Method, device and equipment for identifying regional power grid resilience based on night light data

Abstract

The application discloses a kind of based on night light data regional power grid resilience identification method, device and equipment, method includes obtaining the night light remote sensing data and multi-source power grid data of target area, create multiple adjacency matrix, and generate target space-time sequence prediction model in combination with adaptive adjacency matrix training;Call model to carry out cascading failure simulation scanning to each power grid node in target area, extract the evolution balance step number and final light loss of each power grid node from simulation result;According to each evolution balance step number and final light loss, phase space clustering is carried out, determine multiple clustering centers, and based on its quadrant distribution characteristics, give the characteristic label corresponding to the characteristic cluster to which each clustering center belongs;According to characteristic label and cascading failure simulation result, extract cascading risk spread area, output the power grid resilience risk zoning map of target area, to accurately depict the dynamic cascading failure process of power grid under the spatiotemporal impact of typhoon persistence, realize the fine evaluation of power grid resilience.

Description

Technical Field

[0001] This invention relates to the field of power grid resilience identification technology, and in particular to a method, apparatus and equipment for regional power grid resilience identification based on nighttime light data. Background Technology

[0002] To support regional power grid disaster prevention and mitigation planning, disaster resistance reinforcement, and post-disaster emergency repair, nighttime light remote sensing data can be used to assess power grid resilience under typhoon disasters. The strong winds and torrential rains accompanying typhoons often lead to widespread power grid failures, and the spatiotemporal dynamic changes in nighttime light brightness can effectively characterize the power grid's operating status, providing an observational basis for analyzing the mechanisms of fault occurrence and propagation.

[0003] Existing solutions mainly focus on static differences in nighttime light brightness in disaster areas, preliminary extraction of power outage ranges, and time-series prediction of power grid status based on simple geographical topology.

[0004] However, the existing solutions mentioned above are often limited to a single spatial distance or basic physical connection, failing to accurately reflect the deep topological relationship of the power grid under complex terrain constraints and multi-dimensional functional attributes. Consequently, it is difficult to accurately depict the dynamic cascading failure process of the power grid under the continuous spatiotemporal impact of typhoons, and it is difficult to achieve a refined assessment of the power grid resilience. Summary of the Invention

[0005] This invention provides a method, apparatus, and equipment for identifying regional power grid resilience based on nighttime light data. It solves the technical problem that existing solutions are often limited to a single spatial distance or basic physical connection, failing to accurately reflect the deep topological relationship of the power grid under complex terrain constraints and multi-dimensional functional attributes. Consequently, it is difficult to accurately characterize the dynamic cascading failure process of the power grid under the continuous spatiotemporal impact of typhoons, and it is difficult to achieve a refined assessment of power grid resilience.

[0006] This invention provides a method for identifying the resilience of a regional power grid based on nighttime light data, comprising:

[0007] Acquire nighttime light remote sensing data and multi-source power grid data for the target area, and create multiple adjacency matrices;

[0008] The initial spatiotemporal sequence prediction model is trained using all the adjacency matrices and the adaptive adjacency matrix, and the target spatiotemporal sequence prediction model is generated.

[0009] Based on the target spatiotemporal sequence prediction model, a cascade failure simulation scan is performed on each power grid node in the target area, and the evolution equilibrium step number and final light loss corresponding to each power grid node are extracted according to the cascade failure simulation results.

[0010] Phase space clustering is performed based on the evolutionary equilibrium steps and the final light loss, multiple cluster centers are determined, and feature labels corresponding to the feature clusters to which each cluster center belongs are assigned based on the quadrant distribution characteristics of each cluster center.

[0011] Based on the feature labels and the cascade failure simulation results, the cascade risk propagation region is extracted, and the power grid resilience risk zoning map corresponding to the target region is output.

[0012] Optionally, the step of acquiring nighttime light remote sensing data and multi-source power grid data of the target area and creating multiple adjacency matrices includes:

[0013] Acquire nighttime light remote sensing data and multi-source power grid data for the target area;

[0014] The nighttime light remote sensing data is preprocessed to obtain preprocessed remote sensing data;

[0015] Based on the preprocessed remote sensing data and the multi-source power grid data, a geographical adjacency matrix, a functional adjacency matrix, and a semantic similarity adjacency matrix are constructed, respectively.

[0016] Optionally, the nighttime light remote sensing data includes pre-disaster nighttime light remote sensing data and post-disaster nighttime light remote sensing data; the step of training the initial spatiotemporal sequence prediction model using all the adjacency matrices and the adaptive adjacency matrix to generate the target spatiotemporal sequence prediction model includes:

[0017] A four-dimensional matrix feature pool is constructed using all the aforementioned adjacency matrices and the adaptive adjacency matrix;

[0018] Construct an initial spatiotemporal sequence prediction model;

[0019] Based on matrix ablation experiments, target adjacency matrix combinations are extracted from the feature pool of the four-dimensional matrix;

[0020] The initial spatiotemporal sequence prediction model is trained using the target adjacency matrix combination, the pre-disaster nighttime light remote sensing data, and the post-disaster nighttime light remote sensing data to generate the target spatiotemporal sequence prediction model.

[0021] Optionally, the step of constructing the initial spatiotemporal sequence prediction model includes:

[0022] Construct a hybrid graph convolutional layer;

[0023] Create and sequentially connect the exogenous feature embedding layer and the linear mapping layer;

[0024] The hybrid graph convolutional layer and the linear mapping layer are connected to the splicing layer;

[0025] The splicing layer, the attention-based GRU encoding / decoding module, and the fully connected layer are connected in sequence to obtain the initial spatiotemporal sequence prediction model.

[0026] Optionally, the step of performing cascaded failure simulation scanning on each power grid node in the target area based on the target spatiotemporal sequence prediction model, and extracting the evolutionary equilibrium step number and final light loss corresponding to each power grid node based on the cascaded failure simulation results, includes:

[0027] Using the target spatiotemporal sequence prediction model, rolling time series extrapolation is performed based on the interference-free initial nighttime light remote sensing data to generate a baseline prediction evolution sequence;

[0028] The power grid nodes within the target area are traversed, and each power grid node is taken as an attack target node. In each time step of the rolling simulation, the nighttime light remote sensing data corresponding to the attack target node is continuously forced to zero to generate forced input data.

[0029] Using the target spatiotemporal sequence prediction model, a single-step forward extrapolation is performed based on the forced input data to generate the propagation prediction evolution data for the current time step, and the deviation between the propagation prediction evolution data and the baseline prediction evolution sequence at the corresponding time step is calculated.

[0030] Determine whether the deviation of the other power grid nodes, excluding the target node, is greater than a preset dual filtering threshold, which includes a historical baseline fluctuation threshold and a model inherent error threshold.

[0031] If there are other power grid nodes whose deviation exceeds the dual filtering threshold, they are identified as newly added secondary failure nodes; the propagation prediction evolution data of the newly added secondary failure nodes at the current time step is extracted and injected as a state update value into the forced input data at the next time step; the process jumps to execute the step of using the target spatiotemporal sequence prediction model, performing single-step forward extrapolation based on the forced input data, and generating the propagation prediction evolution data for the current time step;

[0032] If there are no other power grid nodes with a deviation greater than the dual filtering threshold in the current time step, it is determined that the evolution of the current cascaded failure branch has reached a balanced state, and the rolling timing deduction of the current attack target node is stopped.

[0033] Based on the total number of simulation cycles required to reach the equilibrium state, and the final value of the network-wide nighttime light attenuation caused by all failed nodes, the evolutionary equilibrium steps and final light loss corresponding to the attack target node are extracted.

[0034] Optionally, the step of performing phase space clustering based on the evolutionary equilibrium steps and the final light loss to determine multiple cluster centers, and assigning feature labels to the feature clusters to which each cluster center belongs based on the four-quadrant distribution characteristics of each cluster center, includes:

[0035] After normalizing the evolutionary equilibrium steps and the final light loss, a two-dimensional spatiotemporal resilience evolution phase space is constructed with the evolutionary equilibrium steps as the horizontal axis and the final light loss as the vertical axis.

[0036] Based on the Euclidean distance between each pair of feature points in the two-dimensional spatiotemporal resilience evolution phase space, spatial clustering is performed on each feature point to obtain multiple feature clusters;

[0037] Extract the cluster centers of each of the aforementioned feature clusters;

[0038] Based on the four-quadrant distribution characteristics of each cluster center in the two-dimensional spatiotemporal resilience evolution phase space, feature labels corresponding to the feature clusters to which each cluster center belongs are assigned.

[0039] Optionally, the step of extracting the cascading risk propagation region based on the feature labels and the cascading failure simulation results, and outputting the power grid resilience risk zoning map corresponding to the target region, includes:

[0040] Locate the target cluster from the feature clusters based on each of the aforementioned feature labels;

[0041] Based on the target cluster and the cascade failure simulation results, the high-frequency failure topological edges corresponding to each feature point within the target cluster are determined.

