, and real-time load; The health status vector Including physical integrity and functional status The physical integrity refers to the degree of integrity of the physical structure. The functional state 1.0 indicates normal function, and 0.0 indicates function failure. The calculation formula is: in, This indicates that the functional state is constrained by the physical integrity of the system. To power nodes Communication neighbor node that sends control signals A set; For product operations, it represents any key control signal. All failures will lead to Functionality failed; Communication node set This includes base stations and fiber optic nodes, whose node attributes include: static characteristics. Dynamic operating status and health status vector ; The static features This includes latitude and longitude, design disaster resistance level, and whether there is a backup power supply; The dynamic operating state Includes bandwidth, latency, packet loss rate, and remaining backup power time. ; The health status vector Including physical integrity and functional status The functional state The calculation formula is: in, This indicates that functionality is constrained by physical integrity. To communicate with nodes Power supply neighbor nodes A set; This indicates that the value is taken from the one with the best functional state among all main power supplies; Let Hevyside step function be the function of the step function. The minimum threshold for starting up a backup power supply; Operation representation Its functionality depends on its physical integrity, and its power source is the best of the two power sources: the primary power source and the backup power source. Traffic node set Including major intersections, charging stations, and emergency centers, their node attributes include: static characteristics. Dynamic operating status and health status vector ; The static features Includes latitude and longitude, road grade, and design capacity; The dynamic operating state Includes traffic flow, congestion index, and traffic light status; The health status vector Including physical integrity and functional status The functional state 1.0 indicates fully functional or working properly, while 0.0 indicates complete failure or malfunction. The calculation formula is: in, To traffic nodes Power supply neighbor nodes gather.

4. The method for predicting multi-network cascading fault propagation under disasters based on heterogeneous graph neural networks according to claim 3, characterized in that, Multi-type inter-layer edges also include: Power supply dependency edge: power-communication edge, from power node Point to communication node The directed edges, whose function is to provide... Provides the power required for operation, and its functional status quilt Used to calculate its own functional state ; Power-transportation edge, from power node Point to traffic node The directed edges, whose function is to provide... Provides the power required for operation, and its functional status quilt Used to calculate its own functional state ; Information control edge: Communication-power edge, from communication node Point to power node The directed edge, whose function is to transmit... The status monitoring data and scheduling control instructions required for operation, and their functional status quilt Used to calculate its own functional state ; Communication-Traffic Edge, from Communication Node Point to traffic node The directed edge, whose function is to transmit... The control commands required for operation; Running load edge: Traffic-Power edge, from traffic node Point to power node The directed edge, its function is to connect... The dynamic electrical load generated at the location is applied to and update Dynamic operating status ; Transportation-communication edge, from transportation nodes Point to communication node The directed edge, its function is to connect... The sudden surge in communication data traffic generated at the location is applied to and update Dynamic operating status ; Function-Recovery-Impedance Edge: From Traffic Node Point to geographically nearby power nodes or communication node The directed edge, whose function is to... or During the recovery phase of a functional failure, transmission Dynamic operating status As a measure of the traffic delay for the quantified emergency repair teams to reach the target node, this hinders... or Health status vector recover.

5. The method for predicting multi-network cascading fault propagation under disasters based on heterogeneous graph neural networks according to claim 4, characterized in that, The propagation mechanism of energy dependence includes: At any given moment, the source power node Functional status ;exist At any given moment, the power outage signal is transmitted along the edge to the target communication node via a message passing mechanism. or traffic nodes ; or The aggregation function receives this signal and immediately updates its internal state to trigger its functional state. or Recalculation; The information control and propagation mechanism includes: At any given time, the source communication node Functional status ; At any moment, transmit the out-of-control signal to the target power node. or traffic nodes ; or The aggregation function receives this signal and updates its internal state to out of control. The spatiotemporal heterogeneous graph neural network model learns the out-of-control state, thus... Its own functional state or Functional status The probability of failure increases significantly; The state self-evolution propagation mechanism includes: communication nodes If the power outage occurs at time t, the remaining time of its backup power supply is... To make its functional state ; in t At any given time, without the need for cross-node messages, the temporal state update module of the spatiotemporal heterogeneous graph neural network model executes its learned internal state transition function. To ensure the remaining time of its backup power supply The formula for calculating the self-evolution is: in, for exist The dynamic operating state at any given moment; evolution continues until... At that time, the communication node Functional status The mutation to 0.0 triggers the aforementioned information control propagation mechanism; The function-recovery-impedance propagation mechanism includes: at time t, a failed power node or communication node In the recovery phase; or The aggregation function receives data from traffic nodes via the impedance side. Dynamic operating status information; As a penalty or gating signal, extend or Physical integrity Time step required to restore to 1.0 .

