Graph neural network-based disaster chain risk early warning method and system for cross-zone distribution network

Abstract

The application discloses a cross-regional power distribution network disaster chain risk early warning method and system based on a graph neural network, and relates to the technical field of power system safety operation and maintenance. The method comprises the following steps: collecting multi-source management data of a cross-regional power distribution network, constructing a multi-dimensional feature vector of a plurality of electric power nodes and initial weight parameters of each associated edge; constructing a heterogeneous space-time graph model, performing message passing and time dimension modeling through a graph neural network, and obtaining a risk amplification factor of each node and a plurality of disaster chain transmission paths; combining the risk amplification factor to perform multi-scale parallel early warning and spatiotemporal fusion on the results to obtain a comprehensive risk matrix; combining the comprehensive risk matrix, the space-time embedding representation and the disaster chain transmission path, calculating an anomaly score through an encoder network, determining an early warning level according to the anomaly score, and outputting a risk early warning result. The application realizes dynamic transmission modeling and multi-scale early warning of disaster chain risk, and improves the accuracy of risk identification and the adaptability of early warning.

Description

Technical Field

[0001] This invention relates to the field of power system safety operation and maintenance technology, specifically to a method and system for early warning of disaster chain risks in cross-regional distribution networks based on graph neural networks. Background Technology

[0002] With the high proportion of new energy sources being integrated into the power grid, urban distribution networks at the receiving end exhibit characteristics of high coupling between power sources, grids, and loads, and multiple overlapping risks. When local disturbances occur, they may be amplified step by step through the power grid topology, forming a cross-regional disaster chain transmission phenomenon, which seriously threatens the safe and stable operation of the power grid.

[0003] Existing risk early warning methods for power distribution networks can be mainly divided into three categories. The first category is based on offline simulation, which struggles to adapt to changes in the real-time operating status of the power grid. The second category is based on empirical thresholds, using fixed thresholds for anomaly detection, resulting in high false alarm and false negative rates across different scenarios. The third category is based on deep learning, but it still suffers from the following shortcomings: insufficient characterization of the dynamic transmission and amplification mechanisms of disaster chains; weak ability to model the coupling of multiple risks such as overload and voltage exceedances; and the inability to adaptively adjust the early warning threshold, leading to low reliability of early warnings under complex disturbance scenarios. Therefore, there is an urgent need for a method that can characterize the dynamic transmission of risks, integrate multi-dimensional coupled risks, and achieve adaptive threshold early warning. Summary of the Invention

[0004] This invention addresses the technical problems in existing technologies, such as insufficient characterization of the dynamic transmission and amplification mechanism of disaster chains, weak modeling ability of multiple risk coupling, and inability to adaptively adjust early warning thresholds. It provides a disaster chain risk early warning method and system for cross-regional distribution networks based on graph neural networks.

[0005] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: In a first aspect, the present invention provides a method for early warning of disaster chain risks in cross-regional power distribution networks based on graph neural networks, including: Multi-source management data of cross-regional power distribution networks are collected through multiple data sources. The multi-source management data is preprocessed, and based on the preprocessing results, multi-dimensional feature vectors of multiple power element nodes and initial weight parameters of each associated edge are constructed. Based on the distribution network topology in the multi-source management data, a heterogeneous spatiotemporal graph model is constructed by combining the multi-dimensional feature vector and the initial weight parameters. Message passing and time dimension modeling are performed through graph neural networks to obtain the risk amplification factor of each power element node and the transmission path of multiple target disaster chains. By combining the risk amplification factor with the multi-source management data, multi-scale parallel early warning is performed on multiple target disaster chain transmission paths, and the results of the multi-scale parallel early warning are spatiotemporally fused to obtain a comprehensive risk matrix; Combining the comprehensive risk matrix, the spatiotemporal embedding representation output by the graph neural network, and the target disaster chain transmission path, anomaly scores are calculated through a pre-constructed encoder network, and the warning level is determined based on the anomaly scores, outputting risk warning results.

[0006] Secondly, this invention provides a cross-regional power distribution network disaster chain risk early warning system based on graph neural networks, including: The data acquisition module is used to collect multi-source management data of cross-regional power distribution networks through multiple data sources, preprocess the multi-source management data, and construct multi-dimensional feature vectors of multiple power element nodes and initial weight parameters of each associated edge based on the preprocessing results. The heterogeneous spatiotemporal graph model module is used to construct a heterogeneous spatiotemporal graph model based on the distribution network topology in the multi-source management data, combined with the multi-dimensional feature vector and the initial weight parameters, and to obtain the risk amplification factor of each power element node and the transmission path of multiple target disaster chains by using a graph neural network for message passing and time dimension modeling. The multi-scale early warning fusion module is used to combine the risk amplification factor with the multi-source management data to perform multi-scale parallel early warning for multiple target disaster chain transmission paths, and to perform spatiotemporal fusion of the multi-scale parallel early warning results to obtain a comprehensive risk matrix. The early warning level determination module is used to combine the comprehensive risk matrix, the spatiotemporal embedding representation output by the graph neural network, and the target disaster chain transmission path, calculate the anomaly score through the pre-constructed encoder network, determine the early warning level based on the anomaly score, and output the risk early warning result.

[0007] The beneficial effects of this invention are: Compared to existing technologies, this invention first constructs a heterogeneous spatiotemporal graph model comprising five types of nodes (substations, lines, loads, distributed power sources, and protection devices) and four types of edges (electrical connections, information communication, geographical proximity, and logical dependencies), thus characterizing differentiated risk transmission paths and solving the problem of isomorphic graphs being unable to distinguish transmission semantics. Secondly, it employs a heterogeneous attention message passing mechanism, combined with the time decay kernel encoding delay effect specific to edge types, and quantifies the key transmission positions of nodes in the disaster chain through a risk amplification factor, achieving accurate extraction of disaster chain paths. Thirdly, it establishes a four-channel parallel early warning framework ranging from minutes to weeks. After spatiotemporal fusion, the input is a variational autoencoder to calculate anomaly scores. A triple adaptive threshold system—sliding window statistics, dual-window comparison detection, and quantile dynamic calibration—is used to determine the early warning level, solving the problem of high false alarm and false negative rates caused by fixed thresholds. Finally, it achieves multi-scale dynamic early warning and key transmission node identification for cross-regional distribution network disaster chain risks. Attached Figure Description

[0008] Figure 1A flowchart illustrating the cross-regional power distribution network disaster chain risk early warning method based on graph neural networks provided by this invention; Figure 2 This is a schematic diagram of the structure of the cross-regional power distribution network disaster chain risk early warning system based on graph neural networks provided by the present invention.

[0009] In the attached diagram, the components represented by each number are as follows: Data acquisition module 11, heterogeneous spatiotemporal graph model module 12, multi-scale early warning fusion module 13, and early warning level determination module 14. Detailed Implementation

[0010] Example 1, as Figure 1 As shown, this embodiment of the invention provides a method for early warning of disaster chain risks in cross-regional power distribution networks based on graph neural networks, including: S10: Collect multi-source management data of cross-regional power distribution networks through multi-source data sources, preprocess the multi-source management data, and construct multi-dimensional feature vectors of multiple power element nodes and initial weight parameters of each associated edge based on the preprocessing results; First, multi-source management data for inter-regional distribution networks is collected through multiple data sources. Inter-regional distribution networks are medium- and low-voltage distribution network systems connecting multiple administrative regions or power supply zones, responsible for distributing electrical energy from substations to end users. These multi-source data sources include distribution network dispatch automation systems, meteorological monitoring and forecasting systems, equipment asset management systems, and online equipment status monitoring systems. The distribution network dispatch automation system provides electrical measurement data, including active power, reactive power, voltage amplitude, current amplitude, and frequency at each node; the meteorological monitoring and forecasting system provides meteorological environmental data, including wind speed, precipitation intensity, icing thickness, and temperature; the equipment asset management system provides equipment topology and ledger data, including equipment capacity, service life, insulation class, and geographical location; and the online equipment status monitoring system provides equipment status monitoring data, including equipment operating temperature, vibration, and partial discharge.

[0011] By collecting data from the aforementioned multi-source data sources, we can obtain panoramic operational status information of cross-regional power distribution networks, providing a data foundation for subsequent feature construction and risk modeling.

[0012] Secondly, preprocessing is performed on the multi-source management data. Since the multi-source data comes from different acquisition systems, each with different temporal resolutions and data formats, and contains issues such as missing values ​​and outliers, unified processing is necessary. This preprocessing involves operations such as linear imputation of missing values, threshold truncation or removal of outliers, normalization of continuous features to the 0-1 range, and one-hot encoding of discrete features. These operations transform the multi-source heterogeneous data into a spatiotemporally aligned, dimensionally unified structured format. Finally, resampling is performed at a unified time step to form a standardized input sequence.

[0013] Furthermore, based on the preprocessing results, multidimensional feature vectors for multiple power element nodes and initial weight parameters for each associated edge are constructed. Power element nodes refer to key equipment units in inter-regional distribution networks, including at least five types: substation nodes, line nodes, load nodes, distributed power generation nodes, and protection automation device nodes. The multidimensional feature vector of each power element node is composed of three parts: electrical operation characteristics, including real-time power, voltage, current, and frequency; environmental status characteristics, including wind speed, precipitation, temperature, and icing thickness; and static attribute characteristics, including equipment capacity, service life, insulation class, and geographic location code. Initial weight parameters for associated edges are calculated separately for different edge types.

[0014] Through the above process of feature vector construction and initial weight parameter setting, multi-source heterogeneous power grid operation data, meteorological environment data, and equipment ledger data are uniformly abstracted into node features and edge weights in a graph structure, which can provide complete node and edge attribute information for the subsequent establishment of heterogeneous spatiotemporal graph models.

[0015] Specifically, multi-source management data of cross-regional distribution networks is collected through multi-source data sources. This multi-source management data is preprocessed, and based on the preprocessing results, multi-dimensional feature vectors of multiple power element nodes and initial weight parameters for each associated edge are constructed, including: Multidimensional operational data of the cross-regional distribution network are collected from the distribution network dispatch automation system, meteorological monitoring and forecasting system, equipment asset management system and equipment status online monitoring system. The multidimensional operational data includes at least: electrical measurement data, meteorological environment data, equipment topology and ledger data, equipment status monitoring data and historical fault event record data. The multidimensional operating data is preprocessed, and feature extraction is performed based on the preprocessing results to obtain the multidimensional feature vector of each power element node. The multidimensional feature vector includes at least electrical operating features, environmental status features, and static attribute features. Based on the meteorological environment data, the initial weight parameters of each associated edge are determined through the prior meteorological-vulnerability mapping curve.

[0016] First, multi-dimensional operational data of the cross-regional distribution network are collected from the distribution network dispatch automation system, meteorological monitoring and forecasting system, equipment asset management system, and equipment status online monitoring system. Specifically, the distribution network dispatch automation system provides real-time electrical measurement data, including active power, reactive power, voltage amplitude, phase angle, line current, and frequency at each node; the meteorological monitoring and forecasting system provides meteorological environmental data, including wind speed, precipitation intensity, temperature, humidity, icing thickness, and irradiance; the equipment asset management system provides equipment topology and ledger data, including substation location, line connection relationships, equipment capacity, years of operation, and insulation class; and the equipment status online monitoring system provides equipment status monitoring data, including main transformer oil temperature, winding temperature, partial discharge, and circuit breaker operation count.

[0017] In addition, historical fault event records need to be collected, including fault time, fault location, fault type, protection action sequence, and cascading fault range, for subsequent model training and path confidence assessment.

[0018] Secondly, the multidimensional operational data is preprocessed, and feature extraction is performed based on the preprocessing results to obtain the multidimensional feature vectors of each power element node. The preprocessing includes operations such as linear imputation of missing values, threshold truncation of outliers, normalization of continuous features to the 0-1 range, and one-hot encoding of discrete features. Feature extraction is the process of extracting key features that characterize the state of the power element nodes from the preprocessed data.

[0019] Specifically, the preprocessing steps include: The collected multidimensional operational data undergoes operations such as missing value imputation, outlier handling, and continuous feature normalization. For missing value imputation, missing values ​​in electrical measurement data are filled using linear interpolation between adjacent valid time points. For consecutive missing values ​​exceeding a preset threshold (e.g., more than 5 consecutive time steps), a validity mask vector is created for that feature dimension. The mask vector has the same dimension as the feature dimension, with valid positions set to 1 and missing positions set to 0. The mask vector is then concatenated with the original feature vector and input into the model. For outlier detection and handling, the mean and standard deviation of each feature dimension are calculated to obtain the Z-score. When the absolute value of the Z-score is greater than 3, it is considered an outlier, and the value is brought back to the threshold range of the mean plus or minus three times the standard deviation. For continuous feature normalization, for continuous variables such as active power, reactive power, and voltage amplitude in the electrical measurement data, the minimum and maximum values ​​on the training set are calculated, and the minimum-maximum normalization method is used to scale the values ​​to the 0-1 range.

[0020] Furthermore, the feature extraction steps include: Based on the preprocessed data, multidimensional feature vectors are extracted for each power element node. These multidimensional feature vectors include at least three parts: electrical operation features, environmental status features, and static attribute features. Each part consists of multiple specific feature dimensions.