[0042] Each feature point and each high-frequency failure topological edge within the target cluster are mapped to a geographic coordinate system, and the kernel density value corresponding to each high-frequency failure topological edge is calculated using a kernel density estimation algorithm to generate a cascaded failure risk density raster surface.

[0043] A strip-shaped region that meets the preset risk conditions is selected from the cascaded failure risk density grid surface as a cascaded risk propagation corridor;

[0044] Using the cascaded risk contagion corridor, all the feature clusters, and the feature labels, a power grid resilience risk zoning map corresponding to the target area is generated and output.

[0045] A second aspect of the present invention provides a regional power grid resilience identification device based on nighttime light data, comprising:

[0046] The matrix creation unit is used to acquire nighttime light remote sensing data and multi-source power grid data of the target area and create multiple adjacency matrices;

[0047] The model training unit is used to train the initial spatiotemporal sequence prediction model using all the adjacency matrices and the adaptive adjacency matrix, and to generate the target spatiotemporal sequence prediction model.

[0048] The feature extraction unit is used to perform cascade failure simulation scanning on each power grid node in the target area based on the target spatiotemporal sequence prediction model, and extract the evolution equilibrium step number and final light loss corresponding to each power grid node according to the cascade failure simulation results;

[0049] The feature label assignment unit is used to perform phase space clustering based on the evolutionary equilibrium steps and the final light loss, determine multiple cluster centers, and assign feature labels to the feature clusters to which each cluster center belongs based on the quadrant distribution characteristics of each cluster center.

[0050] The power grid resilience risk zoning map generation unit is used to extract the cascading risk propagation region based on the feature labels and the cascading failure simulation results, and output the power grid resilience risk zoning map corresponding to the target region.

[0051] A third aspect of the present invention provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the regional power grid resilience identification method based on nighttime light data as described in any of the first aspects of the present invention.

[0052] As can be seen from the above technical solutions, the present invention has the following advantages:

[0053] This process involves acquiring nighttime light remote sensing data and multi-source power grid data for the target area, creating multiple adjacency matrices, and training an adaptive adjacency matrix to generate a spatiotemporal sequence prediction model for the target area. The model is then used to perform cascade failure simulation scanning on each power grid node within the target area. The evolutionary equilibrium steps and final light loss for each node are extracted from the simulation results. Phase space clustering is performed based on these evolutionary equilibrium steps and final light loss to determine multiple cluster centers. Based on their quadrant distribution characteristics, feature labels are assigned to the feature clusters to which each cluster center belongs. Based on the feature labels and the cascade failure simulation results, cascade risk propagation areas are extracted, and a power grid resilience risk zoning map of the target area is output. This accurately depicts the dynamic cascade failure process of the power grid under the continuous spatiotemporal impact of typhoons, achieving a refined assessment of power grid resilience. Attached Figure Description

[0054] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1 A flowchart illustrating the steps of a regional power grid resilience identification method based on nighttime light data provided in this embodiment of the invention;

[0056] Figure 2 A model architecture diagram of an initial spatiotemporal sequence prediction model provided in an embodiment of the present invention;

[0057] Figure 3 A schematic diagram illustrating the simulation results verification of a target spatiotemporal sequence prediction model provided in an embodiment of the present invention;

[0058] Figure 4 A schematic diagram of the target node propagation effect in an example provided by an embodiment of the present invention;

[0059] Figure 5 A schematic diagram illustrating the result comparison of an example provided in an embodiment of the present invention;

[0060] Figure 6 This is a structural block diagram of a regional power grid resilience identification device based on nighttime light data, provided as an embodiment of the present invention.

[0061] Figure 7 This is a structural block diagram of an electronic device for performing a regional power grid resilience identification method based on nighttime light data, provided in an embodiment of the present invention. Detailed Implementation

[0062] This invention provides a method, apparatus, and device for identifying regional power grid resilience based on nighttime light data. It addresses the technical problem that existing solutions are often limited to a single spatial distance or basic physical connection, failing to accurately reflect the deep topological relationships of the power grid under complex terrain constraints and multi-dimensional functional attributes. Consequently, they are difficult to accurately characterize the dynamic cascading failure process of the power grid under the continuous spatiotemporal impact of typhoons, and thus difficult to achieve a refined assessment of power grid resilience.

[0063] To make the objectives, features, and advantages of this invention more apparent and understandable, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the embodiments described below are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0064] Please see Figure 1 , Figure 1 A flowchart illustrating the steps of a regional power grid resilience identification method based on nighttime light data, provided in an embodiment of the present invention.

[0065] This invention provides a method for identifying the resilience of a regional power grid based on nighttime light data, comprising:

[0066] Step 101: Obtain nighttime light remote sensing data and multi-source power grid data for the target area, and create multiple adjacency matrices;

[0067] Nighttime Light (NTL) remote sensing data refers to the radiation information of artificial light sources on the Earth's surface captured by optical sensors carried by polar-orbiting satellites or space stations under cloudless conditions at night. It is usually stored in the form of gridded digital images after radiometric calibration and geographical correction, and can characterize the power load status and power supply continuity of the area covered by the power grid.

[0068] Multi-source power grid data refers to multi-dimensional heterogeneous datasets related to the physical structure, operating environment, and disaster-causing disturbances of the power grid in the target area. These include, but are not limited to, four categories: power grid physical topology data, geographic environment data, typhoon-induced disaster impact factor time-series data, and disaster-bearing body basic attribute data. Specifically, power grid physical topology data includes substation coordinates, high-voltage line lengths, and voltage levels within the target area; geographic environment data includes elevation, slope, and maximum rainfall within the target area; typhoon-induced disaster impact factor time-series data is used to assist model learning and quantify the spatiotemporal disturbance characteristics of the external disaster environment on the power grid state, including wind force levels, dynamic spatial distances between nodes and the eye of the storm, and other typhoon time-series evolution data; disaster-bearing body basic attributes are socio-economic and infrastructure base data characterizing the vulnerability and disaster resistance of grid nodes, including but not limited to population density, GDP per capita, historical average nighttime light radiation baseline, impervious surface area ratio, road network density, distance to high-voltage lines and substations, power grid node density, and the area ratio of various electricity-consuming areas (industrial, commercial, residential, and public service).

[0069] An adjacency matrix, in the field of graph neural networks, is a square matrix that quantitatively expresses the connectivity and strength between nodes in a graph. The row and column indices of the matrix elements correspond to the node numbers, and the element values ​​reflect the weights of the edges between corresponding node pairs. A zero value indicates no direct connection. This adjacency matrix can include, but is not limited to, geographical adjacency matrices (characterizing spatial connectivity under terrain constraints), functional adjacency matrices (characterizing power grid topology under power transmission physical rules), and semantic similarity matrices (characterizing implicit correlations in disaster response patterns between non-adjacent nodes).

[0070] In this embodiment, long-term nighttime light remote sensing data of the target area, from a preset pre-disaster duration to a preset post-disaster duration, is acquired from a publicly available or authorized remote sensing data platform. Simultaneously, multi-source power grid data, including power grid physical topology data, geographic environment data, typhoon time-series evolution data, and basic attribute data of disaster-bearing bodies, are collected within the target area. After preprocessing the nighttime light remote sensing data and the multi-source power grid data, multiple adjacency matrices are constructed based on both.

[0071] In one example of the present invention, step 101 may include the following sub-steps:

[0072] Acquire nighttime light remote sensing data and multi-source power grid data for the target area;

[0073] Data preprocessing is performed on nighttime light remote sensing data to obtain preprocessed remote sensing data;

[0074] Based on preprocessed remote sensing data and multi-source power grid data, a geographic adjacency matrix, a functional adjacency matrix, and a semantically similar adjacency matrix are constructed, respectively.

[0075] In this embodiment, long-term nighttime light remote sensing data covering the pre-disaster stable period to the post-disaster recovery period is acquired. Simultaneously, multi-source power grid data on the power grid physical topology, geographical environment, typhoon-induced disaster time-series evolution, and basic attributes of disaster-bearing bodies within the target area are collected. The acquired NTL data is then subjected to directional preprocessing, including cloud removal and noise reduction, radiometric calibration, mutual correction, and spatial projection transformation, to generate preprocessed remote sensing data that matches the spatial resolution of the multi-source power grid data. Simultaneously, spatial coordinate registration, outlier removal, and feature normalization of the multi-source power grid data are completed.

[0076] Taking into account both spatial distance and terrain barriers between nodes, a terrain slope penalty factor is introduced to accurately reflect the physical connectivity cost and environmental exposure of transmission lines in complex terrains such as windward slopes and high altitudes. Using equal-resolution grid cells from preprocessed remote sensing data as the carriers of power grid nodes, a geographical adjacency matrix is ​​constructed based on geographic environment data using a Gaussian kernel function with elevation penalty. Specifically, the planar geographic coordinates of node i and node j can be... and The average elevations of their corresponding grids are respectively and First, calculate the squared horizontal Euclidean distance between the two nodes. Square of elevation difference :

[0077] ;

[0078] ;

[0079] Introducing a Gaussian kernel function with terrain penalty to calculate spatial elevation constraint weights :

[0080] ;

[0081] In the formula, To control the distance attenuation bandwidth factor of the spatial domain; It is the vertical elevation penalty coefficient, used to characterize the nonlinear amplification effect of terrain undulation on the cost of physical connectivity.