6. The method for predicting multi-network cascading fault propagation under disasters based on heterogeneous graph neural networks according to claim 1, characterized in that, The computation of the spatiotemporal heterogeneous graph neural network model includes: heterogeneous graph node set any node According to static attributes Health status running status Feature fusion encoding is performed, and the calculation formula is as follows: in, Represents the node at time t The initial fused feature vector before inputting into the heterogeneous graph neural network model; MLP is a multilayer perceptron. Node type; static property Includes node latitude and longitude, design disaster resistance level, rated capacity, road grade, and backup power supply type; Indicates feature concatenation operation; Cross-network message passing is performed using encoded fusion features, based on a defined set of inter-layer edge types. Through aggregation and nodes There are different types of relationships neighboring nodes News This updates the node. fusion feature vector , Here is the layer index of the heterogeneous graph neural network; the message construction formula is as follows: For learnable neural networks, for The neighboring nodes, which are also the source nodes that send information, can be used as inputs to the learnable neural network, including the source nodes that send messages. fusion feature vector The target node as the message receiver fusion feature vector ,node , static features and the dynamic edge attributes at time t ; Employing attention mechanisms Aggregation from the same relation The messages are then aggregated by weighted summation of all related messages to generate a comprehensive spatial aggregation vector at time t. , This represents the total number of layers in a heterogeneous graph neural network. Indicates the process The final layer node fusion feature vector output after layer message passing uses an attention mechanism to make the sensitivity of power nodes, communication nodes, and transportation nodes to different neighbors different due to their inherent structural characteristics. Gated recurrent units (GRUs) are used to model the nodes in the time dimension, for each node. Maintain a hidden state Aggregate vectors in the integrated space It serves as the input to the current time step of the gated loop unit and receives the input from the previous time step t. Hidden state As a historical memory, the final hidden state at time t is calculated. : go through After the spatial and temporal updates of the layer, the final hidden state output by the GRU is received using a dual-head prediction layer. Predicting t Node state at time step ; For continuous states, use a regression layer Predicting the dynamic operating state at the next moment Including the remaining time of backup power and dynamic running status The formula is: For discrete states, a classification layer is used. Predicting health status in the next moment Including physical integrity and functional status The failure probability is given by the formula: 。 7. The method for predicting multi-network cascading fault propagation under disasters based on heterogeneous graph neural networks according to claim 1, characterized in that, Training a spatiotemporal heterogeneous graph neural network model includes: Construct a cascading failure dataset generated from historical disasters or simulations, wherein the dataset contains paired input sequences. and target sequence ( ); For node status, To characterize The adjacency matrix of the heterogeneous graph model's topology at time step, which corresponds to the heterogeneous graph model. dynamic edge set in ; Constructing a weighted combination loss function Used at each time step The loss function minimizes the error between the predicted value and the true target value. Regression loss including continuous state heads of the double-headed prediction layer Classification loss of discrete event heads in a dual-head prediction layer The calculation formula is: in, and These are the predicted values ​​of the dynamic operating status and health status at time t, respectively. and These represent the actual target values ​​for the dynamic running state and health state at time t, respectively. Let the mean squared error loss function be used. The cross-entropy loss function; and These are the trainable loss weight coefficients; The spatiotemporal heterogeneous graph neural network model optimization employs the backpropagation algorithm to minimize the cumulative loss over all time steps N. The goal is to optimize all learnable parameters of a spatiotemporal heterogeneous graph neural network model.