[0021] Electrical operating characteristics reflect the real-time operating status of power element nodes at the current moment, and include at least: active power, reactive power, voltage amplitude, voltage phase angle, current amplitude, frequency, power factor, active load rate, reactive load rate, and three-phase imbalance. Taking a substation node as an example, the electrical operating characteristics consist of 12 dimensions, including the main transformer active load rate, main transformer reactive load rate, per-unit voltage of the high-voltage side bus, per-unit voltage of the low-voltage side bus, main transformer oil temperature, main transformer winding temperature, reactive power reserve, power factor, three-phase imbalance, short-circuit capacity, frequency deviation, and harmonic distortion rate.

[0022] Environmental situation characteristics are used to reflect the meteorological and environmental conditions of the geographical location of power element nodes, and include at least: wind speed, precipitation, temperature, humidity, icing thickness, lightning density, and geological disaster risk index. Taking overhead line nodes as an example, the environmental situation characteristics consist of 8 dimensions, including the wind speed, precipitation, temperature, humidity, icing thickness, lightning density, geological disaster risk index, and flood risk index of the area.

[0023] Static attribute features are used to reflect the inherent equipment parameters and spatial location information of power element nodes, and include at least: equipment capacity, years of operation, insulation class, voltage level code, equipment type code, geographical location longitude, and geographical location latitude. Taking a substation node as an example, the static attribute features have 7 dimensions, including rated capacity, voltage level code, years of operation, insulation class code, geographical location longitude, geographical location latitude, and equipment type code.

[0024] For discrete features, such as voltage level coding, equipment type coding, and weather type, traditional one-hot coding is not used; instead, a learnable embedding layer is employed. For example, voltage levels have four categories: 10 kV, 35 kV, 110 kV, and 220 kV. With an embedding dimension of 4, each voltage level corresponds to a learnable four-dimensional embedding vector, which is optimized synchronously during model training.

[0025] Finally, the electrical operation features, environmental status features, and static attribute features are concatenated to obtain a multi-dimensional feature vector for each power element node. For example, the complete feature vector of a substation node has the following dimensions: 12 dimensions for electrical operation features, 8 dimensions for environmental status features, and 7 dimensions for static attribute features. Adding the additional dimensions introduced by learnable embedding encoding, the total dimensions are approximately 30 to 40. This multi-dimensional feature vector serves as the initial feature representation for each node in the heterogeneous spatiotemporal graph model, and is input into the subsequent graph neural network for message passing and spatiotemporal modeling.

[0026] Furthermore, based on meteorological environmental data, the initial weight parameters of each associated edge are determined through a priori meteorological-vulnerability mapping curve. The meteorological-vulnerability mapping curve is a pre-established functional relationship based on historical meteorological disaster and equipment failure statistics, describing the mapping pattern between equipment downtime probability and meteorological elements under different meteorological conditions.

[0027] For example, the curve is constructed as follows: Historical meteorological disaster event data (including typhoons, rainstorms, icing, lightning strikes, etc.) from the past 10 years and equipment failure records within the corresponding time periods are collected. Meteorological elements (wind speed, precipitation intensity, icing thickness, etc.) are discretized into multiple intervals. The failure rate of equipment within each meteorological element interval is calculated. Logistic regression or piecewise linear fitting methods are used to fit the mapping function between meteorological elements and failure probabilities. Taking the wind speed-failure probability curve as an example, the lower limit of the design wind speed is set to 25 meters per second, and the upper limit of the design wind speed is set to 40 meters per second. When the wind speed is below 25 meters per second, the failure probability is 0; when the wind speed is above 40 meters per second, the failure probability is 1. The intermediate region is smoothly transitioned using a quadratic function. For example, the physical vulnerability coefficient is equal to the wind speed minus the square of the lower limit of the design wind speed divided by the upper limit of the design wind speed minus the square of the lower limit of the design wind speed.

[0028] Specifically, for geographically adjacent edges, their initial weights reflect the probability of synchronous impact of meteorological disasters on multiple devices within the same area, and can be calculated based on the geographical distance between nodes combined with the meteorological-vulnerability mapping curve. Optionally, the initial weight calculation formula for geographically adjacent edges is: the initial weight is equal to the negative exponential function value of the geographical distance between nodes, that is, the weight is equal to e raised to the power of negative (geographical distance divided by the distance decay coefficient).

[0029] For example, with a distance attenuation coefficient of 2 kilometers, when the geographical distance between two nodes is 0.5 kilometers, the initial weight is e^(-0.25) approximately equal to 0.78; when the distance is 2 kilometers, the initial weight is e^(-1) approximately equal to 0.37; and when the distance is 5 kilometers, the initial weight is e^(-2.5) approximately equal to 0.08. Simultaneously, the initial weight of the geographical edge is corrected using the meteorological-vulnerability mapping curve: when both nodes are located within the area affected by severe weather, the weight is multiplied by a meteorological synchronization factor. For example, when the wind speed exceeds 30 meters per second, the synchronization factor is set to 1.5, further increasing the initial weight of the geographical edge to reflect the actual physical law that the probability of nearby equipment being simultaneously affected by severe weather conditions increases.

[0030] For electrical connection edges, their initial weights are set based on line susceptance or power transmission limits. Specifically, the initial weight of an electrical connection edge is equal to the absolute value of the line susceptance divided by the absolute value of the maximum susceptance of all electrical edges, normalizing the weight to the range of 0 to 1. For example, if a line susceptance is 0.05 Siemens and the maximum susceptance of all electrical edges is 0.1 Siemens, then the initial weight of that electrical edge is 0.5. For information communication edges, their initial weights are set based on communication link reliability and average delay. The initial weight is equal to the historical availability rate of the communication link multiplied by a delay attenuation factor, where the delay attenuation factor is equal to e raised to the negative power of (average delay divided by a baseline delay of 50 milliseconds). For logically dependent edges, their initial weights are set based on protection coordination levels or switching success rates. The initial weight is equal to the switching success rate divided by the difference in protection coordination levels.

[0031] In this way, multi-source heterogeneous data can be transformed into node feature vectors and initial edge weights in a graph structure, providing standardized input for the subsequent construction of heterogeneous spatiotemporal graph models.

[0032] S20: Based on the distribution network topology in the multi-source management data, a heterogeneous spatiotemporal graph model is constructed by combining the multi-dimensional feature vector and the initial weight parameters. Message passing and time dimension modeling are performed through a graph neural network to obtain the risk amplification factor of each power element node and the transmission path of multiple target disaster chains. Secondly, based on the distribution network topology in multi-source management data, a heterogeneous spatiotemporal graph model is constructed by combining multi-dimensional feature vectors and initial weight parameters. The distribution network topology describes the connection relationships between power element nodes, including the electrical connection between substations and lines, the information association between protection devices and protected equipment, the relationship between geographically adjacent equipment, and the timing coordination relationship between protection devices.

[0033] The heterogeneous spatiotemporal graph model abstracts inter-regional distribution networks into a heterogeneous graph composed of multiple node types and edge types. Node types include at least five categories: substation nodes, line nodes, load nodes, distributed generation nodes, and protection automation device nodes. Edge types include at least four categories: electrical connection edges, information communication edges, geographical proximity edges, and logical dependency edges. Each node type has its own independent multi-dimensional feature vector dimension, and each edge type has its own independent initial weight parameters and message passing parameters. This heterogeneous spatiotemporal graph model is used to transform the diverse devices and their complex relationships in the physical power grid into graph-structured data, enabling subsequent graph neural networks to learn the propagation patterns of risks in the topology based on this structure.

[0034] Furthermore, message passing and temporal dimension modeling are performed using a graph neural network. This graph neural network adopts a heterogeneous spatiotemporal graph attention network architecture, which includes multiple layers of heterogeneous attention message passing layers and a temporal dimension modeling module. The heterogeneous spatiotemporal graph model defines the static structure of the graph, while the graph neural network performs dynamic message passing computation on this static structure, realizing the topological propagation and temporal evolution modeling of risk information.

[0035] Specifically, message passing is performed using dedicated transformation matrices and attention mechanisms corresponding to different edge types to propagate risk information from neighboring nodes along the topology to target nodes, capturing the spatial transmission characteristics of disaster chains. Temporal modeling is also performed using time coding and time-series networks to sequentially model node states across multiple time steps, capturing the time lag effect and evolution trend of risks. Finally, the risk amplification factor of each power element node and the transmission paths of multiple target disaster chains are obtained.

[0036] The risk amplification factor, a quantitative indicator calculated based on spatiotemporal embedding representation, characterizes the critical transmission position of a node in the disaster chain. When the risk amplification factor is significantly greater than 1, it indicates that the node's overall risk is much higher than its individual risk, placing it at a critical transmission position in the disaster chain and requiring close attention. The target disaster chain transmission path characterizes the complete path sequence of risk propagating from the source node along the topology to downstream nodes, including the node sequence, attention weights for each transmission segment, cumulative risk values, and estimated transmission times, providing dispatchers with a visual tracking basis for risk evolution.

[0037] Specifically, based on the distribution network topology in the multi-source management data, a heterogeneous spatiotemporal graph model is constructed by combining the multi-dimensional feature vectors and the initial weight parameters. A graph neural network is then used for message passing and time dimension modeling to obtain the risk amplification factor of each power element node and the transmission paths of multiple target disaster chains, including: The distribution network topology is obtained, and a heterogeneous spatiotemporal graph model containing multiple node types and multiple edge types is constructed by combining the multidimensional feature vectors and the initial weight parameters. By combining preset multi-level early warning targets, the heterogeneous spatiotemporal graph model is divided into discrete time steps, generating a corresponding multi-granularity graph snapshot sequence: The pre-trained hierarchical spatiotemporal graph attention network performs forward propagation based on multi-layer heterogeneous attention message passing based on the multi-granularity graph snapshot sequence to obtain the spatiotemporal embedding representation of each power element node; Based on the spatiotemporal embedding representation and combined with the multi-source management data, the risk amplification factor of each power element node is calculated and obtained; The risk amplification factor traverses the heterogeneous spatiotemporal graph model to perform reverse path search, obtains multiple target disaster chain transmission paths, and represents them in a structured manner, including at least: node sequence, cumulative transmission risk, estimated total transmission time, and historical hit rate based on historical data.

[0038] In the heterogeneous spatiotemporal graph model, the node types include at least substation nodes, line nodes, load nodes, distributed power generation nodes, and protection automation device nodes; the edge types include at least electrical connection edges, information communication edges, geographical proximity edges, and logical dependency edges, and each edge type corresponds to independent message transformation parameters and attention calculation parameters.

[0039] First, the distribution network topology is obtained, and a heterogeneous spatiotemporal graph model containing multiple node types and edge types is constructed by combining multi-dimensional feature vectors and initial weight parameters. The distribution network topology is extracted from the power grid GIS or CIM model and is used to describe the connection relationships between substations, lines, loads, distributed power sources, and protection devices.

[0040] In the heterogeneous spatiotemporal graph model, node types include at least five categories: substation nodes, line nodes, load nodes, distributed power generation nodes, and protection automation device nodes. Each node type has its own independent multi-dimensional feature vector dimension. Edge types include at least four categories: electrical connection edges, information communication edges, geographical proximity edges, and logical dependency edges. Each edge type corresponds to independent message transformation parameters and attention calculation parameters, and has its own initial weight parameters. Electrical connection edges represent power flow coupling relationships, information communication edges represent the transmission relationships of control signals and measurement data, geographical proximity edges represent geographical proximity relationships, and logical dependency edges represent the temporal dependencies of protection coordination and redundancy switching.

[0041] Specifically, the process of constructing the heterogeneous spatiotemporal graph model is as follows: Each power element node in the distribution network is treated as a vertex in the graph, and its node type is assigned to the corresponding node category based on its equipment type. For each vertex, the multi-dimensional feature vector extracted in step S10 is assigned to that node as its initial feature representation. The feature vector dimensions can differ for different node types; for example, the feature vector dimension of a substation node is 26 dimensions, a line node is 25 dimensions, a load node is 17 dimensions, a distributed power generation node is 19 dimensions, and a protection automation device node is 13 dimensions. Simultaneously, the type code for each node is recorded for subsequent use by the heterogeneous graph neural network to distinguish different node types.

[0042] For the relationships between nodes, edges are established based on the actual connection methods in the distribution network topology. Specifically, electrical connection edges are established based on the substations or loads connected to both ends of the line; information communication edges are established based on the flow of control signals and measurement data between the protection device and the protected equipment; geographical proximity edges are established based on whether the geographical distance between devices is less than a preset threshold; and logical dependency edges are established based on the timing coordination relationship between protection devices or the redundancy switching relationship between primary and backup equipment. For each established edge, the initial weight parameters calculated in step S10 are assigned to that edge as its initial weight. The calculation methods for the initial weight parameters of different types of edges are different. The initial weight of electrical connection edges is set based on the line susceptance or power transmission limit; the initial weight of information communication edges is set based on the reliability of the communication link and the average delay; the initial weight of geographical proximity edges is set based on a negative exponential function of geographical distance; and the initial weight of logical dependency edges is set based on the protection coordination level or switching success rate.