[0082] The above constraint weight matrix is ​​compared with the binary adjacency matrix representing the connection relationship of the basic space. After performing the Adamas product and eliminating self-loop effects, row normalization is applied to prevent gradient explosion caused by excessively high local node degrees, ultimately yielding the geographical adjacency matrix. :

[0083] ;

[0084] A functional adjacency matrix is ​​constructed based on a voltage driving force and resistance attenuation model using power grid physical topology data. Specifically, spatial intersection analysis is used to determine that grid nodes passing through the same high-voltage transmission line have basic physical connectivity. For collinear node pairs... Extract the voltage level V of its associated line and calculate the horizontal spatial distance between the two nodes. and terrain slope item :

[0085] ;

[0086] Introducing a combined resistance attenuation model based on voltage driving force and distance / slope, the functional connection weights are calculated. :

[0087] ;

[0088] In the formula, It is a voltage weighting index used to characterize the dominant role of high-voltage lines in energy transmission; The resistance attenuation index characterizes the penalty effect of line exposure and transmission loss under long-distance and steep terrain.

[0089] When multiple high-voltage lines connect two nodes, the maximum weight is taken as the final functional connection strength of that node pair.

[0090] ;

[0091] Similarly, after adding self-loops, the matrix is ​​normalized to obtain the final functional adjacency matrix. .

[0092] Based on the basic attribute data of disaster-bearing bodies, a semantically similar adjacency matrix is ​​constructed through cosine similarity calculation and adaptive K-Nearest Neighbor (KNN) selection. Specifically, the feature aggregation mechanism of graph neural networks can be used to extract feature vectors based on the basic attributes of disaster-bearing bodies in the region where a node is located, capturing implicitly related nodes that are not spatially adjacent but have highly consistent typhoon impact patterns. Let the original node feature matrix containing multi-dimensional basic attributes such as population, economy, and nighttime light benchmarks be X. To eliminate the influence of dimensions, the feature matrix is ​​first normalized by Min-Max and L2 norm to obtain standard feature vectors. Then, the cosine similarity between global nodes is calculated. :

[0093] ;

[0094] To ensure both graph sparsity and computational efficiency, an adaptive K-nearest neighbor algorithm is employed, retaining only the K most similar hidden neighbors for each node. By comprehensively considering isolated node penalties, average degree constraints, and global similarity scores, the optimal number of nearest neighbors is automatically determined. .

[0095] After selecting the optimal neighbors, the properties of the undirected graph are ensured through a symmetric operation that takes the maximum value:

[0096] ;

[0097] In the formula, This is the similarity matrix after KNN sparsification. Finally, self-loops are added and normalized to output the semantic similarity matrix. .

[0098] Step 102: Train the initial spatiotemporal sequence prediction model using the full adjacency matrix and the adaptive adjacency matrix to generate the target spatiotemporal sequence prediction model;

[0099] Adaptive adjacency matrices are a type of purely data-driven association matrix without prior topological constraints. They are generated dynamically during model training by randomly initializing two sets of learnable node embedding vectors and then performing inner product, nonlinear activation, and probability mapping. They are used to capture implicit spatiotemporal dependencies that cannot be covered by fixed multidimensional matrices.

[0100] The initial spatiotemporal sequence prediction model refers to the Adaptive Graph Convolutional Sequence-to-Sequence (AGC_Seq2Seq) model, which is a deep spatiotemporal prediction network that integrates graph convolutional spatial coding, exogenous variable embedding, and temporal encoding and decoding structures without completed parameter training.

[0101] The target spatiotemporal sequence prediction model refers to a mature prediction model that has converged and possesses reliable multi-step prediction capabilities after multiple rounds of iterative optimization.

[0102] In this embodiment, after obtaining multiple generated adjacency matrices, an adaptive adjacency matrix that automatically learns weights based on node historical NTL time-series data is introduced to construct a matrix feature pool containing four association dimensions.

[0103] During the encoding stage, the NTL features of historical moments are subjected to graph diffusion operations with each adjacency matrix, and the summation is used to obtain a feature tensor that integrates multi-source spatial constraints. At the same time, the typhoon impact factor is mapped through an independent linear embedding layer and concatenated with the spatial features, and then fed into the gate control cyclic unit encoder to generate the hidden state sequence of each time step.

[0104] In the decoding phase, typhoon forecast covariates for future time steps are read, and historical encoded states are adaptively aligned and weighted using an attention mechanism to form a context vector. The decoder's gated recurrent unit recursively outputs NTL predictions for multiple future steps based on this context vector and exogenous shock information. Using pre-disaster nighttime light remote sensing data, the deviation between predicted and actual values ​​is minimized using gradient descent, all network parameters are iteratively updated, and an ablation experiment is introduced to determine the optimal configuration combination of multiple adjacency matrices, ultimately obtaining the target spatiotemporal sequence prediction model. Through the fusion training of multidimensional fixed adjacency matrices and adaptive adjacency matrices, the model simultaneously captures the explicit physical topological constraints and implicit data-driven correlations of the power grid. Combined with the embedding and fusion of typhoon exogenous variables, the model accurately learns the temporal evolution patterns of nighttime lights at power grid nodes.

[0105] The node NTL evolution curve under the disaster-free assumption, as deduced by the model, is defined as the baseline predicted evolution sequence.

[0106] In one example of the present invention, the nighttime light remote sensing data includes pre-disaster nighttime light remote sensing data and post-disaster nighttime light remote sensing data; step 102 may include the following sub-steps S11-S14:

[0107] S11. Construct a four-dimensional matrix feature pool using the full adjacency matrix and the adaptive adjacency matrix;

[0108] S12. Construct an initial spatiotemporal sequence prediction model;

[0109] Furthermore, S12 may include the following sub-steps:

[0110] Construct a hybrid graph convolutional layer;

[0111] Create and sequentially connect the exogenous feature embedding layer and the linear mapping layer;

[0112] Connect the hybrid graph convolutional layer and the linear mapping layer to the stitching layer;

[0113] By sequentially connecting the splicing layer, the attention-based GRU encoding / decoding module, and the fully connected layer, an initial spatiotemporal sequence prediction model is obtained.

[0114] like Figure 2 As shown, Figure 2 This diagram illustrates the model architecture of an initial spatiotemporal sequence prediction model provided by an embodiment of the present invention. It belongs to AGC_Seq2Seq (Adaptive Graph Convolutional Sequence-to-Sequence, attention graph convolutional sequence-to-sequence model). This model includes a hybrid graph convolutional layer, a typhoon exogenous variable fusion mechanism, and a GRU encoding / decoding module based on an attention mechanism.

[0115] Specifically, addressing the complexity of the power grid topology, a hybrid graph convolutional layer receives multiple adjacency matrices while simultaneously initializing the node embedding vectors of two science departments. This matrix is ​​then dynamically generated using their inner product and ReLU and Softmax activation functions. The hybrid graph convolutional layer performs a weighted sum of the graph convolution outputs of each matrix, thereby simultaneously capturing both the explicit physical constraints and implicit spatiotemporal dependencies of the power grid. The module processing procedure is as follows:

[0116] Assume the input nighttime light characteristics at time t are as follows: Hybrid graph convolutional layers utilize learnable node embedding vectors. and Dynamically generate adaptive adjacency matrix:

[0117] ;

[0118] Combine it with K fixed-dimensional matrices Perform joint graph convolution to output the spatial encoding features at time t:

[0119] .

[0120] To accurately characterize the impact of external disasters, the typhoon exogenous variable fusion mechanism is implemented by combining an exogenous feature embedding layer and a linear mapping layer. After linearly mapping the typhoon exogenous variable sequence composed of wind force level, typhoon distance, etc., it is spliced ​​with the spatial joint features output by the hybrid graph convolutional layer in the feature dimension to form a disaster-power grid joint representation, which is then input to the GRU encoder.

[0121] The exogenous shock variable of the typhoon at time t After linear mapping, it is concatenated with spatial features to serve as a joint representation of disaster and power grid. :

[0122] ;

[0123] The encoder uses a gated recirculation unit (GRU), which internally updates the gate. and reset door Update the hidden state at time t The formula is as follows:

[0124] ;

[0125] ;

[0126] ;

[0127] ;

[0128] Output the full historical hidden state during the encoding phase. and final state .

[0129] The encoder uses a gated recurrent unit network to extract historical temporal memory from the fused joint features; the decoder introduces a temporal attention mechanism, which calculates the attention weights of the decoder's hidden state and the encoder's output at each historical moment, dynamically aligns and extracts the context vector most relevant to the future state, and then combines the exogenous typhoon prediction features of the future period to recursively output the nighttime light prediction values ​​of each node at multiple time steps in the future.