8. The method for predicting multi-network cascading fault propagation under disasters based on heterogeneous graph neural networks according to claim 4, characterized in that, Reconstructing cross-domain cascading failure propagation paths includes: Traversing the health status time series output by the dual-head prediction layer Identify nodes that undergo functional state mutations. and its failure time Mutation is defined as a node Satisfy its functional state The predicted value is intact or the probability is higher than the threshold, while The predicted value is invalid or the probability is higher than the threshold; For in Every node that fails at any time Check them in order of priority. Causal attribution is performed based on the state of neighboring nodes and the state of the node itself at time t; the causal attribution rules are as follows: (1) Disaster impact attribution: if Then the node The failure is attributed to the initial shock transmitted from the disaster-physical edge; (2) Energy supply dependence attribution: if Check the source power nodes connected to the power supply dependency edge. ;like satisfy Then the node The failure is attributed to power supply interruption; (3) Information control attribution: if (2) is not satisfied, check the source communication node connected to the information control edge. ;like satisfy Then the node Failure attribution is due to information control failure; (4) State self-evolution attribution: if (2) and (3) are not satisfied, check the node. In itself Continuous state at time step ;like For communication nodes And its backup power remaining time If the failure is not satisfied, the failure will be attributed to the depletion of backup power; (5) Operating load attribution: If (2), (3) and (4) are not satisfied, check the source traffic node connected to the operating load edge. ;like exist Dynamic load at any time Exceeding the preset load threshold ,lead to Dynamic operating status If the limit is exceeded, it is attributed to load surge; Aggregate all identified failure events and their attribution results to construct and output a directed graph of cascading failures. , The directed edges in the diagram represent the path from the cause to the failure node. The causal relationship.

9. A multi-network cascading fault propagation prediction system based on heterogeneous graph neural networks under disaster conditions, characterized in that, A method for predicting multi-network cascading fault propagation under disasters based on heterogeneous graph neural networks as described in any one of claims 1-8, comprising: The heterogeneous graph model construction module acquires multi-source data from disaster, power, communication, and transportation networks to construct a four-layer heterogeneous graph model comprising disaster, power, communication, and transportation layers. ,in These are nodes in a heterogeneous graph, which include disaster nodes, power nodes, communication nodes, and transportation nodes. For edges of a heterogeneous graph, For time steps, the heterogeneous graph model includes intra-layer edges and multiple types of inter-layer edges; The initial fault state prediction module establishes a mapping relationship between the disaster feature vector of the disaster node and the health status of power, communication and transportation nodes in the heterogeneous graph model, so as to predict the initial fault state at the initial moment caused by the disaster. The cross-domain cascading fault propagation rule construction module constructs cross-domain cascading fault propagation rules based on the physical mechanism of inter-layer edge type activation. The rules include power supply dependency propagation, information control propagation, state self-evolution propagation, and function-recovery-impedance propagation. The cascading failure propagation prediction module constructs and trains a spatiotemporal heterogeneous graph neural network model. The spatiotemporal heterogeneous graph neural network model stacks a heterogeneous graph neural network module and a gated recurrent unit module, and uses a dual-head prediction layer to predict the continuous operating state and discrete functional state of nodes in an autoregressive manner to simulate cascading failure propagation and reconstruct cross-domain cascading failure propagation paths.

10. A computer storage medium having a computer program stored thereon, characterized in that, When the computer executes the computer program, it implements the multi-network cascading fault propagation prediction method based on heterogeneous graph neural networks under disasters as described in any one of claims 1-8.

11. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor runs the computer program stored in the memory. When the computer program is executed, it implements the multi-network cascading fault propagation prediction method based on heterogeneous graph neural networks under disasters as described in any one of claims 1-8.

Images