[0043] Through the above assignment process, the multidimensional feature vectors of the nodes and the initial weight parameters of the edges extracted in step S10 are respectively attached to the vertices and edges of the graph structure, forming a complete heterogeneous spatiotemporal graph model. The structured representation of this heterogeneous spatiotemporal graph model is G=(V,E,X,W), where V is the set of vertices, E is the set of edges, X is the feature vector matrix of all nodes, and W is the initial weight matrix of all edges.

[0044] Secondly, based on the pre-set multi-level early warning objectives, the obtained heterogeneous spatiotemporal graph model is divided into discrete time steps, generating a multi-granularity graph snapshot sequence. The pre-set multi-level early warning objectives include four time scales: minute-level, hour-level, day-level, and week-level, each corresponding to different early warning requirements. Minute-level early warnings require graph snapshots with a 1-minute time step for real-time electrical quantity anomaly detection; hour-level early warnings require graph snapshots with a 15-minute time step for weather-driven equipment outage prediction; day-level early warnings require graph snapshots with a 1-hour time step for cumulative damage cascading failure assessment; and week-level early warnings require graph snapshots with a 1-day time step for topological vulnerability evolution trend analysis.

[0045] Specifically, the preset multi-level early warning targets are comprehensively set based on the actual needs of power system dispatch operations for risk early warning at different time scales and the temporal characteristics of various disaster propagation. Minute-level early warnings are used to capture instantaneous anomalies in electrical quantities, corresponding to the response time of relay protection and emergency control; hourly-level early warnings are used to address the risk of equipment outages caused by sudden weather changes, corresponding to the operational window for dispatchers to take preventive and control measures; daily-level early warnings are used to assess the medium- and long-term impact of cumulative damage on system reliability, corresponding to the equipment maintenance plan formulation cycle; and weekly-level early warnings are used to analyze the evolution trend of topological vulnerability, corresponding to the evaluation cycle of power grid planning and transformation decisions.

[0046] The heterogeneous spatiotemporal graph model is divided into discrete time steps. A graph snapshot is generated at each time step. The graph snapshot contains the feature vectors of all nodes and the weight parameters of all edges at that time, forming a multi-granularity graph snapshot sequence.

[0047] Furthermore, the pre-trained hierarchical spatiotemporal graph attention network, based on multi-granularity graph snapshot sequences, performs forward propagation based on multi-layer heterogeneous attention message passing to obtain the spatiotemporal embedding representation of each power element node. The pre-trained hierarchical spatiotemporal graph attention network is a deep graph neural network constructed using a heterogeneous attention mechanism. This network is pre-trained using historical fault event data and normal operation data to extract spatiotemporal features from graph structure data. The core function of this hierarchical spatiotemporal graph attention network is to transform the initial feature vector of each node in the input multi-granularity graph snapshot sequence into a comprehensive embedding representation that integrates spatial topological propagation information and temporal evolution patterns through multi-layer message passing and temporal modeling. For graph snapshot sequences of different granularities, the hierarchical spatiotemporal graph attention network shares the same set of parameters but performs forward propagation independently, outputting the spatiotemporal embedding representation of nodes at the corresponding time scale for subsequent risk amplification factor calculation and disaster chain path extraction.

[0048] For example, the specific architecture and parameters of the hierarchical spatiotemporal graph attention network are as follows. The network consists of three heterogeneous attention message passing layers. Each layer contains independent message transformation matrices, attention parameters, and gating fusion parameters for four edge types. The input dimension of the message transformation matrix is ​​equal to the embedding dimension of the current layer node, and the output dimension is uniformly set to 128 dimensions. The input dimension of the first layer is equal to the original feature dimension of each node type: 26 dimensions for substation nodes, 25 dimensions for line nodes, 17 dimensions for load nodes, 19 dimensions for distributed power generation nodes, and 13 dimensions for protection automation device nodes. Through type-specific input transformation matrices, the original features of different dimensions are uniformly mapped to 128 dimensions. The input and output dimensions of the second and third layers are both 128 dimensions. The attention mechanism uses four heads, and the key dimension of each head is set to 32 dimensions. The time constant of the time decay kernel is set according to the edge type: 0.1 seconds for electrical edges, 5 seconds for information edges, 7200 seconds for geographical edges, and 30 seconds for logical edges. In gated fusion, the gate vector for each edge type is calculated by concatenating the current node embedding and the aggregation result of that type of edge, followed by linear transformation and the Sigmoid function. Residual connections and layer normalization are performed after message passing at each layer. The temporal modeling module employs a two-layer Transformer encoder, each layer containing four attention heads. The feedforward network has a dimension of 512, and the input sequence length is the number of time steps corresponding to each channel: 60 steps for minute-level channels, 96 steps for hour-level channels, 24 steps for day-level channels, and 7 steps for week-level channels. Through the above architecture and parameter configuration, the network can efficiently extract spatiotemporal features from multi-granularity graph snapshot sequences.

[0049] Specifically, the pre-trained hierarchical spatiotemporal graph attention network performs forward propagation based on the multi-granularity graph snapshot sequence and multi-layer heterogeneous attention message passing to obtain the spatiotemporal embedding representation of each power element node, including: For each edge type, a type-specific linear transformation function is used to transform the features of neighboring nodes into message vectors; The message vector is weighted by time delay using a time decay function constructed based on a time decay kernel; The attention weights of the neighboring nodes of the power element node are calculated through a multi-head attention mechanism. Based on the time-delay weighting results, intra-class aggregation based on edge type and gated cross-type fusion are performed sequentially to obtain the fused node embedding. The fusion nodes at each time step are embedded into the pre-trained time dimension modeling module to obtain the spatiotemporal embedding representation of each power element node, wherein the time dimension modeling module is established based on the time series architecture.

[0050] Step 1: For each edge type, a type-specific linear transformation function is used to transform the features of neighboring nodes into message vectors. Since different types of relationships convey different risk information, separate transformation matrices are needed to extract the corresponding feature patterns. Specifically, for electrical connection edges, the linear transformation matrix maps the features of neighboring nodes to electrical coupling risk messages, such as "the upstream substation is severely overloaded and may transfer power flow to this node." For information communication edges, the linear transformation matrix maps the features of neighboring nodes to communication-related risk messages, such as "increased communication delay of the protection device may lead to failure to operate." For geographically proximate edges, the linear transformation matrix maps the features of neighboring nodes to geographically proximate risk messages, such as "adjacent substations within the same typhoon corridor have been shut down." For logically dependent edges, the linear transformation matrix maps the features of neighboring nodes to logically dependent risk messages, such as "the upstream protection has been activated, and it is this protection's turn to assume the responsibility for preventing overstepping." Each edge type uses an independent parameter matrix to ensure that the model can learn the differentiated semantics of different types of relationships.

[0051] Step two involves applying time-delay weighting to the message vector using a time decay function constructed based on a time decay kernel. Disaster propagation exhibits a time-delay effect; failure events occurring at different times have varying degrees of impact on the current risk assessment. Events occurring more recently have a greater impact, while the impact of events occurring further back in time decays exponentially over time. The time decay function is defined as an exponential decay form, with the decay rate determined by the characteristic time constant corresponding to the edge type.

[0052] Among them, different types of fault propagation have completely different time scales. For example, fault propagation on the electrical connection side can be completed in milliseconds, so the time constant is set to 0.1 seconds; the control signal and measurement data transmission on the information communication side have a delay of seconds, so the time constant is set to 5 seconds; the duration of meteorological disaster impact on the geographically adjacent side is on the hour level, so the time constant is set to 7200 seconds; the protection coordination and switching actions on the logically dependent side are on the second to minute level, so the time constant is set to 30 seconds.

[0053] By using time decay weighting, the impact of failure events occurring at different times on the current risk assessment is negatively correlated with the time interval between their occurrence, thus achieving differentiated weighting of risk events occurring at different times.

[0054] Step 3: Calculate the attention weights of the neighboring nodes of the power element node through a multi-head attention mechanism, and perform intra-class aggregation based on edge type according to the time-delay weighted results.

[0055] First, for each edge type, the attention score between the current node and each of its neighbors is calculated. The attention score is calculated by concatenating the current node's feature vector with the time-delay-weighted message vectors of its neighbors, performing a learnable attention vector transformation, and then applying the LeakyReLU activation function to obtain the unnormalized attention score. Subsequently, the attention scores of all neighbors are normalized using Softmax to obtain the attention weight. This attention weight reflects the strength of risk propagation from the neighbor to the current node; a larger weight indicates that the risk from the neighbor is more easily propagated to the current node.

[0056] Next, the attention weight is multiplied by the time-delay-weighted message vector, and then the weighted sum of all neighboring nodes under that edge type is performed to obtain the aggregate vector for that edge type. Through this step, four aggregate vectors are output for each of the four edge types: the electrical edge aggregate vector represents risk information from the electrical coupling direction, the information edge aggregate vector represents risk information from the communication association direction, the geographic edge aggregate vector represents risk information from the geographic proximity direction, and the logical edge aggregate vector represents risk information from the logical dependency direction.

[0057] Preferably, after intra-class aggregation, residual connections and layer normalization are performed. The previous-layer embedding of the current node is added to the intra-class aggregation result to achieve residual connections. When neighbor information is scarce or attention weights are dispersed, the aggregation result may contain very little information. Residual connections ensure that even if neighbors do not convey valuable information, the node's own features can be fully preserved in the next layer. Subsequently, layer normalization is performed to unify the vectors after residual connections to the same numerical range, enabling subsequent gating fusion to fairly compare the importance of each type.

[0058] Step four involves cross-type fusion using a gating mechanism, integrating the aggregation results of the four edge types into a unified node embedding. For each edge type, a gating vector is calculated. The gating vector is calculated by concatenating the previous-level embedding of the current node with the aggregation vector of that edge type, followed by a linear transformation and a sigmoid activation function, outputting a gating value between 0 and 1. The gating value represents the importance of the edge type's information in the current operating scenario. For example, during typhoon weather, the gating value for geographical edges should be increased because the synchronous impact of meteorological disasters is a major risk source; in communication failure scenarios, the gating value for information edges should be decreased due to unreliable communication connections. The fused node embedding is obtained by multiplying the aggregation vectors of each edge type by their corresponding gating values ​​and then summing the results.

[0059] Step 5: Embed the fused nodes from each time step into the pre-trained temporal modeling module to obtain the spatiotemporal embedding representation of each power element node. The temporal modeling module is built based on a time-series architecture and can employ a Transformer encoder or a gated recurrent unit. The input to this module is a time series, i.e., the fused node embeddings of the same node over multiple past time steps. The Transformer encoder captures the dependencies between different time steps through a self-attention mechanism, enabling the model to identify trends in risk accumulation or decay over time. The gated recurrent unit controls the flow of information by updating and resetting gates, effectively addressing the gradient vanishing problem in long sequences. The output is a spatiotemporal embedding representation that integrates spatial propagation information and temporal evolution information. This representation includes both the risk propagation information received by the node from its neighbors and the changes in the node's own state over time, used for subsequent risk amplification factor calculation and disaster chain path extraction.

[0060] For example, the temporal modeling module employs a two-layer Transformer encoder, each layer containing four attention heads, with the feedforward network hidden dimension set to 512. The input sequence length is T, for example, 60 time steps for a minute-level channel, and the embedding dimension of the fusion node at each time step is dh, for example, 128 dimensions. The Transformer encoder captures the dependencies between different time steps through a self-attention mechanism, outputs the encoding result of each time step, and takes the output of the last time step as the spatiotemporal embedding representation of the current moment. If a gated recurrent unit (GRU) is used, two layers of GRU are set, with a hidden layer dimension of 128. The fusion node embeddings of each time step are input sequentially, and the hidden state of the last time step is taken as the spatiotemporal embedding representation.

[0061] Furthermore, based on the spatiotemporal embedding representation obtained through forward propagation, physical constraints are also embedded during model training to ensure the physical rationality of the data-driven model. These constraints include three categories: power balance constraints, thermal stability constraints, and propagation causality constraints.

[0062] Power balance constraints require that the node risk scores predicted by the graph neural network should not violate basic power flow balance relationships. Specifically, based on the relationship between node injected power and node voltage phase angle, the deviation between the predicted value and the calculated value from the power flow equation is calculated, and the sum of the squares of this deviation is taken as the power balance constraint loss. This power balance constraint is used to ensure that the risk score output by the model is consistent with the physical operating laws of the power grid, avoiding unreasonable predictions that violate Kirchhoff's laws.

[0063] The thermal stability constraint penalty graph neural network predicts situations where the thermal stability limit of the line is exceeded. Specifically, for each electrical connection edge, based on the difference between the predicted line current and the maximum allowable current carrying capacity of the line, when the predicted current exceeds the maximum allowable current carrying capacity, the square of the excess is calculated as a thermal stability constraint penalty; otherwise, the penalty is zero. This thermal stability constraint is used to guide the model to avoid outputting unreasonable results that violate the thermal stability limit when predicting risks.

[0064] The propagation causality constraint ensures that the propagation direction reflected by the attention weights is consistent with the fault timeline. Specifically, based on historical fault event data, the probability of fault propagation direction is statistically analyzed. A positive reward is given when the propagation direction indicated by the attention weights is consistent with the historical statistical direction, and a penalty is given when they are inconsistent. This propagation causality constraint is used to ensure that the risk transmission direction learned by the model conforms to causal logic, that is, the causal event occurs before the result event.