[0130] Among them, in predicting the future During each step, a dot product attention mechanism is introduced to calculate the current state of the decoder. Alignment weights with historical states Generate dynamic context vectors :

[0131] ;

[0132] ;

[0133] ;

[0134] Finally, With the future Known exogenous features The concatenation is input to the decoder GRU unit and outputs a baseline prediction value. :

[0135] ;

[0136] ;

[0137] ;

[0138] S13. Extracting target adjacency matrix combinations from a four-dimensional matrix feature pool based on matrix ablation experiments;

[0139] During the model training phase, since different regions have different power grid topologies and different geographical environments, and different typhoon disasters will have different impacts, the impact of the adjacency matrix combination method in the four-dimensional matrix feature pool on the node NTL time series prediction accuracy can be tested one by one through matrix ablation experiments, so as to select the target adjacency matrix combination for a specific typhoon in the target area power grid.

[0140] The results of the matrix ablation experiment are shown in Table 1 below:

[0141] Table 1

[0142]

[0143] Where geo is the geographic adjacency matrix; func is the functional adjacency matrix; sem is the semantic similarity adjacency matrix; and adp is the adaptive adjacency matrix.

[0144] In this embodiment, the hyperparameter architecture of the initial spatiotemporal sequence prediction model can be fixed. The subsets of adjacency matrices to be evaluated, as shown in Table 1 above, are sequentially fed into the model as graph convolution input matrices. Using continuous time series of pre-disaster nighttime light data, supervised sample pairs are constructed according to historical and future prediction windows. The mean squared error of the multi-step prediction output is used as the loss function. The model weights are iteratively updated by backpropagating gradients along the time axis until each configuration converges independently. The mean absolute error and coefficient of determination of all matrix combinations in the independent validation period are recorded and compared. The combination with the smallest error and the highest coefficient of determination is determined as the optimal structure scheme for the target region, i.e., the target adjacency matrix combination.

[0145] S14. The initial spatiotemporal sequence prediction model is trained using the target adjacency matrix combination, pre-disaster nighttime light remote sensing data, and post-disaster nighttime light remote sensing data to generate the target spatiotemporal sequence prediction model.

[0146] In this embodiment, the target adjacency matrix combination is used as the spatial feature input basis of the model, and the pre-disaster nighttime light remote sensing data is used as the temporal input subject and training ground truth label of the model. Simultaneously, the typhoon-related exogenous variables of the corresponding time period are embedded as auxiliary features into the model and input into the initial AGC_Seq2Seq model. The pre-disaster nighttime light time series data is divided into training set, validation set and test set according to a preset ratio. The backpropagation algorithm is used to iteratively update the learnable parameters of the model. The fitting accuracy and generalization ability of the model are monitored in real time through the validation set. At the same time, the prediction optimality of the target adjacency matrix combination is verified again through ablation experiments. When the prediction error of the model on the test set drops to within the preset threshold and the evolution trend of the time series prediction curve is highly consistent with the actual observation data, the training is stopped and all parameters of the model are solidified, generating a target spatiotemporal sequence prediction model that can stably output the ideal stable prediction evolution sequence of each power grid node in the target area under the absence of disaster impact. By combining multi-dimensional topological constraints of target adjacency matrices and learning the temporal patterns of pre-disaster stable period data, the model can simultaneously capture the explicit physical topological associations and implicit temporal evolution dependencies of the power grid, completely eliminating the interference of disaster disturbances on the model's pattern learning. The output disaster-free baseline prediction evolution sequence can accurately reflect the normal operating patterns of power grid nodes.

[0147] For a detailed implementation, please refer to Figure 3 The model employs a 2×2 four-subgraph layout, fully covering extreme boundary node testing and global average effect verification. The horizontal axis represents a time step of 0-160 days, fully encompassing the three independent stages of model training, verification, and testing. The vertical axis represents NTL data characterizing the power supply status and load continuity of power grid nodes. The gray solid line represents the actual nighttime light observations acquired by satellite remote sensing, the red dashed line represents the time-series prediction results output by the model, and the green and orange dashed lines are the interval division lines for training / verification and verification / testing, respectively. MAE is the core accuracy evaluation index; the smaller the value, the smaller the deviation between the model prediction and the actual value, and the higher the accuracy. To verify the model's generalization ability and reliability, this embodiment not only extracts the global average prediction results but also conducts rigorous extreme boundary tests, selecting the three extreme nodes with the largest global prediction mean absolute error. Figure 3 A full-time comparison was performed on (a), (b), and (c) in the simulation. The simulation results show that:

[0148] Rigorous testing was conducted on the three extreme boundary nodes with the largest prediction errors out of 9123 nodes in the entire network, corresponding to the three subgraphs (a), (b), and (c) in the figure. Even under extreme error scenarios, the predicted curve output by the model was still highly synchronized with the peaks, troughs, and overall oscillation time series pattern of the actual observed curve, with perfect phase matching. There was only a small deviation in the absolute value of the peak value, with no trend deviation. Figure 3As shown in (d), on a global average level, the baseline prediction path output by the model closely matches the temporal evolution of the actual observed data. More importantly, even at the extreme nodes with the largest network error, the peaks, troughs, and overall oscillation patterns of the prediction curve (red line) remain highly phase-synchronized with the natural fluctuation trajectory of the actual surface illuminance (gray line). This fully demonstrates that the AGC_Seq2Seq model constructed in this invention has not fallen into superficial numerical fitting, but has truly learned and captured the deep temporal evolution mechanism of nighttime lighting at power grid nodes under complex environmental constraints. This result objectively proves the model's extremely strong robustness across different topological nodes, providing an extremely stable and reliable disaster-free comparison benchmark for the dynamic determination of the actual damage state of the power grid under disaster impacts using a dual-threshold method.

[0149] Step 103: Based on the target spatiotemporal sequence prediction model, perform cascade failure simulation scanning on each power grid node in the target area, and extract the evolution equilibrium step number and final light loss corresponding to each power grid node according to the cascade failure simulation results.

[0150] Cascaded failure simulation scanning refers to using the N-1 traversal principle to assume that a single node in the power grid suffers a devastating blow (the node characteristics are forcibly set to a fault state), and using this as the initial boundary condition to drive the target spatiotemporal sequence prediction model to perform multi-step rolling deduction, and to examine the dynamic process of the fault propagating to other nodes in the entire network along the multidimensional adjacency matrix.

[0151] Evolutionary equilibrium steps refer to the total number of discrete time steps from the initial fault injection time until the system no longer generates new failure nodes and the power grid reaches a new steady state after multiple rounds of dynamic propagation.

[0152] Final light loss refers to the spatial sum of the nighttime light (NTL) attenuation of each affected grid node relative to the baseline predicted evolution sequence when the cascading failure time-series simulation reaches an evolutionary equilibrium state. This value is used to characterize the final spatial damage depth caused when the fault propagation reaches its limit and the network's resilience.

[0153] In this embodiment, each power grid node within the target area is traversed and sequentially used as a scanning object. The associated nighttime light remote sensing data, i.e., NTL features, is set to zero and input into the target spatiotemporal sequence prediction model. The propagation process of the fault along the multidimensional topology is dynamically simulated. The failure of secondary nodes is determined by dual thresholds. The simulation is iterated until the entire network reaches an evolutionary equilibrium state with no newly failed nodes. After completing the single-node simulation, the evolutionary equilibrium step number and final light loss corresponding to the target node are extracted. The entire network is traversed to complete the full scan and index extraction. With a propagation-filtering-update feedback closed-loop mechanism, the simulation of the cascading propagation of a single fault source under multidimensional constraints is effectively covered.

[0154] In one example of the present invention, step 103 may include the following sub-steps S21-S27:

[0155] S21. Using the target spatiotemporal sequence prediction model, rolling time series extrapolation is performed based on the initial nighttime light remote sensing data without interference to generate a baseline prediction evolution sequence.

[0156] In this embodiment, after obtaining the target spatiotemporal sequence prediction model, historical data before the disaster and exogenous variables of future typhoon impact can be input. Without injecting any assumptions of human node failure, the model deduces the predicted state sequence of each node at multiple subsequent time steps, thus obtaining the baseline predicted evolution sequence.

[0157] S22. Traverse the power grid nodes within the target area and treat each power grid node as an attack target node one by one; in each time step of the rolling simulation, continuously force the nighttime light remote sensing data corresponding to the attack target node to zero to generate forced input data.

[0158] The target node refers to the single power grid node selected as the initial disturbance source of the fault in the N-1 cascaded failure simulation.

[0159] In this embodiment, a unified time-series simulation window for the whole-network cascade simulation is first defined. Each independent power grid node in the target area that has completed grid division is taken as the traversal object. According to the preset traversal order, each individual power grid node is selected as the target node for this simulation. In the simulation window, the standardized nighttime light remote sensing feature value corresponding to the target node is forcibly set to zero, replacing the normal operation light data of the original node, generating forced input data injected with the initial fault of a single point, which is used as the input for this single-node cascade simulation.

[0160] S23. Using the target spatiotemporal sequence prediction model, a single-step forward extrapolation is performed based on the forced input data to generate the propagation prediction evolution data at the current time step, and the deviation between the propagation prediction evolution data and the benchmark prediction evolution sequence at the corresponding time step is calculated.

[0161] In this embodiment, the target spatiotemporal sequence prediction model with fixed parameters is imported, and rolling time-series recursive deduction is performed in units of preset time steps. After the node state prediction of each time step is completed, the prediction result of that time step is updated in real time to the input data of the next time step, and the recursive deduction continues until the preset maximum deduction window boundary, finally generating a propagating prediction evolution sequence covering all grid nodes and all deduction time steps.

[0162] For each time step, the difference between corresponding elements of the baseline predicted evolution sequence and the propagation predicted evolution sequence is calculated node by node, forming a deviation tensor of the same dimension as the spatiotemporal grid. Since the baseline sequence has internalized the normal fluctuation patterns such as holiday power consumption patterns and periodic maintenance through model learning, this differential operation can completely separate these common system fluctuations from cascading damage signals, so that the deviation only retains the newly added power outage loss component caused purely by the initial cause failure of the target node propagating through the multidimensional topology, providing a net physical failure signal that is not affected by background noise for the dual threshold determination.

[0163] S24. Determine whether the deviation of the other power grid nodes, excluding the target node, is greater than the preset dual filtering threshold. The dual filtering threshold includes the historical baseline fluctuation threshold and the model inherent error threshold.

[0164] The historical baseline threshold and the model inherent error threshold serve as the dual threshold determination criteria in this embodiment. The historical baseline threshold is calculated by statistically analyzing the natural fluctuation patterns of nighttime light data for each power grid node in the period leading up to the disaster, calculating twice the standard deviation of its historical fluctuations, and selecting values ​​within the 95% confidence interval. The model inherent error threshold is obtained by extracting the average absolute error of the aforementioned target spatiotemporal sequence prediction model during non-disaster periods. This dual threshold setting effectively filters out normal natural temporal fluctuations and model prediction noise interference.

[0165] S25. If there are other power grid nodes with deviations greater than the double filtering threshold, they are identified as newly added secondary failure nodes. Extract the propagation prediction evolution data of the newly added secondary failure nodes at the current time step, and inject it as the state update value into the forced input data of the next time step. Jump to execute the step of using the target spatiotemporal sequence prediction model, performing single-step forward extrapolation based on the forced input data, and generating the propagation prediction evolution data of the current time step.

[0166] In this embodiment, a failed node refers to a power grid node whose deviation exceeds two independent judgment thresholds simultaneously during the cascading propagation process. The model identifies it as suffering physical damage that cannot be absorbed by the system redundancy, updates its nighttime light (NTL) characteristics to the actual attenuation prediction value obtained in the current simulation, and uses it as the state input for the next time step of the cascading simulation.

[0167] In this embodiment, when the deviation exceeds the node's own historical baseline threshold and the global model's inherent error threshold at a certain time step, it is confirmed that the secondary node's light decay is not caused by normal fluctuations or model errors, but by a real physical cascading impact caused by a preceding fault propagating through multidimensional topology. The node is immediately marked as a failed node, its Nighttime Light (NTL) characteristics are updated to the currently derived actual attenuation prediction value, and this is used as the state input for the next time step of the cascading derivation. The rolling derivation in step S22 continues, repeating until, after several time steps, the deviation in the new prediction no longer triggers the dual-threshold judgment, and the entire network stops generating new failed nodes. This dual-threshold AND logic simultaneously avoids both false alarms from conservative models and false negatives from aggressive models, driving the cascading evolution path to naturally converge to a stable state while balancing physical rationality and numerical completeness.

[0168] S26. If there are no other power grid nodes with deviations greater than the double filtering threshold in the current time step, it is determined that the evolution of the current cascaded failure branch has reached a balanced state, and the rolling timing deduction of the current attack target node is stopped.

[0169] In this embodiment, if there are no other grid nodes with deviations greater than the dual-filter threshold at a certain time step, it indicates that the light drop at the target node is still within the range that can be explained by daily natural fluctuations or model prediction residuals. The system absorbs the impact of the target node failure through grid redundancy and load transfer, and the cascaded energy fails to form an effective destructive cross-domain propagation in this direction. At this time, the simulation of the cascaded branch is actively terminated, computing resources are released, the simulation is rolled back to the new target node, and the next round of N-1 scanning is executed.

[0170] S27. Based on the total number of simulation cycles taken to reach the equilibrium state, and the final value of the network-wide nighttime light attenuation caused by all failed nodes, extract the evolutionary equilibrium steps and the final light loss corresponding to the attack target node.

[0171] Furthermore, S27 may include the following sub-steps:

[0172] The cumulative time from the start of the rolling time series simulation until no new failure nodes appear is obtained as the evolution balance step number for each power grid node.

[0173] The total nighttime light loss of each power grid node when the evolutionary equilibrium step number is reached is obtained as the corresponding final light loss.

[0174] In this embodiment, for each power grid node serving as the initial target node, after its corresponding cascade simulation reaches an evolutionary equilibrium state, the total number of simulation time steps experienced by the target node from the initial fault injection to the network reaching equilibrium is counted and extracted as the evolutionary equilibrium step number corresponding to that node. This index precisely characterizes the spatial propagation trajectory of the fault on the multidimensional heterogeneous topology, representing the overall vulnerability of the system to continuous cascading impacts, as well as the cumulative destructive effect caused by the energy transfer of typhoon-induced disasters.

[0175] Simultaneously, the spatial sum of the nighttime light attenuation deviations of affected nodes across the entire network that meet the judgment threshold when the secondary cascade simulation evolves to equilibrium is extracted, serving as the final light loss for that node. This indicator, combined with the number of evolutionary equilibrium steps, characterizes the dynamic buffer depth of the system in absorbing external disaster impacts and preventing the cross-regional spread of faults from two dimensions: spatial damage limit and temporal blocking period. After completing the simulation scan of all nodes across the entire network as fault sources, a dataset of two resilience identification indicators covering all nodes in the power grid—the number of evolutionary equilibrium steps and the final light loss—is generated.

[0176] Step 104: Perform phase space clustering based on the number of evolutionary equilibrium steps and the final light loss to determine multiple cluster centers, and assign feature labels to the feature clusters to which each cluster center belongs based on the quadrant distribution characteristics of each cluster center.

[0177] Phase space refers to a two-dimensional feature space constructed with the normalized evolution equilibrium step count as the horizontal axis and the normalized final light loss as the vertical axis. It is used to uniformly represent the resilience characteristics of power grid nodes from both temporal and spatial dimensions.

[0178] Feature labels refer to the engineering roles assigned based on the specific location of the cluster center in the four quadrants of phase space, including "Global Catastrophe Spread Hub", "Transient Collapse Trigger Point", "Long-Term Hysteresis Isolation Zone" and "Resilient Absorption Sink".

[0179] In this embodiment, the evolutionary equilibrium steps and final lighting loss extracted from all power grid nodes are subjected to maximum-minimum normalization to eliminate the dimensional differences between the two indicators. A two-dimensional spatiotemporal resilience evolution phase space is constructed with the normalized evolutionary equilibrium steps as the horizontal axis and the final lighting loss as the vertical axis, mapping all power grid nodes as feature points within the phase space. An unsupervised clustering algorithm is used to cluster the feature points within the phase space, adaptively converging to obtain multiple highly cohesive and loosely coupled feature clusters, and determining the coordinates of the cluster center corresponding to each feature cluster. Based on the four-quadrant distribution characteristics of each cluster center in the two-dimensional coordinate system of the phase space, a corresponding power grid physical role feature label is matched to each feature cluster, completing the resilience functional community division of all network nodes, thereby transforming the purely numerical features of discrete nodes into functional role classifications with clear physical meaning.

[0180] In one example of the present invention, step 104 may include the following sub-steps:

[0181] After normalizing each evolutionary equilibrium step and each final light loss, a two-dimensional spatiotemporal resilience evolution phase space is constructed with the evolutionary equilibrium step as the horizontal axis and the final light loss as the vertical axis.

[0182] Based on the Euclidean distance between each pair of feature points in the two-dimensional spatiotemporal resilience evolution phase space, spatial clustering is performed on each feature point to obtain multiple feature clusters;

[0183] Extract the cluster centers of each feature cluster;

[0184] Based on the four-quadrant distribution characteristics of each cluster center in the two-dimensional spatiotemporal resilience evolution phase space, feature labels are assigned to the feature clusters to which each cluster center belongs.

[0185] In this embodiment, the evolution balance steps and final light loss of all target nodes are subjected to max-min normalization to eliminate dimensional differences. The normalized evolution balance steps are mapped to the horizontal axis, and the main compensation cumulative loss is mapped to the vertical axis to construct a two-dimensional spatiotemporal resilience evolution phase space. Each grid node is mapped to a corresponding phase space feature point in the phase space.