[0065] The model's total loss function is a weighted sum of the warning task loss and the losses from the three types of physical constraints mentioned above. The warning task loss includes the cross-entropy loss from node fault classification and the mean squared error loss from edge propagation weighted regression. The total loss function is calculated as follows: the total loss equals the warning task loss plus the power balance constraint loss multiplied by the first weight coefficient, plus the thermal stability constraint loss multiplied by the second weight coefficient, plus the propagation causality constraint loss multiplied by the third weight coefficient, plus the variational autoencoder loss multiplied by the fourth weight coefficient. Each weight coefficient can be adjusted based on performance on the validation set. For example, the first weight coefficient is set to 0.1, the second weight coefficient to 0.1, the third weight coefficient to 0.05, and the fourth weight coefficient to 1.0.

[0066] By embedding the aforementioned physical constraints, the model can optimize prediction accuracy while ensuring the physical rationality of the prediction results. This avoids prediction outputs that violate basic physical laws that may be generated by purely data-driven models, thereby improving the credibility of the early warning results and the acceptability of the dispatchers.

[0067] Furthermore, based on the spatiotemporal embedding representation and combined with multi-source management data, the risk amplification factor of each power element node is calculated. This risk amplification factor is used to quantify the node's critical transmission position in the disaster chain. When the risk amplification factor is much greater than 1, it indicates that the node is at a critical transmission position in the disaster chain, and its overall risk is much higher than its independent risk.

[0068] Specifically, based on the spatiotemporal embedding representation and combined with the multi-source management data, the risk amplification factor for each power element node is calculated and obtained, including: Based on the spatiotemporal embedding representation, the basic risk value of each power element node is calculated through a risk scoring function model; Based on the meteorological environment data and equipment status monitoring data in the multi-source management data, the physical vulnerability coefficient of each power element node is calculated; Using the attention weight between each power element node and its neighbor nodes as the weighting weight, the basic risk values ​​of multiple neighbor nodes and the corresponding physical vulnerability coefficients are weighted and summed to obtain the neighbor risk weighting value for each power element node. Based on the basic risk value and the corresponding neighbor risk weighting value, a comprehensive risk value is calculated, and the ratio of the comprehensive risk value to the basic risk value is output as the risk amplification factor for each power element node.

[0069] First, based on spatiotemporal embedding representation, the basic risk value of each power element node is calculated through a risk scoring function model. Spatiotemporal embedding representation is a node feature vector output by a graph neural network after multi-layer heterogeneous attention message passing and time dimension modeling. This vector integrates the node's own original features, risk information propagated by neighboring nodes, and the temporal evolution law.

[0070] The risk scoring function model is a learnable multilayer perceptron module embedded at the output of a graph neural network (Graph Neural Network). It is jointly trained with the entire Graph Neural Network, and its parameters are automatically learned from historical fault data during training via backpropagation. Specifically, the risk scoring function model employs a multilayer perceptron structure, taking spatiotemporal embeddings as input. It progressively reduces dimensionality through two fully connected layers: the first layer maps from the hidden layer dimension to 64 dimensions, and the second layer maps from 64 dimensions to 32 dimensions. Finally, a linear layer and a sigmoid activation function output a scalar value between 0 and 1. During model training, real fault labels from historical fault event data are used as supervision signals. The parameters of the entire network, including the risk scoring function model, are optimized by minimizing the cross-entropy loss between the predicted risk value and the real fault label. After training, the risk scoring function model possesses the ability to map the spatiotemporal embeddings of nodes to the basic risk value.

[0071] The final output scalar value is the basic risk value of that node, representing the independent failure probability of a power element node when only its own factors are considered, without considering the impact of topology propagation. For example, a basic risk value of 0.3 for a substation node means that the substation has a 30% probability of failure based solely on its own condition assessment; the higher the value, the worse the node's own condition and the greater the risk of independent failure.

[0072] Second, based on meteorological and environmental data and equipment status monitoring data from multi-source management data, the physical vulnerability coefficient of each power element node is calculated. The physical vulnerability coefficient is used to indirectly characterize the risk transmission capability of each associated edge, that is, how easily a fault can be transmitted to the current node through different types of associated edges when a neighboring node fails. The calculation method for the physical vulnerability coefficient differs for different edge types.

[0073] Specifically, for electrical connection edges, the physical vulnerability coefficient is the minimum of the line load rate and 1, that is, the ratio of the current line transmission power to the line's maximum transmission capacity. When the line load rate is close to 1, it indicates that the line is near full load, and power fluctuations from neighboring nodes can easily be transmitted to the current node through this line, resulting in a physical vulnerability coefficient close to 1. When the line load rate is low, the physical vulnerability coefficient is small, indicating that the line has a certain power margin and weaker risk transmission capability. For example, if a line currently transmits 80 MW of power and has a maximum transmission capacity of 100 MW, then the load rate is 0.8, and the physical vulnerability coefficient is 0.8.

[0074] For information communication links, the physical vulnerability coefficient is calculated by subtracting the historical availability rate of the communication link from 1, which represents the communication unreliability. A low availability rate indicates that control signals or measurement data are easily lost or delayed, resulting in strong risk transmission and a high physical vulnerability coefficient. For example, if the availability rate of a communication link is 99.9%, the physical vulnerability coefficient is 0.001.

[0075] For geographically proximate edges, the physical vulnerability coefficient is calculated based on real-time wind speed and the design wind speed threshold. Specifically, the real-time wind speed of the node's location is obtained from meteorological monitoring data. When the wind speed is below the lower limit of the design wind speed, the physical vulnerability coefficient is 0; when the wind speed is above the upper limit of the design wind speed, the physical vulnerability coefficient is 1; intermediate areas are smoothly transitioned using a quadratic function. For example, if the lower limit of the design wind speed is 25 m / s, the upper limit of the design wind speed is 40 m / s, and the real-time wind speed is 35 m / s, then the physical vulnerability coefficient is 0.44.

[0076] For logically dependent edges, the physical vulnerability coefficient is 1 minus the handover success rate. A low handover success rate indicates lower reliability of backup equipment or backup protection, and stronger risk transmission capability. For example, if the handover success rate of a certain automatic transfer switch for backup power is 95%, then the physical vulnerability coefficient is 0.05.

[0077] Third, using the attention weights between each power element node and its neighboring nodes as weighting weights, the basic risk values ​​of multiple neighboring nodes are summed with their corresponding physical vulnerability coefficients to obtain the neighbor risk weighting value for each power element node. The attention weights are calculated during the forward propagation of the graph neural network and reflect the strength of risk transmission from neighboring nodes to the current node. For each neighboring node, its basic risk value is multiplied by its corresponding physical vulnerability coefficient, and then multiplied by the attention weight between the current node and that neighbor to obtain the neighbor's risk contribution to the current node. The risk contributions of all neighboring nodes are summed to obtain the neighbor risk weighting value. This neighbor risk weighting value represents the total amount of risk transmitted from neighboring nodes to the current node through topology propagation.

[0078] For example, if a node has three neighbors with base risk values ​​of 0.3, 0.2, and 0.1, corresponding physical vulnerability coefficients of 0.8, 0.5, and 0.2, and attention weights of 0.5, 0.3, and 0.2, then the weighted average risk of the neighbors is 0.3×0.8×0.5+0.2×0.5×0.3+0.1×0.2×0.2=0.154.

[0079] Fourth, based on the base risk value and the corresponding neighbor risk weighting value, a comprehensive risk value is calculated, and the ratio of the comprehensive risk value to the base risk value is output as the risk amplification factor for each power element node. The comprehensive risk value equals the base risk value plus the neighbor risk weighting value, representing the overall risk of the node after considering topology propagation. The risk amplification factor equals the comprehensive risk value divided by the base risk value, used to quantify the degree to which topology propagation amplifies the node's risk.

[0080] When the risk amplification factor is approximately 1, it indicates that the neighbor risk weighting value is much smaller than the base risk value, and the node's risk is mainly determined by its own factors, with minimal impact from topology propagation. When the risk amplification factor is greater than 1, it indicates that the neighbor risk weighting value contributes significantly to the overall risk, and the node is on the propagation path of the disaster chain. When the risk amplification factor is much greater than 1, such as greater than 2, it indicates that the node is at a critical position in the disaster chain, and its overall risk is much higher than its independent risk, requiring close monitoring.

[0081] For example, if a node has a base risk value of 0.2 and a neighbor risk weighting value of 0.3, then the overall risk value is 0.5 and the risk amplification factor is 2.5. This means that the node's independent failure probability is only 20%, but the neighbor risk received through topology propagation increases its overall risk to 50%, and the risk is amplified by 2.5 times. This node is in a critical transmission position in the disaster chain.

[0082] The above calculations yield a risk amplification factor for each power element node, which can be used for subsequent screening of key transmission nodes and extraction of disaster chain transmission paths. Nodes with higher risk amplification factors are more important in the disaster chain and should be considered key monitoring targets.

[0083] Furthermore, using the power element nodes with the highest preset quantiles in the risk amplification factor as source nodes, the heterogeneous spatiotemporal graph model is traversed to perform reverse path search, thereby obtaining multiple target disaster chain transmission paths.

[0084] Specifically, the risk amplification factor traverses the heterogeneous spatiotemporal graph model to perform a reverse path search, obtaining multiple target disaster chain transmission paths, including: The power element nodes ranked at the top preset quantile of the risk amplification factor are taken as the source node set; Starting from each source node, perform a beam search along the associated edge with the largest attention weight; During the search process, the beam search results are pruned and delooped based on the following: the cumulative transmission risk is lower than the preset risk threshold, the cumulative transmission time exceeds the preset time limit, and duplicate power element nodes appear in the path. The paths with the highest cumulative transmission risk values ​​are retained as the target disaster chain transmission paths, and the confidence level is calculated for each target disaster chain transmission path based on the hit rate of similar paths in historical disaster event data.

[0085] The first step involves selecting the power element nodes with the highest risk amplification factors (ranked by a preset quantile) as the source node set. The risk amplification factor reflects the critical transmission position of each node in the disaster chain; a higher risk amplification factor indicates that the node is more closely positioned on the critical path of risk amplification. All nodes are sorted by their risk amplification factors from highest to lowest, and the nodes with the highest risk amplification factors (ranked by a preset quantile) are selected as source nodes. This preset quantile can be set according to actual needs, for example, selecting the top 10% of nodes as source nodes. These source nodes are potential starting points or key intermediate nodes in the disaster chain. Performing a reverse path search starting from these nodes can efficiently locate the main disaster chain transmission paths.

[0086] The second step involves performing a beam search starting from each source node and following the associated edge with the highest attention weight. Beam search is a heuristic search algorithm that retains only the currently optimal few paths at each step, thus reducing computational complexity while maintaining search quality.

[0087] Specifically, starting from the source node, all outgoing edges connected to that node are examined and sorted in descending order of attention weight. Only the first b edges are retained for expansion, where b is called the beamwidth. The beamwidth represents the number of candidate paths retained at each layer during the search process. This beamwidth is set based on a balance between computational resources and search accuracy; for example, it can be set to 3. A larger beamwidth searches for more paths, increasing the likelihood of finding the globally optimal path, but also increasing computational cost. A smaller beamwidth is more computationally efficient but may miss optimal paths. The process extends along each retained edge to the next node, recording the attention weight, edge type, and propagation time for that segment. This process is repeated layer by layer, expanding outwards to form multiple candidate paths.

[0088] The third step involves pruning and loop removal of beam search results based on factors such as cumulative transmission risk being lower than a preset risk threshold, cumulative transmission time exceeding a preset time limit, and duplicate power element nodes appearing in the path.

[0089] Cumulative transmission risk refers to the product of the attention weight and physical vulnerability coefficient of all segments along the path from the source node to the current node. When the cumulative transmission risk is lower than a preset risk threshold, it indicates that the risk of subsequent transmission along the path is already very small, and further expansion is not meaningful; therefore, the path is pruned. The preset risk threshold is a lower limit of cumulative transmission risk used to determine whether a path still has value for further expansion. It is set comprehensively based on the actual physical laws of disaster chain transmission and the risk sensitivity requirements of the early warning system. For example, it is set to 0.01, meaning that expansion is terminated when the cumulative risk is below 1%. The lower the preset risk threshold, the longer the searched path, but it may include many low-risk paths with little practical significance; the higher the preset risk threshold, the higher the search efficiency, but it may miss low-risk paths that are still transmitting.

[0090] Cumulative propagation time refers to the sum of the propagation times of each segment along the path from the source node to the current node. The propagation time differs for different edge types; electrical edges have propagation times in milliseconds, while geographical edges have propagation times in hours. This propagation time is calculated using a time decay function or set based on historical statistical data; for example, the propagation time for electrical edges can be set to 0.1 seconds, and for geographical edges, it can be set to 2 hours.

[0091] When the cumulative transmission time exceeds the preset time limit, it indicates that the transmission time of the path has exceeded the warning focus window, and further expansion is not meaningful; therefore, the path is pruned. The preset time limit is the maximum cumulative transmission time used to determine whether a path is still within the warning focus window. It is set comprehensively according to the time scale requirements of the warning task, and can be set according to the warning objective. For example, the time limit for minute-level warnings is set to 1 hour, and the time limit for hourly warnings is set to 24 hours. The smaller the preset time limit, the more focused the search is on recent risk transmission paths, and the faster the warning response; the larger the preset time limit, the longer the search coverage, but it may include distant, low-urgency paths.