[0186] Subsequently, the Euclidean distances between all pairwise feature points in the phase space are calculated. Unsupervised clustering algorithms based on centroid iteration, such as K-means, are used to spatially cluster the scattered points. Multiple feature clusters are automatically converged and divided by minimizing the intra-cluster squared error function. The arithmetic mean coordinates of the members within each cluster are extracted as the cluster centers. Based on the four-quadrant distribution characteristics of each cluster center in the orthogonal coordinate system of the two-dimensional phase space, a corresponding physical role feature label with clear significance for power grid disaster prevention engineering is matched to the feature cluster to which each cluster center belongs, completing the adaptive division of the resilience functional groups of the entire power grid nodes. Through the unified phase space representation of dual-dimensional resilience indicators, the scattered node-level quantitative data is integrated into a system-level resilience feature space. Combined with unsupervised clustering, objective adaptive grouping of massive nodes is achieved. Finally, the transformation from mathematical features to engineering disaster prevention roles is completed through four-quadrant distribution mapping.

[0187] Specifically, assigning feature labels based on the four-quadrant distribution characteristics can be achieved in the following ways:

[0188] If both the horizontal and vertical values ​​of the central coordinates approach extreme values, the community is identified as a highly destructive "global catastrophe spread hub" that is prone to triggering a global avalanche; if the cumulative loss is high and the number of evolution steps is short, it is identified as a "transient collapse trigger point" that is prone to triggering a shock-like power outage; if the cumulative loss is small and the number of evolution steps is long, it is identified as a "long-term hysteresis isolation zone" where the system oscillates repeatedly; if both the horizontal and vertical values ​​are in the low value range, it is identified as a "resilient absorption sink" with extremely strong fault isolation capabilities.

[0189] Step 105: Based on the feature labels and cascading failure simulation results, extract the cascading risk propagation region and output the power grid resilience risk zoning map corresponding to the target region.

[0190] The cascaded risk spread area refers to the strip-shaped high vulnerability area jointly delineated in real geographical space by all high-frequency topological connections that are judged to be failure spread in the cascaded failure simulation of all power grid nodes.

[0191] A power grid resilience risk zoning map is a comprehensive spatial visualization output map that uses a geographic base map as a foundation, overlays continuous risk surfaces using color gradients or contour lines, and symbolically classifies and identifies nodes in different physical functional groups.

[0192] In this embodiment, after obtaining the feature labels, target clusters with high disaster risk are selected based on the feature labels. Combining the high-frequency fault propagation topology edges traced during the cascading failure simulation, the target clusters and high-frequency fault edges are mapped to the real geographic coordinate system of the target area. The kernel density estimation algorithm is used to perform spatial density calculation on the high-frequency fault edges, transforming the discrete vector fault paths into continuous spatial risk probability surfaces, and extracting continuous cascading risk spread areas. Thus, the kernel density estimation realizes the transformation from discrete fault paths to continuous risk areas, breaking through the limitations of traditional discrete fault point representation and accurately extracting the core spread channel of typhoon disaster energy.

[0193] By integrating the distribution of various clusters and cascading risk spread areas with the geographic base map of the target area, layer overlay and symbolic rendering are completed, and the corresponding power grid resilience risk zoning map of the target area is output.

[0194] In one example of the present invention, step 105 may include the following sub-steps:

[0195] Locate the target cluster from the feature clusters based on each feature label;

[0196] Based on the simulation results of target clustering and cascade failure, the high-frequency failure topological edges corresponding to each feature point within the target cluster are determined.

[0197] Each feature point and each high-frequency failure topological edge within the target cluster is mapped to a geographic coordinate system, and the kernel density value corresponding to each high-frequency failure topological edge is calculated by combining the kernel density estimation algorithm to generate a cascaded failure risk density raster surface.

[0198] Select a strip-shaped area that meets the preset risk conditions from the cascade failure risk density grid surface as the cascade risk propagation corridor;

[0199] By employing cascaded risk contagion corridors, all feature clusters, and feature labels, a power grid resilience risk zoning map corresponding to the target area is generated and output.

[0200] In this embodiment, feature clusters labeled "global catastrophe propagation hub" and "transient collapse trigger point" are selected as target clusters, and the power grid nodes and their identifiers are extracted. Then, the entire N-1 cascaded failure simulation results are traversed, and the cumulative number of times each topology edge is marked as an effective propagation path is counted. After sorting by frequency in descending order, the top-ranked high-frequency edge set is extracted, and edges with at least one end belonging to a member of the target cluster are identified as high-frequency failure topology edges. Specifically, the upper quartile of the total trigger frequency of all topology edges in the entire network can be used as the high-frequency determination threshold; topology edges with a total trigger frequency not lower than this threshold are identified as high-frequency failure topology edges.

[0201] The vector datasets of node coordinates within the target clusters and high-frequency failure topological edges are uniformly transformed to a spatial coordinate system completely consistent with the preprocessed nighttime light data. Using edge segments as input elements for the kernel density estimation algorithm, each high-frequency failure topological edge is first discretized into uniform sampling points at intervals no greater than half the target raster resolution. Kernel density calculation is then performed based on these sampling points. The cumulative trigger frequency of the high-frequency failure topological edges is used as a weight field input to the calculation model. The kernel function bandwidth is adaptively determined using the Silverman empirical rule, or a fixed bandwidth is set to match the average span of the power grid in the study area. The resolution and range of the output raster are completely consistent with the preprocessed nighttime light remote sensing data. The initial output density raster is then normalized to eliminate dimensional differences, ultimately generating a normalized cascaded failure risk density raster surface.

[0202] Based on this, the grid surface is reclassified by quantile, and continuous pixels with density values ​​higher than the preset quantile threshold are merged into polygon objects. Morphological closing operations are used to fill internal holes and remove fragments and patches with too small an area, while retaining the main connected strip area as a cascaded risk spread corridor.

[0203] Finally, the corridor polygon is filled with a semi-transparent gradient thermal color. All feature cluster nodes are configured with different shapes and color symbols according to four types of labels. This is then overlaid and refined with the base map of the region boundary, road network, water system, etc., to output the power grid resilience risk zoning map corresponding to the target region. In the specific implementation, the layer system can be built in the following hierarchical order: geographic base layer, power grid base layer, risk zoning layer, and annotation element layer. The geographic base layer uses the officially published topographic map or satellite image base map of the target region. The power grid base layer overlays the basic vector data of substations, high-voltage transmission lines, and power grid grid units in the target region. The risk zoning layer overlays the full feature cluster node hierarchical rendering layer and the cascaded risk spread corridor layer from bottom to top. The node layer uses differentiated color rendering according to feature labels: global disaster spread hubs are red, transient collapse trigger points are orange, long-term hysteresis isolation zones are yellow, and strong resilience absorption sinks are green. The cascaded risk spread corridor uses red gradient thermal rendering to distinguish between extremely high and high risk levels. The annotation element layer covers the core high-risk node names, risk corridor coverage, regional boundaries, and key geographic markers. This provides a highly intuitive and targeted decision-making guide for flood control and typhoon relief departments in developing adaptive island delineation plans before typhoon landfall and in accurately allocating resources after disasters.

[0204] Please see Figure 4 In one example of this invention, an N-1 time-cascaded simulation scan is performed on each node in the entire network, and the spread effect of the top ten target nodes with the highest destructive force (total cumulative loss) is spatially visualized. To verify the reliability of the cascaded simulation results in the real physical world, this study uses Open Street Map street block and infrastructure data to spatially map and verify the high-risk nodes identified in phase space with the four major power plants in region L. Figure 5 As shown, the comparison results demonstrate a high degree of agreement with the model's deduction:

[0205] First, among the 9123 nodes in the entire network, the "Global Disaster Spread Hub" (node ​​188), which has the greatest impact on the system and ranks first, precisely corresponds in real geographical space to power plant a, which has the largest installed capacity and the most core position in the system in the southern part of region L. Second, the nodes where power plant b (node ​​615, ranked 218 / 9123) is located in the south and power plant c (node ​​8835, ranked 175 / 9123) is located in the north are also both within the top 2.5% of the total network's cumulative losses.

[0206] This quantitative rating result profoundly reveals the extremely vulnerable physical topology of the power grid in the study area. Historical disaster operation records show that the power grids in Region L suffer from a severe structural defect of "south-to-north power transmission": over 70% of the island's power generation is concentrated in large power plants such as power plant a and power plant b in the south, while the core power load is concentrated in the area where power plant d is located in the north. During the hurricane of xx year and several other major disasters, it was precisely because the main power sources, such as power plant a, and their cross-mountain transmission lines were damaged first that the lifeline of energy transmission for the entire grid was severed, ultimately triggering a cascading blackout that affected the entire island and lasted for several months.

[0207] In summary, the multidimensional heterogeneous graph sequence model constructed in this embodiment, without inputting any prior labels for the power plant, relies purely on deep learning of the spatiotemporal evolution trajectory of nighttime lighting and multiple network topologies to extremely accurately capture hidden structural vulnerabilities, precisely pinpointing the true power supply heart of the system as the highest-risk node from a disaster prevention perspective. This confirms the extremely high reliability and practical engineering value of this dynamic cascaded assessment mechanism in macro-level disaster prevention zoning and blind location of high-risk hubs.