[0092] When duplicate power element nodes appear in the path, it indicates that the path has formed a loop, and the risk is circulated and transmitted in the loop, which does not conform to the unidirectional transmission characteristic of the disaster chain. Therefore, the path is de-looped and the expansion is terminated.

[0093] The fourth step is to retain the paths with the highest cumulative transmission risk values ​​as the target disaster chain transmission paths. Among all candidate paths, they are sorted from highest to lowest cumulative transmission risk, and the top-ranked paths are retained, for example, the top 5 paths. The structured representation of each path includes at least: a node sequence, i.e., a sequential list of power element nodes traversed from the source node to the end node; cumulative transmission risk, i.e., the product of the transmission risk values ​​of each segment on the path; estimated total transmission time, i.e., the sum of the transmission times of each segment on the path; and attention weights and edge types for each segment, used to explain the specific path and mechanism of risk transmission.

[0094] Simultaneously, based on the hit rate of similar paths in historical disaster event data, a confidence score is calculated for each target disaster chain propagation path. Specifically, the confidence score is calculated as follows: in the historical disaster event database, disaster chain paths with similar topological structures to the current path are retrieved, and the proportion of actual disaster chain propagation in similar paths is statistically analyzed. This proportion is used as the confidence score of the current path. The criteria for determining similar paths are: calculating the node sequence edit distance between each historical path in the historical disaster event database and the currently extracted path. The edit distance is defined as the minimum number of single-element edit operations required to convert one sequence into another. When the edit distance is less than a preset similarity threshold, the two paths are considered similar. This preset similarity threshold can be set according to the node sequence length, for example, 20% of the sequence length. For example, if the current path node sequence is [A,B,C,D], the historical path node sequence is [A,B,E,D], the edit distance is 1 (replacing C with E), the sequence length is 4, and the similarity threshold is 0.8, the edit distance 1 is less than 0.8 × 4 = 3.2, and the paths are considered similar. Using the above method, disaster chain paths with similar topological structures to the current path can be retrieved from the historical database, and the proportion of actual disaster chain propagation in similar paths can be counted as the confidence level of the current path.

[0095] For example, if a similar topological structure has occurred 10 times in history, and the propagation of a disaster chain along that path was actually observed in 8 of them, then the confidence level of that path is 80%. The higher the confidence level, the higher the historical hit rate of that path, and the stronger the credibility of the current prediction.

[0096] Through the above steps, multiple high-confidence disaster chain transmission paths can be extracted from the heterogeneous spatiotemporal graph model. Each path contains complete node sequences, cumulative risks, transmission time, and confidence information, providing a traceable risk transmission basis for subsequent multi-scale early warning and scheduling decisions.

[0097] S30: Combining the risk amplification factor with the multi-source management data, perform multi-scale parallel early warning for multiple target disaster chain transmission paths, and perform spatiotemporal fusion of the multi-scale parallel early warning results to obtain a comprehensive risk matrix; Furthermore, by combining risk amplification factors with multi-source management data, multi-scale parallel early warning is conducted for the transmission paths of multiple target disaster chains. The evolution of disaster chain risks involves multiple time scales, ranging from minute-level electrical transient processes to weekly equipment degradation trends. Early warning at a single time scale cannot fully reflect the risk situation of the system. Therefore, it is necessary to establish a multi-scale parallel early warning framework and conduct specialized modeling for the risk characteristics at different time scales.

[0098] Specifically, the multi-scale parallel early warning framework includes at least four parallel early warning channels: minute-level, hourly-level, daily-level, and weekly-level. Each channel outputs a channel risk score for each power element node based on input data at different time granularities and different risk assessment logics. The minute-level channel, based on real-time electrical measurement data, detects immediate electrical anomalies such as overload and voltage exceedances by predicting thermal stability margin, voltage margin, and frequency margin. The hourly-level channel, based on weather forecast data, predicts the probability of equipment outages in the next few hours using a weather-vulnerability mapping curve and simulates cascading outages using a risk amplification factor. The daily-level channel, based on equipment condition monitoring data, assesses the cumulative damage and remaining lifespan of equipment using a health index degradation model and simulates cascading failures using the Monte Carlo method. The weekly-level channel, based on topological characteristics, calculates indicators such as the rate of change of betweenness centrality, risk amplification factor ranking fluctuation, and graph connectivity resilience to assess the evolution trend of system structural vulnerability.

[0099] Secondly, the multi-scale parallel early warning results are spatiotemporally fused to obtain a comprehensive risk matrix. The rows of this comprehensive risk matrix correspond to power element nodes, and the columns correspond to different time scales. Matrix elements represent the comprehensive risk score of the node at the corresponding time scale. The comprehensive risk matrix provides a unified format of input features for the subsequent variational autoencoder anomaly detection module. It also serves as a quantitative basis for dispatchers to assess the risk status of each power element node at different time scales, enabling dispatchers to quickly identify nodes with high risk across different time dimensions such as minutes, hours, days, and weeks, as well as nodes with risk only at specific time scales. This allows for the development of targeted, tiered early warning strategies and differentiated inspection plans.

[0100] Specifically, combining the risk amplification factor with the multi-source management data, multi-scale parallel early warning is performed on multiple target disaster chain transmission paths, and the multi-scale parallel early warning results are spatiotemporally fused to obtain a comprehensive risk matrix, further including: Establish a multi-scale parallel early warning framework, including at least: minute-level channels, hour-level channels, day-level channels and week-level channels, which output channel risk scores for each power element node in parallel based on input data of different time granularities and different risk assessment logics. The channel risk scores of each power element node output by the multi-scale parallel early warning framework are normalized, and then multiplied by the preset scene channel weight and the normalized risk amplification factor, and summed to obtain the spatial fusion risk value of each power element node. Based on the prior time risk regression model corresponding to each early warning channel, the output results of each early warning channel are extrapolated along their respective time scales to obtain the predicted risk increment of each power element node at multiple scales. The spatial fusion risk value and the predicted risk increment are superimposed to generate a comprehensive risk matrix for each power element node.

[0101] First, establish a multi-scale parallel early warning framework, including at least four parallel early warning channels: minute-level channel, hour-level channel, day-level channel, and week-level channel.

[0102] The minute-level channel, based on real-time electrical measurement data and using a 1-minute time step, detects immediate electrical anomalies such as overload, voltage exceedance, and frequency deviation by predicting thermal stability margin, voltage margin, and frequency margin, and outputs a minute-level risk score for each power element node. The thermal stability margin refers to the distance between the current current and the maximum allowable current carrying capacity of the line. The calculation formula is: thermal stability margin = 1 minus the current current divided by the maximum allowable current carrying capacity. The closer the margin is to 0, the closer the line is to overload; a negative margin indicates that the line is already overloaded. For example, if a line has a maximum allowable current carrying capacity of 100 amps and a current of 95 amps, the thermal stability margin is 0.05, indicating that the line is close to full load. When the thermal stability margin falls below a preset threshold, such as below 0.1, a minute-level warning is triggered. The reciprocal of the thermal stability margin, after normalization, is used as the contribution to thermal stability risk.

[0103] Voltage margin refers to the minimum distance between the current voltage of a node and its upper and lower voltage limits. The calculation formula is: voltage margin equals the smaller of the lower voltage limit minus the current voltage and the current voltage minus the upper voltage limit. For example, if the lower voltage limit is 0.95 times the per-unit value, the upper voltage limit is 1.05 times the per-unit value, and the current voltage is 0.96 times the per-unit value, then the difference between the lower voltage limit and the current voltage is negative, and the difference between the current voltage and the upper voltage limit is also negative. The larger of the two, the smaller absolute value, results in a voltage margin of 0.01. When the voltage margin falls below a preset threshold, for example, below 0.02, a minute-level warning is triggered. The reciprocal of the voltage margin, after normalization, is used as the voltage risk contribution. Frequency margin refers to the absolute value of the difference between the rated frequency and the current frequency. For example, if the rated frequency is 50 Hz and the current frequency is 49.8 Hz, then the frequency margin is 0.2 Hz. When the frequency margin exceeds a preset threshold, for example, exceeding 0.2 Hz, a minute-level warning is triggered. The frequency margin, after normalization, is used as the contribution to frequency risk.

[0104] The contributions of thermal stability risk, voltage risk, and frequency risk are weighted and summed according to preset weights to obtain a minute-level risk score. The weighting coefficients are comprehensively set based on the relative importance of each risk factor to the safe operation of the power grid, as well as the actual operating characteristics of different voltage levels and regional power grids. For example, for urban power grids characterized by dense loads and limited line carrier frequency, the thermal stability weight is set to 0.5, the voltage weight to 0.3, and the frequency weight to 0.2. For regional power grids characterized by high penetration of new energy sources and frequent frequency fluctuations, the frequency weight can be appropriately increased; for example, the thermal stability weight could be set to 0.4, the voltage weight to 0.2, and the frequency weight to 0.4. Through this weighted summation, the minute-level risk score can comprehensively reflect the overall risk level of each power element node in the dimension of electrical quantity anomalies.

[0105] The hourly channel, based on meteorological forecast data and using a 15-minute time step, predicts the probability of equipment outages within the next 4 to 24 hours using a meteorological-vulnerability mapping curve. It also simulates cascading outages using a risk amplification factor, outputting an hourly risk score for each power element node. The meteorological-vulnerability mapping curve is a pre-established functional relationship based on historical meteorological disaster and equipment failure statistics, describing the mapping pattern between equipment outage probabilities and meteorological elements under different meteorological conditions. For example, with wind speed, the outage probability is 0 when the wind speed is below 25 meters per second, and 1 when it is above 40 meters per second, with a smooth transition in the intermediate region following a quadratic function.

[0106] Equipment vulnerability assessment is a quantitative evaluation of the equipment's inherent disaster resistance capabilities. The equipment vulnerability coefficient is obtained by multiplying three components: service life vulnerability, insulation condition vulnerability, and geographical location vulnerability. Service life vulnerability equals 1 plus the service life coefficient multiplied by the positive part of the difference between the service life and the design life. For example, if the service life coefficient is 0.02, the design life is 20 years, and the actual service life is 25 years, then the service life vulnerability is 1.1. Insulation condition vulnerability is the reciprocal of the insulation health index. For example, if the insulation health index is 0.9, then the insulation condition vulnerability is 1.11. Geographical location vulnerability is set based on micro-topography and micro-meteorological characteristics; for example, equipment located in windy or icing-prone areas has a vulnerability coefficient of 1.2, while equipment in ordinary areas has a coefficient of 1.0. The overall equipment vulnerability coefficient is the product of these three factors.

[0107] The equipment downtime probability is equal to the weather downtime probability output by the weather-vulnerability mapping curve multiplied by the equipment's overall vulnerability coefficient, and then cropped to the range of 0 to 1. For example, if the weather downtime probability is 0.44 and the equipment's overall vulnerability coefficient is 1.1, then the equipment downtime probability is 0.484. Specifically, for equipment whose predicted downtime probability exceeds a preset threshold, such as 0.3, its status is marked as faulty, the graph snapshot features are updated, and the graph neural network forward propagation is rerun to obtain the updated risk values ​​of other nodes. This process is repeated until no new faulty equipment is added or the maximum number of iterations is reached, such as 10 times.

[0108] The final risk score is a weighted sum of the probability of equipment failure and the risk value after cascading propagation. The weighting coefficients are set based on the different contributions of equipment failure and cascading propagation to the overall risk. For example, the weight of the equipment failure probability is set to 0.4, and the weight of the risk value after cascading propagation is set to 0.6, ensuring that the cascading propagation effect dominates the hourly risk score. When the probability of equipment failure is high but the risk of cascading propagation is low, the overall risk score is mainly contributed by the equipment's own factors. When the probability of equipment failure is moderate but the equipment is in a critical transmission position, the weight of cascading propagation significantly increases the overall risk score, thus more accurately reflecting the characteristics of the disaster chain transmission.

[0109] The daily-level channel, based on equipment condition monitoring data and using a 1-hour time step, assesses the cumulative damage and remaining lifespan of equipment through a health index degradation model. It employs the Monte Carlo method to simulate cascading failures and outputs a daily-level risk score for each power element node. The health index degradation model describes the decline in equipment health status over time. The equipment health index is a dimensionless indicator between 0 and 1, where 1 represents brand-new equipment and 0 represents complete failure. The initial health index is set to 1. The health index decreases non-linearly with service time and the accumulation of stress factors. Its calculation formula is: the current health index equals the initial health index minus the degradation rate coefficient multiplied by the service life raised to the power of β, and then multiplied by the exponentially weighted sum of the cumulative exposure to each stress factor. The degradation rate coefficient is related to the equipment type and manufacturing quality; for example, the degradation rate coefficient for transformers can be taken as 0.02. β is the degradation nonlinearity exponent; when β is greater than 1, it indicates accelerated degradation; for example, it can be taken as 1.2. Stress factors include average load rate, temperature cycle count, cumulative humidity exposure, and number of short-circuit impacts. The influence coefficients of each stress factor are obtained by fitting historical failure data; for example, the influence coefficient of average load rate can be taken as 0.3, and the influence coefficient of temperature cycle count can be taken as 0.2. When the health index is lower than the failure threshold of 0.3, the equipment is considered to have reached the end of its service life and has a high risk of failure.

[0110] Based on the health index degradation model, an inverse Gaussian distribution is used to describe the remaining lifespan distribution of equipment. The inverse Gaussian distribution is a probability distribution commonly used to describe the time to reach a certain threshold; its probability density function is determined by a drift parameter and a diffusion parameter. The drift parameter reflects the average rate of health index degradation, while the diffusion parameter reflects the degree of random fluctuation in the degradation process. After fitting historical data to obtain the drift and diffusion parameters for each piece of equipment, the remaining lifespan probability distribution of that equipment can be determined. For example, if a transformer has a drift parameter of 0.05 and a diffusion parameter of 0.01, its expected remaining lifespan is 20 years, with a standard deviation of approximately 4.5 years.

[0111] Finally, the Monte Carlo method is used to simulate the cascading effects of equipment failure. First, independent random sampling is performed from the remaining lifetime distribution of each device to obtain a sampled remaining lifetime value for each device. Devices with a sampled remaining lifetime less than the assessment time window are marked as the initial fault set. Starting from the initial fault set, power flow calculations are run to update the grid power distribution. Lines overloaded due to power flow shifts are added to the fault set, and the operation of protection devices is simulated according to protection action logic. This process is repeated until no new faulty devices are added or the maximum number of iterations is reached. The above sampling and simulation process is repeated N times, and the frequency of each device appearing in the final fault set is counted as the daily risk score for that device. For example, with an assessment time window of 24 hours and N sampling times of 10,000, if a device appears in the final fault set in 2,500 samplings, its daily risk score is 0.25. This score comprehensively considers the cumulative damage and degradation of the device itself and the cascading propagation effect after the failure.

[0112] The weekly channel, based on topological features and with a time step of one day, calculates indicators such as the rate of change of betweenness centrality, ranking fluctuation of risk amplification factor, graph connectivity resilience, and minimum cut set risk to assess the evolution trend of system structural vulnerability and output a weekly risk score for each power element node. Betweenness centrality reflects the frequency with which a node acts as a bridge in all shortest paths. The formula for betweenness centrality is: the number of shortest paths passing through node i divided by the total number of shortest paths between all node pairs. A higher betweenness centrality indicates a more important bridging role for the node in the network, and a greater impact of its failure on system connectivity. The rate of change of betweenness centrality refers to the relative change in a node's betweenness centrality at the current moment compared to one week ago. The formula is: the difference between the current value and the value one week ago divided by the value one week ago. A positive rate of change indicates that the node's importance in the network is increasing.

[0113] Risk amplification factor ranking fluctuation refers to the magnitude of change in a node's risk amplification factor ranking within one week. The risk amplification factor reflects a node's critical transmission position in the disaster chain; a higher ranking indicates a more important node. Greater ranking fluctuation indicates a rapidly changing criticality for the node, requiring close monitoring. The formula for calculating ranking fluctuation is the absolute value of subtracting the ranking from one week ago from the current ranking.

[0114] Graph connectivity resilience refers to the probability that a system remains connected after removing several of the highest-risk nodes. The specific calculation method is as follows: nodes are sorted from highest to lowest risk amplification factor, and the top K highest-risk nodes are removed sequentially. After each removal, the system's connectivity is checked. The connectivity check is implemented using the connected component algorithm in graph theory. If the system remains connected after removal, the resilience is high; if the system splits into multiple isolated subgraphs, the resilience is low. The resilience score is 1 minus the degree of splitting.

[0115] Minimal cut set risk refers to the minimum risk cost required to remove nodes or edges that resolves a system. A cut set is a set of nodes or edges that, when removed, splits the system into two or more disconnected subgraphs. The minimum cut set is the one with the lowest risk cost among all cut sets. The minimum cut set risk is calculated using the minimum cut theorem in network flow algorithms, where node risk values ​​are used as node weights to find the minimum weighted cut set that resolves the system.

[0116] Finally, the indicators such as the rate of change of betweenness centrality, the ranking volatility of the risk amplification factor, the graph connectivity resilience, and the minimum cut set risk are normalized and then summed according to preset weights to obtain the weekly risk score. For example, the weight of the rate of change of betweenness centrality is set to 0.3, the weight of the ranking volatility of the risk amplification factor is set to 0.3, the weight of the graph connectivity resilience is set to 0.2, and the weight of the minimum cut set risk is set to 0.2.

[0117] By using the parallel early warning systems of the four channels mentioned above, multi-scale, comprehensive monitoring of disaster chain risks can be achieved, ranging from minute-level to week-level. The minute-level channel captures immediate electrical anomalies, the hour-level channel captures weather-driven equipment downtime risks, the day-level channel captures equipment degradation risks caused by cumulative damage, and the week-level channel captures the vulnerability evolution trends of the system topology. The outputs of the four channels are then spatiotemporally fused to form a comprehensive risk matrix.

[0118] Secondly, the channel risk scores of each power element node output by the multi-scale parallel early warning framework are normalized and multiplied by preset scenario channel weights and a normalized risk amplification factor, then summed to obtain the spatial fusion risk value of each power element node. Since the risk scores output by different channels may have different dimensions and numerical ranges, normalization is first performed to map each channel score to the 0-1 range. The preset scenario channel weights are used to adjust the importance of different early warning channels in spatial fusion and can be dynamically adjusted according to the current operating scenario. For example, in typhoon weather, the weight of geographically related channels should be increased; in communication failure scenarios, the weight of information-related channels should be decreased. The normalized risk amplification factor serves as a spatial correlation weighting coefficient, giving nodes at key transmission positions in the disaster chain higher risk weights. Multiplying the channel risk score by the scenario channel weight and then by the risk amplification factor, and summing the results, yields the spatial fusion risk value. This spatial fusion risk value comprehensively reflects the total amount of risk received by the node from different risk sources at the current moment.

[0119] Furthermore, based on the prior time risk regression model corresponding to each early warning channel, the output results of each early warning channel are extrapolated along their respective time scales to obtain the predicted risk increment of each power element node at multiple scales.

[0120] The prior time-based risk regression model is a time-series prediction model pre-built based on historical risk data, used to describe the trend of risk scores for each power element node over time. An independent regression model is established for each early warning channel and each power element node. The training data for the model comes from the historical risk score sequence of that node over a past period, for example, data from the past 30 days.

[0121] Taking a linear regression model as an example, the input is the time step number, and the output is a risk score. The linear regression model takes the form: the risk score equals the intercept term plus the slope coefficient multiplied by the time step number. The slope coefficient reflects the trend direction of risk over time: a positive slope indicates an upward trend, a negative slope indicates a downward trend, and a larger absolute value of the slope indicates a more pronounced trend. Model parameters are obtained by fitting historical data using the least squares method. Taking an exponential smoothing model as another example, the model captures the trend component of the time series using a weighted average. The exponential smoothing model includes three parameters: a level parameter, a trend parameter, and a seasonality parameter. The level parameter controls the baseline value of the series, the trend parameter controls the upward or downward trend of the series, and the seasonality parameter controls periodic fluctuations. Model parameters are obtained through maximum likelihood estimation of historical data.

[0122] For minute-level channels, the historical risk score sequence of the most recent 60 minutes is taken, with a time step of 1 minute, totaling 60 time points. Using the time step number as the independent variable and the risk score as the dependent variable, a linear regression model or an exponential smoothing model is fitted. The risk change trend over the next 5 to 15 minutes is extrapolated, i.e., the risk score is predicted for the next 5, 10, and 15 minutes. The predicted risk increment equals the predicted risk score at the future time minus the risk score at the current time. For example, if the current risk score is 0.3, and the model predicts a risk score of 0.45 10 minutes from now, then the predicted risk increment is 0.15, indicating an upward trend in risk.

[0123] For hourly channels, the historical risk score sequence of the most recent 24 hours is taken, with a time step of 15 minutes, totaling 96 time points. The risk change trend is extrapolated for the next 1 to 4 hours, i.e., the risk score is predicted for the next 1, 2, 3, and 4 hours. The predicted risk increment equals the predicted risk score at the future time minus the risk score at the current time.

[0124] For the daily-level channel, the historical risk score sequence of the most recent 30 days is taken, with a time step of 1 hour, totaling 720 time points. The risk change trend for the next 1 to 7 days is extrapolated, i.e., the risk scores are predicted for the next 1, 3, and 7 days. The predicted risk increment equals the predicted risk score at the future time minus the risk score at the current time.

[0125] For weekly channels, the historical risk score sequence of the most recent 52 weeks is used, with a time step of 1 day, totaling 364 time points. The risk change trend over the next 1 to 4 weeks is extrapolated, i.e., the risk scores are predicted for the next 1, 2, 3, and 4 weeks. The predicted risk increment equals the predicted risk score at future moments minus the risk score at the current moment.

[0126] Specifically, a positive predicted risk increment indicates that the risk at that node is trending upwards in the future, requiring enhanced monitoring; a negative predicted risk increment indicates a downward trend, allowing for a reduction in attention; the larger the absolute value of the predicted risk increment, the more drastic the trend change. By introducing the predicted risk increment, spatiotemporal fusion not only considers the current risk status but also predicts the future evolution direction of the risk, giving the early warning system foresight. For example, if a node currently has a low risk score, but the model predicts that the risk score will increase by 0.2 within the next hour, it indicates a possibility of rapid deterioration, triggering an early warning. Using this method, the predicted risk increments of each power element node at multiple scales can be obtained, which can then be superimposed with the spatially fused risk values ​​to generate a comprehensive risk matrix.

[0127] Finally, the spatial fusion risk value and the predicted risk increment are superimposed to generate a comprehensive risk matrix for each power element node. The spatial fusion risk value at the current moment is added to the predicted risk increment at each time scale to obtain the comprehensive risk score for that node at different future time scales. The rows of the comprehensive risk matrix correspond to power element nodes, and the columns correspond to four time scales: minute, hour, day, and week. Matrix elements represent the comprehensive risk score for that node at the corresponding time scale. This comprehensive risk matrix includes both the current risk status of the node and the evolution trend of risk at various future time scales. It provides a unified format of input features for the subsequent variational autoencoder anomaly detection module and provides dispatchers with multi-time-dimensional risk situation awareness.

[0128] S40: Combining the comprehensive risk matrix, the spatiotemporal embedding representation output by the graph neural network, and the target disaster chain transmission path, the anomaly score is calculated through the pre-constructed encoder network, and the warning level is determined based on the anomaly score, and the risk warning result is output.

[0129] Furthermore, by combining the comprehensive risk matrix, the spatiotemporal embedding representation output by the graph neural network, and the target disaster chain propagation path, anomaly scores are calculated using a pre-constructed encoder network. The comprehensive risk matrix contains the comprehensive risk scores of each power element node at four time scales: minute, hour, day, and week, reflecting the current risk status and future evolution trend of the node. The spatiotemporal embedding representation is the node feature vector output by the graph neural network after multi-layer heterogeneous attention message passing and time dimension modeling, containing the node's location information and risk propagation relationship in the topology. The target disaster chain propagation path contains the complete path sequence of risk propagation from the source node to downstream nodes along the topology, the cumulative propagation risk, the estimated total propagation time, and confidence information.

[0130] The encoder network employs a variational autoencoder architecture. By learning the distribution characteristics of historical normal operation data, it compares the current state with the normal mode to calculate an anomaly score. The anomaly score is a comprehensive indicator that quantitatively assesses the degree to which the current power grid operation deviates from the normal operation mode, reflecting the likelihood of a disaster chain risk in the system. A higher anomaly score indicates a more significant deviation from the normal mode, and a greater probability of a disaster chain risk occurring. Finally, based on the anomaly score, a warning level is determined, and a risk warning result is output.

[0131] Specifically, combining the comprehensive risk matrix, the spatiotemporal embedding representation output by the graph neural network, and the target disaster chain propagation path, anomaly scores are calculated through a pre-constructed encoder network, and a warning level is determined based on the anomaly scores, outputting a risk warning result, including: The comprehensive risk matrix, the spatiotemporal embedding representation, and the structured features of the target disaster chain path are concatenated into a comprehensive feature vector; The synthesized feature vector is input into the encoder network to obtain the mean vector and variance vector under the latent representation; The mean vector and variance vector are parameterized and sampled, and the encoder network and corresponding decoder network perform input feature reconstruction. The sum of reconstruction error and latent space deviation is calculated as an anomaly score. A triple adaptive threshold mechanism is used for early warning determination. This mechanism selects the most stringent of the following threshold values ​​as the triple adaptive threshold: The first layer is a statistical threshold based on a sliding window, which dynamically sets the threshold according to the mean and standard deviation of the outlier scores within the window; The second layer is the dual-window comparison detection threshold, which sets the threshold by comparing the offset of the mean abnormal scores of the short time window and the long time window. The third layer is the quantile dynamic calibration threshold based on the empirical cumulative distribution function; The abnormal score is compared with the triple adaptive threshold to determine the warning level, and the warning signal and the list of affected nodes are output, wherein the list of affected nodes is arranged in descending order according to the risk amplification factor.