[0208] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0209] The following describes the regional power grid resilience identification device based on nighttime light data provided in the embodiments of the present invention. The regional power grid resilience identification device based on nighttime light data described below and the regional power grid resilience identification method based on nighttime light data described above can be referred to in correspondence with each other.

[0210] Please see Figure 6 , Figure 6 An embodiment of the present invention provides a regional power grid resilience identification device based on nighttime light data, comprising:

[0211] The matrix creation unit 601 is used to acquire nighttime light remote sensing data and multi-source power grid data of the target area and create multiple adjacency matrices;

[0212] The model training unit 602 is used to train the initial spatiotemporal sequence prediction model using the full adjacency matrix and the adaptive adjacency matrix, and to generate the target spatiotemporal sequence prediction model.

[0213] The feature extraction unit 603 is used to perform cascade failure simulation scanning on each power grid node in the target area based on the target spatiotemporal sequence prediction model, and extract the evolution equilibrium step number and final light loss corresponding to each power grid node according to the cascade failure simulation results.

[0214] The feature label assignment unit 604 is used to perform phase space clustering based on each evolution equilibrium step and each final light loss, determine multiple cluster centers, and assign feature labels to the feature clusters to which each cluster center belongs based on the quadrant distribution characteristics of each cluster center.

[0215] The power grid resilience risk zoning map generation unit 605 is used to extract the cascading risk propagation area based on feature labels and cascading failure simulation results, and output the power grid resilience risk zoning map corresponding to the target area.

[0216] Optionally, the matrix creation unit 601 is specifically used for:

[0217] Acquire nighttime light remote sensing data and multi-source power grid data for the target area;

[0218] Data preprocessing is performed on nighttime light remote sensing data to obtain preprocessed remote sensing data;

[0219] Based on preprocessed remote sensing data and multi-source power grid data, a geographic adjacency matrix, a functional adjacency matrix, and a semantically similar adjacency matrix are constructed, respectively.

[0220] Optionally, the nighttime light remote sensing data includes pre-disaster nighttime light remote sensing data and post-disaster nighttime light remote sensing data; model training unit 602 includes:

[0221] The feature pool construction sub-unit is used to construct a four-dimensional matrix feature pool using the full adjacency matrix and the adaptive adjacency matrix;

[0222] The model creation sub-unit is used to build the initial spatiotemporal sequence prediction model;

[0223] Combining and selecting sub-units is used to extract target adjacency matrix combinations from a four-dimensional matrix feature pool based on matrix ablation experiments;

[0224] The model training subunit is used to train the initial spatiotemporal sequence prediction model using the target adjacency matrix combination, pre-disaster nighttime light remote sensing data, and post-disaster nighttime light remote sensing data, and generate the target spatiotemporal sequence prediction model.

[0225] Optionally, the model creation sub-unit is specifically used for:

[0226] Construct a hybrid graph convolutional layer;

[0227] Create and sequentially connect the exogenous feature embedding layer and the linear mapping layer;

[0228] Connect the hybrid graph convolutional layer and the linear mapping layer to the stitching layer;

[0229] By sequentially connecting the splicing layer, the attention-based GRU encoding / decoding module, and the fully connected layer, an initial spatiotemporal sequence prediction model is obtained.

[0230] Optionally, the feature extraction unit 603 includes:

[0231] The baseline sequence generation subunit is used to generate a baseline prediction evolution sequence by performing rolling time-series extrapolation based on the target spatiotemporal sequence prediction model and uninterrupted initial nighttime light remote sensing data.

[0232] The perturbation addition sub-unit is used to traverse the power grid nodes in the target area and treat each power grid node as an attack target node one by one; in each time step of the rolling simulation, the night light remote sensing data corresponding to the attack target node is continuously forced to zero to generate forced input data;

[0233] The deviation calculation subunit is used to use the target spatiotemporal sequence prediction model to perform single-step forward extrapolation based on the forced input data, generate the propagation prediction evolution data at the current time step, and calculate the deviation between the propagation prediction evolution data and the benchmark prediction evolution sequence at the corresponding time step.

[0234] The threshold judgment subunit is used to determine whether the deviation of the other power grid nodes, excluding the attack target node, is greater than the preset dual filtering threshold, which includes the historical baseline fluctuation threshold and the model inherent error threshold.

[0235] The cyclic update subunit is used to identify any remaining power grid nodes whose deviation exceeds the dual filtering threshold as newly added secondary failure nodes; extract the propagation prediction evolution data of the newly added secondary failure nodes at the current time step, and inject it as a state update value into the forced input data at the next time step; then jump to execute the step of using the target spatiotemporal sequence prediction model, performing single-step forward extrapolation based on the forced input data, and generating the propagation prediction evolution data for the current time step.

[0236] The failure loop subunit is used to determine that the evolution of the current cascaded failure branch has reached a balanced state and stop the rolling timing deduction of the current attack target node if there are no other power grid nodes with the deviation amount greater than the dual filtering threshold in the current time step.

[0237] The feature extraction subunit is used to extract the evolutionary equilibrium steps and final light loss corresponding to the attack target node based on the total number of evolution cycles experienced when reaching the equilibrium state and the final value of the nighttime light attenuation caused by all failed nodes.

[0238] Optionally, the feature extraction subunit is specifically used for:

[0239] The cumulative time from the start of the rolling time series simulation until no new failure nodes appear is obtained as the evolution balance step number for each power grid node.

[0240] The total nighttime light loss of each power grid node at the time of reaching the evolutionary equilibrium step is obtained as the corresponding final light loss.

[0241] Optionally, the feature label assignment unit 604 is specifically used for:

[0242] After normalizing each evolutionary equilibrium step and each final light loss, a two-dimensional spatiotemporal resilience evolution phase space is constructed with the evolutionary equilibrium step as the horizontal axis and the final light loss as the vertical axis.

[0243] Based on the Euclidean distance between each pair of feature points in the two-dimensional spatiotemporal resilience evolution phase space, spatial clustering is performed on each feature point to obtain multiple feature clusters;

[0244] Extract the cluster centers of each feature cluster;

[0245] Based on the four-quadrant distribution characteristics of each cluster center in the two-dimensional spatiotemporal resilience evolution phase space, feature labels are assigned to the feature clusters to which each cluster center belongs.

[0246] Optionally, the power grid resilience risk zoning map generation unit 605 is specifically used for:

[0247] Locate the target cluster from the feature clusters based on each feature label;

[0248] Based on the simulation results of target clustering and cascade failure, the high-frequency failure topological edges corresponding to each feature point within the target cluster are determined.

[0249] Each feature point and each high-frequency failure topological edge within the target cluster is mapped to a geographic coordinate system, and the kernel density value corresponding to each high-frequency failure topological edge is calculated by combining the kernel density estimation algorithm to generate a cascaded failure risk density raster surface.

[0250] Select a strip-shaped area that meets the preset risk conditions from the cascade failure risk density grid surface as the cascade risk propagation corridor;

[0251] By employing cascaded risk contagion corridors, all feature clusters, and feature labels, a power grid resilience risk zoning map corresponding to the target area is generated and output.

[0252] This invention provides an electronic device, including a memory and a processor. The memory stores a computer program, which, when executed by the processor, causes the processor to perform the steps of the regional power grid resilience identification method based on nighttime light data as described in any embodiment of this invention.

[0253] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the above-described device and module can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0254] In the several embodiments provided by this invention, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0255] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0256] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0257] Indicatively, such as Figure 7 As shown, Figure 7 This is a schematic diagram of the internal structure of an electronic device 700 provided in an embodiment of the present invention. The electronic device 700 can be provided as a server. (Refer to...) Figure 7 The electronic device 700 includes a processing component 702, which further includes one or more processors, and memory resources represented by memory 701 for storing instructions, such as application programs, that can be executed by the processing component 702. The application programs stored in memory 701 may include one or more modules, each corresponding to a set of instructions. Furthermore, the processing component 702 is configured to execute instructions to perform the text recognition method of any of the above embodiments.

[0258] The electronic device 700 may also include a power supply component 703 configured to perform power management of the electronic device 700, a wired or wireless network interface 704 configured to connect the electronic device 700 to a network, and an input / output interface 705. The electronic device 700 can operate based on storage.