[0132] First, the comprehensive risk matrix, spatiotemporal embedding representation, and structured features of the target disaster chain path are concatenated into a comprehensive feature vector. The comprehensive risk matrix contains the comprehensive risk scores of each power element node at four time scales: minute, hour, day, and week, which are flattened into a one-dimensional vector in node order. The spatiotemporal embedding representation is the node feature vector output by the graph neural network, reflecting the node's location information and risk propagation relationship in the topology. The structured features of the target disaster chain path include the node sequence on the path, cumulative transmission risk, estimated total transmission time, and confidence level, which are encoded into a fixed-length feature vector. The above three parts are concatenated in sequence to form the comprehensive feature vector, which serves as the input to the encoder network.

[0133] Next, the synthesized feature vector is input into the encoder network to obtain the mean and variance vectors under the latent representation. For example, the encoder network adopts a four-layer fully connected network structure, with the following input and output dimensions for each layer: The first layer has the total dimension of the synthesized feature vector as input, for example, 512 dimensions, and an output dimension of 256 dimensions, using the ReLU activation function and followed by a Dropout layer with a dropout rate of 0.2 to prevent overfitting; the second layer has the same input dimension of 256 dimensions and an output dimension of 128 dimensions, using the ReLU activation function and followed by a Dropout layer with a dropout rate of 0.2; the third layer has the same input dimension of 128 dimensions and an output dimension of 64 dimensions, using the ReLU activation function; the fourth layer splits into two parallel output heads: the mean head has an output dimension of 32 dimensions and uses a linear activation function, and the log-variance head has an output dimension of 32 dimensions and uses a linear activation function.

[0134] The encoder network outputs two vectors: a mean vector and a log-variance vector. The mean vector represents the central location of the latent variable, and the log-variance vector represents the range of uncertainty of the latent variable. Parameterized sampling is performed on the mean and log-variance vectors to generate the latent vector. The sampling method employs a reparameterization technique: a noise vector is randomly sampled from a standard normal distribution, multiplied by the exponential half of the variance vector, and then added to the mean vector to obtain the latent vector. This operation enables the model to perform gradient backpropagation.

[0135] The decoder network adopts a four-layer fully connected network structure symmetrical to the encoder. The input layer receives a 32-dimensional latent vector sampled from the latent space. The first layer has an output dimension of 64 dimensions and uses the ReLU activation function; the second layer has an output dimension of 128 dimensions and uses the ReLU activation function; the third layer has an output dimension of 256 dimensions and uses the ReLU activation function; the output layer has the same output dimension as the input comprehensive feature vector, for example, 512 dimensions, and uses the Sigmoid activation function to constrain the output value to the interval between 0 and 1.

[0136] The encoder network and its corresponding decoder network reconstruct the input features, calculating the sum of the reconstruction error and the latent space deviation as the anomaly score. The decoder network maps the latent vector back to the original feature space, outputting a reconstructed composite feature vector. The reconstruction error equals the square of the Euclidean distance between the input features and the reconstructed features, reflecting the degree of deviation between the current state and the normal operating mode learned by the variational autoencoder. The latent space deviation equals the KL divergence between the latent distribution and the standard normal distribution, reflecting the distance between the current state's position in the latent space and the normal region. The anomaly score is obtained by weighted summing the reconstruction error and the latent space deviation. A higher anomaly score indicates a greater deviation of the current system state from the normal operating mode, and a higher probability of a hazard chain risk.

[0137] Finally, a triple adaptive threshold mechanism is used for early warning determination. This mechanism includes a first threshold based on a sliding window statistical threshold, a second threshold based on a dual-window comparison detection threshold, and a third threshold based on a dynamic quantile calibration threshold based on an empirical cumulative distribution function. The most stringent of the three thresholds is then selected as the final decision threshold.

[0138] The first layer is a sliding window-based statistical threshold: A fixed-length sliding window for outlier scores is maintained, for example, a window length of 1440 minutes corresponding to the past 24 hours. The mean and standard deviation of the outlier scores within the window are calculated, and the mean plus three times the standard deviation is used as the statistical threshold. This threshold is dynamically updated as the window slides, adapting to the gradual changes in the operating status.

[0139] The second layer of dual-window comparison detection threshold: Maintain two sliding windows of different lengths, such as a short window of 30 minutes and a long window of 1440 minutes. Calculate the difference between the mean abnormal scores within the two windows. When the difference exceeds a preset offset threshold, a trend warning is triggered. This mechanism can identify rapid changes in operating patterns even if the absolute value of the abnormal score does not exceed the statistical threshold.

[0140] The third quantile dynamic calibration threshold is based on the empirical cumulative distribution function: the empirical cumulative distribution function of outlier scores is calculated within a sliding window, and the 99th percentile is taken as the quantile threshold. This threshold does not depend on the normal distribution assumption and is suitable for scenarios where outlier score distribution is asymmetric. The maximum value among the three thresholds is used as the final decision threshold.

[0141] Finally, the anomaly score is compared with a triple adaptive threshold to determine the warning level. For example, when the anomaly score exceeds 1.2 times the judgment threshold, it is judged as a Level 1 Red Warning, indicating that the system is in a high-risk state and the risk is rapidly evolving, requiring immediate activation of the emergency plan; when the anomaly score exceeds the judgment threshold but is less than 1.2 times, it is judged as a Level 2 Orange Warning, indicating that the system is in an abnormal state but the risk is relatively stable, requiring close monitoring and preparation of emergency measures; when the anomaly score does not exceed the judgment threshold but the dual-window offset exceeds the preset threshold, it is judged as a Level 3 Yellow Warning, indicating that the system operating mode is changing and there is a potential risk escalation, requiring strengthened inspection and pre-control.

[0142] Simultaneously, it outputs early warning signals and a list of affected nodes. The list of affected nodes is arranged in descending order of risk amplification factor, listing the nodes with the highest risk, their comprehensive risk score, risk amplification factor, and the disaster chain path to which they belong. It also outputs the node sequence, cumulative transmission risk, estimated total transmission time and confidence level for each disaster chain transmission path, as well as targeted handling recommendations.

[0143] In summary, the above steps enable dynamic early warning and precise location of disaster chain risks in cross-regional power distribution networks, providing dispatchers with actionable decision support.

[0144] In summary, the embodiments of this application have at least the following technical effects: This invention first constructs a heterogeneous spatiotemporal graph model to uniformly model diverse nodes in inter-regional power distribution networks, including substations, lines, loads, distributed power sources, and protection automation devices. It designs four types of heterogeneous edges: electrical connection edges, information communication edges, geographical proximity edges, and logical dependency edges, each corresponding to different risk transmission semantics, thus solving the problem that existing homogeneous graph models cannot distinguish differentiated risk transmission paths. Secondly, through the heterogeneous attention message passing mechanism of graph neural networks, combined with the time decay kernel encoding delay effect, it learns the propagation weights of risk in the topology and calculates the risk amplification factor. This achieves accurate extraction of disaster chain transmission paths and quantitative assessment of risk amplification effects, addressing the problem of insufficient characterization of the dynamic transmission and amplification mechanisms of disaster chains.

[0145] Furthermore, a four-channel parallel early warning framework at the minute, hour, day, and week levels is established to perform spatiotemporal fusion of multi-scale early warning results. A variational autoencoder is introduced to calculate anomaly scores, and a triple adaptive threshold mechanism is used to dynamically determine the early warning level. This solves the problems of weak multi-risk coupling modeling capabilities and high false alarm / missed alarm rates caused by fixed thresholds. This invention achieves dynamic early warning and precise location of disaster chain risks in cross-regional power distribution networks.

[0146] Example 2, as Figure 2 As shown, based on the same inventive concept as the cross-regional distribution network disaster chain risk early warning method based on graph neural networks provided in Embodiment 1, this embodiment of the invention also provides a cross-regional distribution network disaster chain risk early warning system based on graph neural networks, including: The data acquisition module 11 is used to collect multi-source management data of cross-regional power distribution network through multi-source data sources, preprocess the multi-source management data, and construct multi-dimensional feature vectors of multiple power element nodes and initial weight parameters of each associated edge based on the preprocessing results. The heterogeneous spatiotemporal graph model module 12 is used to construct a heterogeneous spatiotemporal graph model based on the distribution network topology in the multi-source management data, combined with the multi-dimensional feature vector and the initial weight parameters, and to obtain the risk amplification factor of each power element node and the transmission path of multiple target disaster chains by performing message passing and time dimension modeling through graph neural network. The multi-scale early warning fusion module 13 is used to combine the risk amplification factor with the multi-source management data to perform multi-scale parallel early warning for multiple target disaster chain transmission paths, and to perform spatiotemporal fusion of the multi-scale parallel early warning results to obtain a comprehensive risk matrix. The early warning level determination module 14 is used to combine the comprehensive risk matrix, the spatiotemporal embedding representation output by the graph neural network, and the target disaster chain transmission path, calculate the anomaly score through the pre-constructed encoder network, determine the early warning level based on the anomaly score, and output the risk early warning result.

[0147] Specifically, the data acquisition module 11 is used for: Specifically, multi-source management data of cross-regional distribution networks is collected through multi-source data sources. This multi-source management data is preprocessed, and based on the preprocessing results, multi-dimensional feature vectors of multiple power element nodes and initial weight parameters for each associated edge are constructed, including: Multidimensional operational data of the cross-regional distribution network are collected from the distribution network dispatch automation system, meteorological monitoring and forecasting system, equipment asset management system and equipment status online monitoring system. The multidimensional operational data includes at least: electrical measurement data, meteorological environment data, equipment topology and ledger data, equipment status monitoring data and historical fault event record data. The multidimensional operating data is preprocessed, and feature extraction is performed based on the preprocessing results to obtain the multidimensional feature vector of each power element node. The multidimensional feature vector includes at least electrical operating features, environmental status features, and static attribute features. Based on the meteorological environment data, the initial weight parameters of each associated edge are determined through the prior meteorological-vulnerability mapping curve.

[0148] Specifically, the heterogeneous spatiotemporal graph model module 12 is used for: Specifically, based on the distribution network topology in the multi-source management data, a heterogeneous spatiotemporal graph model is constructed by combining the multi-dimensional feature vectors and the initial weight parameters. A graph neural network is then used for message passing and time dimension modeling to obtain the risk amplification factor of each power element node and the transmission paths of multiple target disaster chains, including: The distribution network topology is obtained, and a heterogeneous spatiotemporal graph model containing multiple node types and multiple edge types is constructed by combining the multidimensional feature vectors and the initial weight parameters. By combining preset multi-level early warning targets, the heterogeneous spatiotemporal graph model is divided into discrete time steps, generating a corresponding multi-granularity graph snapshot sequence: The pre-trained hierarchical spatiotemporal graph attention network performs forward propagation based on multi-layer heterogeneous attention message passing based on the multi-granularity graph snapshot sequence to obtain the spatiotemporal embedding representation of each power element node; Based on the spatiotemporal embedding representation and combined with the multi-source management data, the risk amplification factor of each power element node is calculated and obtained; The risk amplification factor traverses the heterogeneous spatiotemporal graph model to perform reverse path search, obtains multiple target disaster chain transmission paths, and represents them in a structured manner, including at least: node sequence, cumulative transmission risk, estimated total transmission time, and historical hit rate based on historical data.

[0149] In the heterogeneous spatiotemporal graph model, the node types include at least substation nodes, line nodes, load nodes, distributed power generation nodes, and protection automation device nodes; the edge types include at least electrical connection edges, information communication edges, geographical proximity edges, and logical dependency edges, and each edge type corresponds to independent message transformation parameters and attention calculation parameters.

[0150] The pre-trained hierarchical spatiotemporal graph attention network performs forward propagation based on multi-layer heterogeneous attention message passing, using the multi-granularity graph snapshot sequence, to obtain the spatiotemporal embedding representation of each power element node, including: For each edge type, a type-specific linear transformation function is used to transform the features of neighboring nodes into message vectors; The message vector is weighted by time delay using a time decay function constructed based on a time decay kernel; The attention weights of the neighboring nodes of the power element node are calculated through a multi-head attention mechanism. Based on the time-delay weighting results, intra-class aggregation based on edge type and gated cross-type fusion are performed sequentially to obtain the fused node embedding. The fusion nodes at each time step are embedded into the pre-trained time dimension modeling module to obtain the spatiotemporal embedding representation of each power element node, wherein the time dimension modeling module is established based on the time series architecture.

[0151] Specifically, based on the spatiotemporal embedding representation and combined with the multi-source management data, the risk amplification factor for each power element node is calculated and obtained, including: Based on the spatiotemporal embedding representation, the basic risk value of each power element node is calculated through a risk scoring function model; Based on the meteorological environment data and equipment status monitoring data in the multi-source management data, the physical vulnerability coefficient of each power element node is calculated; Using the attention weight between each power element node and its neighbor nodes as the weighting weight, the basic risk values ​​of multiple neighbor nodes and the corresponding physical vulnerability coefficients are weighted and summed to obtain the neighbor risk weighting value for each power element node. Based on the basic risk value and the corresponding neighbor risk weighting value, a comprehensive risk value is calculated, and the ratio of the comprehensive risk value to the basic risk value is output as the risk amplification factor for each power element node.

[0152] The risk amplification factor involves traversing the heterogeneous spatiotemporal graph model to perform a reverse path search, obtaining multiple target disaster chain transmission paths, including: The power element nodes ranked at the top preset quantile of the risk amplification factor are taken as the source node set; Starting from each source node, perform a beam search along the associated edge with the largest attention weight; During the search process, the beam search results are pruned and delooped based on the following: the cumulative transmission risk is lower than the preset risk threshold, the cumulative transmission time exceeds the preset time limit, and duplicate power element nodes appear in the path. The paths with the highest cumulative transmission risk values ​​are retained as the target disaster chain transmission paths, and the confidence level is calculated for each target disaster chain transmission path based on the hit rate of similar paths in historical disaster event data.

[0153] The multi-scale early warning fusion module 13 is specifically used for: Specifically, combining the risk amplification factor with the multi-source management data, multi-scale parallel early warning is performed on multiple target disaster chain transmission paths, and the multi-scale parallel early warning results are spatiotemporally fused to obtain a comprehensive risk matrix, further including: Establish a multi-scale parallel early warning framework, including at least: minute-level channels, hour-level channels, day-level channels and week-level channels, which output channel risk scores for each power element node in parallel based on input data of different time granularities and different risk assessment logics. The channel risk scores of each power element node output by the multi-scale parallel early warning framework are normalized, and then multiplied by the preset scene channel weight and the normalized risk amplification factor, and summed to obtain the spatial fusion risk value of each power element node. Based on the prior time risk regression model corresponding to each early warning channel, the output results of each early warning channel are extrapolated along their respective time scales to obtain the predicted risk increment of each power element node at multiple scales. The spatial fusion risk value and the predicted risk increment are superimposed to generate a comprehensive risk matrix for each power element node.

[0154] The warning level determination module 14 is specifically used for: Specifically, combining the comprehensive risk matrix, the spatiotemporal embedding representation output by the graph neural network, and the target disaster chain propagation path, anomaly scores are calculated through a pre-constructed encoder network, and a warning level is determined based on the anomaly scores, outputting a risk warning result, including: The comprehensive risk matrix, the spatiotemporal embedding representation, and the structured features of the target disaster chain path are concatenated into a comprehensive feature vector; The synthesized feature vector is input into the encoder network to obtain the mean vector and variance vector under the latent representation; The mean vector and variance vector are parameterized and sampled, and the encoder network and corresponding decoder network perform input feature reconstruction. The sum of reconstruction error and latent space deviation is calculated as an anomaly score. A triple adaptive threshold mechanism is used for early warning determination. This mechanism selects the most stringent of the following threshold values ​​as the triple adaptive threshold: The first layer is a statistical threshold based on a sliding window, which dynamically sets the threshold according to the mean and standard deviation of the outlier scores within the window; The second layer is the dual-window comparison detection threshold, which sets the threshold by comparing the offset of the mean abnormal scores of the short time window and the long time window. The third layer is the quantile dynamic calibration threshold based on the empirical cumulative distribution function; The abnormal score is compared with the triple adaptive threshold to determine the warning level, and the warning signal and the list of affected nodes are output, wherein the list of affected nodes is arranged in descending order according to the risk amplification factor.

Claims

1. A method for early warning of disaster chain risks in cross-regional power distribution networks based on graph neural networks, characterized in that, include: Multi-source management data of cross-regional power distribution networks are collected through multiple data sources. The multi-source management data is preprocessed, and based on the preprocessing results, multi-dimensional feature vectors of multiple power element nodes and initial weight parameters of each associated edge are constructed. Based on the distribution network topology in the multi-source management data, a heterogeneous spatiotemporal graph model is constructed by combining the multi-dimensional feature vector and the initial weight parameters. Message passing and time dimension modeling are performed through graph neural networks to obtain the risk amplification factor of each power element node and the transmission path of multiple target disaster chains. By combining the risk amplification factor with the multi-source management data, multi-scale parallel early warning is performed on multiple target disaster chain transmission paths, and the results of the multi-scale parallel early warning are spatiotemporally fused to obtain a comprehensive risk matrix; Combining the comprehensive risk matrix, the spatiotemporal embedding representation output by the graph neural network, and the target disaster chain transmission path, anomaly scores are calculated through a pre-constructed encoder network, and the warning level is determined based on the anomaly scores, outputting risk warning results. This involves collecting multi-source management data from cross-regional power distribution networks through multiple data sources, preprocessing the multi-source management data, and constructing multi-dimensional feature vectors for multiple power element nodes and initial weight parameters for each associated edge based on the preprocessing results, including: Multidimensional operational data of the cross-regional distribution network are collected from the distribution network dispatch automation system, meteorological monitoring and forecasting system, equipment asset management system and equipment status online monitoring system. The multidimensional operational data includes at least: electrical measurement data, meteorological environment data, equipment topology and ledger data, equipment status monitoring data and historical fault event record data. The multidimensional operating data is preprocessed, and feature extraction is performed based on the preprocessing results to obtain the multidimensional feature vector of each power element node. The multidimensional feature vector includes at least electrical operating features, environmental status features, and static attribute features. Based on the meteorological environment data, the initial weight parameters of each associated edge are determined through the prior meteorological-vulnerability mapping curve. Specifically, based on the distribution network topology in the multi-source management data, a heterogeneous spatiotemporal graph model is constructed by combining the multi-dimensional feature vectors and the initial weight parameters. A graph neural network is then used for message passing and time dimension modeling to obtain the risk amplification factor of each power element node and the transmission paths of multiple target disaster chains, including: The distribution network topology is obtained, and a heterogeneous spatiotemporal graph model containing multiple node types and multiple edge types is constructed by combining the multidimensional feature vector with the initial weight parameters. By combining preset multi-level early warning targets, the heterogeneous spatiotemporal graph model is divided into discrete time steps, corresponding to the generation of multi-granularity graph snapshot sequences: The pre-trained hierarchical spatiotemporal graph attention network performs forward propagation based on multi-layer heterogeneous attention message passing based on the multi-granularity graph snapshot sequence to obtain the spatiotemporal embedding representation of each power element node; Based on the spatiotemporal embedding representation and combined with the multi-source management data, the risk amplification factor of each power element node is calculated and obtained; The risk amplification factor traverses the heterogeneous spatiotemporal graph model to perform reverse path search, obtains multiple target disaster chain transmission paths, and represents them in a structured manner, including at least: node sequence, cumulative transmission risk, estimated total transmission time, and historical hit rate based on historical data; The risk amplification factor involves traversing the heterogeneous spatiotemporal graph model to perform a reverse path search, obtaining multiple target disaster chain transmission paths, including: The power element nodes ranked at the top preset quantile of the risk amplification factor are taken as the source node set; Starting from each source node, perform a beam search along the associated edge with the largest attention weight; During the search process, the beam search results are pruned and delooped based on the following: the cumulative transmission risk is lower than the preset risk threshold, the cumulative transmission time exceeds the preset time limit, and duplicate power element nodes appear in the path. The paths with the highest cumulative transmission risk values ​​are retained as the target disaster chain transmission paths, and the confidence level is calculated for each target disaster chain transmission path based on the hit rate of similar paths in historical disaster event data.

2. The method for early warning of disaster chain risks in cross-regional distribution networks based on graph neural networks as described in claim 1, characterized in that, The pre-trained hierarchical spatiotemporal graph attention network performs forward propagation based on multi-layer heterogeneous attention message passing based on the multi-granularity graph snapshot sequence to obtain the spatiotemporal embedding representation of each power element node, including: For each edge type, a type-specific linear transformation function is used to transform the features of neighboring nodes into message vectors; The message vector is weighted by time delay using a time decay function constructed based on a time decay kernel; The attention weights of the neighboring nodes of the power element node are calculated through a multi-head attention mechanism. Based on the time-delay weighting results, intra-class aggregation based on edge type and gated cross-type fusion are performed sequentially to obtain the fused node embedding. The fusion nodes at each time step are embedded into the pre-trained time dimension modeling module to obtain the spatiotemporal embedding representation of each power element node, wherein the time dimension modeling module is established based on the time series architecture.

3. The method for early warning of disaster chain risks in cross-regional distribution networks based on graph neural networks as described in claim 1, characterized in that, Based on the spatiotemporal embedding representation and combined with the multi-source management data, the risk amplification factor for each power element node is calculated and obtained, including: Based on the spatiotemporal embedding representation, the basic risk value of each power element node is calculated through a risk scoring function model; Based on the meteorological environment data and equipment status monitoring data in the multi-source management data, the physical vulnerability coefficient of each power element node is calculated; Using the attention weight between each power element node and its neighbor nodes as the weighting weight, the basic risk values ​​of multiple neighbor nodes and the corresponding physical vulnerability coefficients are weighted and summed to obtain the neighbor risk weighting value for each power element node. Based on the basic risk value and the corresponding neighbor risk weighting value, a comprehensive risk value is calculated, and the ratio of the comprehensive risk value to the basic risk value is output as the risk amplification factor for each power element node.

4. The method for early warning of disaster chain risks in cross-regional distribution networks based on graph neural networks as described in claim 1, characterized in that, Combining the risk amplification factor with the multi-source management data, multi-scale parallel early warning is performed on multiple target disaster chain transmission paths, and the multi-scale parallel early warning results are spatiotemporally fused to obtain a comprehensive risk matrix, further including: Establish a multi-scale parallel early warning framework, including at least: minute-level channels, hour-level channels, day-level channels and week-level channels, which output channel risk scores for each power element node in parallel based on input data of different time granularities and different risk assessment logics. The channel risk scores of each power element node output by the multi-scale parallel early warning framework are normalized, and then multiplied by the preset scene channel weight and the normalized risk amplification factor, and summed to obtain the spatial fusion risk value of each power element node. Based on the prior time risk regression model corresponding to each early warning channel, the output results of each early warning channel are extrapolated along their respective time scales to obtain the predicted risk increment of each power element node at multiple scales. The spatial fusion risk value and the predicted risk increment are superimposed to generate a comprehensive risk matrix for each power element node.

5. The method for early warning of disaster chain risks in cross-regional distribution networks based on graph neural networks as described in claim 1, characterized in that, Combining the comprehensive risk matrix, the spatiotemporal embedding representation output by the graph neural network, and the target disaster chain propagation path, anomaly scores are calculated through a pre-constructed encoder network. Based on these anomaly scores, a warning level is determined, and a risk warning result is output, including: The comprehensive risk matrix, the spatiotemporal embedding representation, and the structured features of the target disaster chain transmission path are concatenated into a comprehensive feature vector; The synthesized feature vector is input into the encoder network to obtain the mean vector and variance vector under the latent representation; The mean vector and variance vector are parameterized and sampled, and the encoder network and corresponding decoder network perform input feature reconstruction. The sum of reconstruction error and latent space deviation is calculated as an anomaly score. A triple adaptive threshold mechanism is used for early warning determination. This mechanism selects the most stringent of the following threshold values ​​as the triple adaptive threshold: The first layer is a statistical threshold based on a sliding window, which dynamically sets the threshold according to the mean and standard deviation of the outlier scores within the window; The second layer is the dual-window comparison detection threshold, which sets the threshold by comparing the offset of the mean abnormal scores of the short time window and the long time window. The third layer is the quantile dynamic calibration threshold based on the empirical cumulative distribution function; The abnormal score is compared with the triple adaptive threshold to determine the warning level, and the warning signal and the list of affected nodes are output, wherein the list of affected nodes is arranged in descending order according to the risk amplification factor.

6. The method for early warning of disaster chain risks in cross-regional distribution networks based on graph neural networks as described in claim 1, characterized in that, In the heterogeneous spatiotemporal graph model, the node types include at least substation nodes, line nodes, load nodes, distributed power generation nodes, and protection automation device nodes; the edge types include at least electrical connection edges, information communication edges, geographical proximity edges, and logical dependency edges, and each edge type corresponds to independent message transformation parameters and attention calculation parameters.

7. A cross-regional power distribution network disaster chain risk early warning system based on graph neural networks, characterized in that, The method for implementing the cross-regional power distribution network disaster chain risk early warning method based on graph neural networks as described in any one of claims 1-6 includes: The data acquisition module is used to collect multi-source management data of cross-regional power distribution networks through multiple data sources, preprocess the multi-source management data, and construct multi-dimensional feature vectors of multiple power element nodes and initial weight parameters of each associated edge based on the preprocessing results. The heterogeneous spatiotemporal graph model module is used to construct a heterogeneous spatiotemporal graph model based on the distribution network topology in the multi-source management data, combined with the multi-dimensional feature vector and the initial weight parameters, and to obtain the risk amplification factor of each power element node and the transmission path of multiple target disaster chains by using a graph neural network for message passing and time dimension modeling. The multi-scale early warning fusion module is used to combine the risk amplification factor with the multi-source management data to perform multi-scale parallel early warning for multiple target disaster chain transmission paths, and to perform spatiotemporal fusion of the multi-scale parallel early warning results to obtain a comprehensive risk matrix. The early warning level determination module is used to combine the comprehensive risk matrix, the spatiotemporal embedding representation output by the graph neural network, and the target disaster chain transmission path, calculate the anomaly score through the pre-constructed encoder network, determine the early warning level based on the anomaly score, and output the risk early warning result.

Images