[0259] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for regional grid resilience identification based on nightlight data, characterized in that, include: Acquire nighttime light remote sensing data and multi-source power grid data for the target area, and create multiple adjacency matrices; The nighttime light remote sensing data includes pre-disaster nighttime light remote sensing data and post-disaster nighttime light remote sensing data; The initial spatiotemporal sequence prediction model is trained using all the adjacency matrices and the adaptive adjacency matrix, and the target spatiotemporal sequence prediction model is generated. Based on the target spatiotemporal sequence prediction model, a cascade failure simulation scan is performed on each power grid node in the target area, and the evolution equilibrium step number and final light loss corresponding to each power grid node are extracted according to the cascade failure simulation results. Phase space clustering is performed based on the evolutionary equilibrium steps and the final light loss, multiple cluster centers are determined, and feature labels corresponding to the feature clusters to which each cluster center belongs are assigned based on the quadrant distribution characteristics of each cluster center. Based on the feature labels and the cascade failure simulation results, the cascade risk propagation region is extracted, and the power grid resilience risk zoning map corresponding to the target region is output. The step of performing cascaded failure simulation scanning on each power grid node in the target area based on the target spatiotemporal sequence prediction model, and extracting the evolutionary equilibrium step number and final light loss corresponding to each power grid node based on the cascaded failure simulation results, includes: Using the target spatiotemporal sequence prediction model, rolling time series extrapolation is performed based on the interference-free initial nighttime light remote sensing data to generate a baseline prediction evolution sequence; The power grid nodes within the target area are traversed, and each power grid node is taken as an attack target node. In each time step of the rolling simulation, the nighttime light remote sensing data corresponding to the attack target node is continuously forced to zero to generate forced input data. Using the target spatiotemporal sequence prediction model, a single-step forward extrapolation is performed based on the forced input data to generate the propagation prediction evolution data for the current time step, and the deviation between the propagation prediction evolution data and the baseline prediction evolution sequence at the corresponding time step is calculated. Determine whether the deviation of the other power grid nodes, excluding the target node, is greater than a preset dual filtering threshold, which includes a historical baseline fluctuation threshold and a model inherent error threshold. If there are other power grid nodes whose deviation exceeds the dual filtering threshold, they are identified as newly added secondary failure nodes; the propagation prediction evolution data of the newly added secondary failure nodes at the current time step is extracted and injected as a state update value into the forced input data at the next time step; the process jumps to execute the step of using the target spatiotemporal sequence prediction model, performing single-step forward extrapolation based on the forced input data, and generating the propagation prediction evolution data for the current time step; If there are no other power grid nodes with a deviation greater than the dual filtering threshold in the current time step, it is determined that the evolution of the current cascaded failure branch has reached a balanced state, and the rolling timing deduction of the current attack target node is stopped. Based on the total number of simulation cycles required to reach the equilibrium state, and the final value of the network-wide nighttime light attenuation caused by all failed nodes, the evolutionary equilibrium steps and final light loss corresponding to the attack target node are extracted.

2. The method of claim 1, wherein, The steps of acquiring nighttime light remote sensing data and multi-source power grid data of the target area, and creating multiple adjacency matrices, include: Acquire nighttime light remote sensing data and multi-source power grid data for the target area; The nighttime light remote sensing data is preprocessed to obtain preprocessed remote sensing data; Based on the preprocessed remote sensing data and the multi-source power grid data, a geographical adjacency matrix, a functional adjacency matrix, and a semantic similarity adjacency matrix are constructed, respectively.

3. The method of claim 1, wherein, The step of training the initial spatiotemporal sequence prediction model using all the adjacency matrices and the adaptive adjacency matrix to generate the target spatiotemporal sequence prediction model includes: A four-dimensional matrix feature pool is constructed using all the aforementioned adjacency matrices and the adaptive adjacency matrix; Construct an initial spatiotemporal sequence prediction model; Based on matrix ablation experiments, target adjacency matrix combinations are extracted from the feature pool of the four-dimensional matrix; The initial spatiotemporal sequence prediction model is trained using the target adjacency matrix combination, the pre-disaster nighttime light remote sensing data, and the post-disaster nighttime light remote sensing data to generate the target spatiotemporal sequence prediction model.

4. The method of claim 3, wherein, The steps for constructing the initial spatiotemporal sequence prediction model include: Construct a hybrid graph convolutional layer; Create and sequentially connect the exogenous feature embedding layer and the linear mapping layer; The hybrid graph convolutional layer and the linear mapping layer are connected to the splicing layer; The splicing layer, the attention-based GRU encoding / decoding module, and the fully connected layer are connected in sequence to obtain the initial spatiotemporal sequence prediction model.

5. The method of claim 1, wherein, The steps of extracting the evolutionary equilibrium steps and final light loss corresponding to each of the power grid nodes include: The cumulative time from the start of the rolling time series deduction until no new failure nodes appear is obtained as the evolution balance step number corresponding to each of the power grid nodes; The total nighttime light loss of each power grid node when the evolutionary equilibrium step number is reached is obtained as the corresponding final light loss.

6. The method of claim 1, wherein, The step of performing phase space clustering based on the evolutionary equilibrium steps and the final light loss to determine multiple cluster centers, and assigning feature labels to the feature clusters to which each cluster center belongs based on the four-quadrant distribution characteristics of each cluster center, includes: After normalizing the evolutionary equilibrium steps and the final light loss, a two-dimensional spatiotemporal resilience evolution phase space is constructed with the evolutionary equilibrium steps as the horizontal axis and the final light loss as the vertical axis. Based on the Euclidean distance between each pair of feature points in the two-dimensional spatiotemporal resilience evolution phase space, spatial clustering is performed on each feature point to obtain multiple feature clusters; Extract the cluster centers of each of the aforementioned feature clusters; Based on the four-quadrant distribution characteristics of each cluster center in the two-dimensional spatiotemporal resilience evolution phase space, feature labels corresponding to the feature clusters to which each cluster center belongs are assigned.

7. The method of claim 1, wherein, The step of extracting the cascading risk propagation region based on the feature labels and the cascading failure simulation results, and outputting the power grid resilience risk zoning map corresponding to the target region, includes: Locate the target cluster from the feature clusters based on each of the aforementioned feature labels; Based on the target cluster and the cascade failure simulation results, the high-frequency failure topological edges corresponding to each feature point within the target cluster are determined. Each feature point and each high-frequency failure topological edge within the target cluster are mapped to a geographic coordinate system, and the kernel density value corresponding to each high-frequency failure topological edge is calculated using a kernel density estimation algorithm to generate a cascaded failure risk density raster surface. A strip-shaped region that meets the preset risk conditions is selected from the cascaded failure risk density grid surface as a cascaded risk propagation corridor; Using the cascaded risk contagion corridor, all the feature clusters, and the feature labels, a power grid resilience risk zoning map corresponding to the target area is generated and output.

8. A regional power grid resilience identification device based on nighttime light data, characterized in that, include: The matrix creation unit is used to acquire nighttime light remote sensing data and multi-source power grid data of the target area and create multiple adjacency matrices; The nighttime light remote sensing data includes pre-disaster nighttime light remote sensing data and post-disaster nighttime light remote sensing data; The model training unit is used to train the initial spatiotemporal sequence prediction model using all the adjacency matrices and the adaptive adjacency matrix, and to generate the target spatiotemporal sequence prediction model. The feature extraction unit is used to perform cascade failure simulation scanning on each power grid node in the target area based on the target spatiotemporal sequence prediction model, and extract the evolution equilibrium step number and final light loss corresponding to each power grid node according to the cascade failure simulation results; The feature label assignment unit is used to perform phase space clustering based on the evolutionary equilibrium steps and the final light loss, determine multiple cluster centers, and assign feature labels to the feature clusters to which each cluster center belongs based on the quadrant distribution characteristics of each cluster center. The power grid resilience risk zoning map generation unit is used to extract the cascading risk propagation region based on the feature labels and the cascading failure simulation results, and output the power grid resilience risk zoning map corresponding to the target region. The feature extraction unit includes: The baseline sequence generation subunit is used to generate a baseline prediction evolution sequence by performing rolling time-series extrapolation based on the target spatiotemporal sequence prediction model and uninterrupted initial nighttime light remote sensing data. The perturbation addition sub-unit is used to traverse the power grid nodes in the target area and treat each power grid node as an attack target node one by one; in each time step of the rolling simulation, the night light remote sensing data corresponding to the attack target node is continuously forced to zero to generate forced input data; The deviation calculation subunit is used to use the target spatiotemporal sequence prediction model to perform single-step forward extrapolation based on the forced input data, generate the propagation prediction evolution data at the current time step, and calculate the deviation between the propagation prediction evolution data and the benchmark prediction evolution sequence at the corresponding time step. The threshold judgment subunit is used to determine whether the deviation of the other power grid nodes, excluding the attack target node, is greater than the preset dual filtering threshold, which includes the historical baseline fluctuation threshold and the model inherent error threshold. The cyclic update subunit is used to identify any remaining power grid nodes whose deviation exceeds the dual filtering threshold as newly added secondary failure nodes; extract the propagation prediction evolution data of the newly added secondary failure nodes at the current time step, and inject it as a state update value into the forced input data at the next time step; then jump to execute the step of using the target spatiotemporal sequence prediction model, performing single-step forward extrapolation based on the forced input data, and generating the propagation prediction evolution data for the current time step. The failure loop subunit is used to determine that the evolution of the current cascaded failure branch has reached a balanced state and stop the rolling timing deduction of the current attack target node if there are no other power grid nodes with the deviation amount greater than the dual filtering threshold in the current time step. The feature extraction subunit is used to extract the evolutionary equilibrium steps and final light loss corresponding to the attack target node based on the total number of evolution cycles experienced when the equilibrium state is reached and the final value of the nighttime light attenuation caused by all failed nodes.

9. An electronic device, characterized in that, The system includes a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the regional power grid resilience identification method based on nighttime light data as described in any one of claims 1-7.

Images