Power grid composite disaster situation linkage early warning method and system based on four-dimensional stereoscopic coordination

Abstract

The application discloses a power grid composite disaster situation linkage early warning method and system based on four-dimensional three-dimensional coordination, the method obtains the visual monitoring data and time series monitoring data of the power grid node, maps the encoder to the unified feature space and performs feature dichotomy decomposition, obtains the aligned features shared across modalities and the context features private to the modalities, adopts an unsupervised multi-constraint alignment mechanism to realize deep fusion of heterogeneous features, generates node fusion features, recursively deduces the extreme weather evolution state based on the compression-decompression architecture, dynamically constructs a node association graph based on feature similarity and aggregates node features similar to the risk mode by using a graph convolution network, generates a power grid state vector containing global spatial coupling information, fuses the deduced meteorological features and power grid state features, outputs four-dimensional disaster situation prediction results, and converts the four-dimensional disaster situation prediction results into natural language prompts to drive a large language model to generate a disposal scheme, and realizes accurate early warning and intelligent decision-making of the composite disaster situation.

Description

Technical Field

[0001] This invention relates to the field of power system safety control and disaster early warning technology. Specifically, it relates to a power grid composite disaster linkage early warning method and system based on four-dimensional three-dimensional collaborative system, which is applicable to accurate early warning, situation assessment and emergency response guidance for power grid disasters under extreme composite weather conditions, and ensures the safe and stable operation of the power system. Background Technology

[0002] Against the backdrop of global climate change, extreme weather events are not only becoming more frequent and numerous, but also exhibiting significant complex and coupled characteristics. Various preconditions and complex events, as well as persistent complex extreme weather, pose a severe threat to the power system. Extreme weather often triggers coupled power accidents. For example, preconditions such as torrential rains following droughts and continuous storms can significantly exacerbate the intensity of disasters such as floods and landslides. In recent years, many parts of my country have frequently experienced persistent cold and wet complex extreme events, namely, cold and wet weather processes where low temperatures and rain / snow occur simultaneously. The formation mechanisms of such complex events are complex and diverse, often accompanied by derivative disasters such as freezing rain, snow accumulation, and road icing. These events not only affect the normal operation of transportation, agriculture, and various infrastructures, but also easily trigger serious power accidents such as transmission line collapses and breaks, and substation equipment damage, leading to large-scale power outages, seriously endangering the safety of electricity use for people's livelihoods and disrupting the stable operation of the social economy.

[0003] Current traditional power emergency response models suffer from several significant drawbacks, making them ill-suited to the high-efficiency requirements of complex disaster prevention and control. First, the lack of a dedicated complex disaster early warning system hinders targeted monitoring and early warning of the coupled characteristics of complex disasters. This makes it difficult to promptly identify early potential hazards and accurately predict the impact of overlapping disaster trends on the power grid. Existing early warning systems primarily target single disaster types, failing to achieve multi-disaster coupling analysis and early prediction. Second, the lack of effective analysis of the development trends and coupled impacts of complex disasters leads to a lack of scientific and systematic assessment of the disaster situation, hindering rapid understanding of the disaster's impact on the power grid and reducing its safety and stability. Third, the inefficiency in generating decision-making solutions makes it difficult to formulate scientifically sound on-site response strategies, failing to provide timely and reliable guidance for complex disaster emergency response and severely restricting the overall effectiveness of power grid disaster prevention and control. Fourth, the fragmentation of multimodal data is a prominent issue. Various related heterogeneous data, such as meteorological, geological, and power equipment status data, are difficult to effectively integrate and finely decouple, failing to fully realize the core value of the data and hindering accurate assessment of complex disasters affecting the power grid.

[0004] Existing power grid disaster early warning technologies mostly employ single monitoring dimensions or simple data splicing methods, failing to adopt advanced technologies such as multimodal data alignment and decoupling. They cannot fully leverage the correlation mechanisms between meteorological big data and power grid operation data, making it difficult to achieve accurate early prediction and coupled analysis of complex disasters, and thus unable to meet the high-precision and high-efficiency requirements for power grid complex disaster prevention and control. Therefore, constructing a coordinated early warning method and system capable of efficient multimodal data fusion, accurate prediction of complex disasters, and intelligent generation of emergency plans has become an urgent technical problem to be solved in the field of power system security and prevention. Summary of the Invention

[0005] To address the aforementioned technical challenges, this invention provides a four-dimensional, integrated, collaborative early warning method and system for power grid composite disasters. By constructing an integrated four-dimensional monitoring system encompassing air, space, ground, and power, it enables early prediction and coupled analysis of composite disasters. Drawing upon the technical approach of decoupling and efficiently fusing multimodal data, it achieves full-process linkage between disaster early warning, situation assessment, and emergency response plan generation. The aim is to improve the accuracy, timeliness, and intelligence of power grid composite disaster early warning, thereby ensuring the safe and stable operation of the power system.

[0006] The technical solution adopted in this invention is: A method for early warning of complex power grid disasters based on four-dimensional three-dimensional collaborative mechanism includes the following steps: Multimodal monitoring data is acquired based on the distribution of power grid nodes, including visual monitoring data and time-series monitoring data. The visual monitoring data and the temporal monitoring data are mapped to a feature space of a unified dimension by a visual encoder and a temporal encoder, respectively, to obtain visual features and temporal features. The visual features and the temporal features are subjected to feature binary decomposition, and the features of each modality are decoupled into alignment features for capturing cross-modal shared semantics and context features for preserving modality-private information. The alignment features and the context features are constrained and optimized by an unsupervised multi-constraint alignment mechanism to obtain quality-enhanced fusion features; Based on the fusion features, a recursive time compression network is used to extract high-order mutation features, and a time decompression network is used to gradually restore the spatiotemporal resolution to deduce the extreme weather state sequence at future moments. Based on the fused features, the feature similarity between nodes is calculated to dynamically construct a node association graph. A graph convolutional network is used for feature propagation and aggregation to generate a power grid state vector containing global spatial coupling information. The extreme weather state sequence is fused with the power grid state vector, and a four-dimensional disaster prediction result containing the disaster location, time, category and level is output through the prediction head network; The four-dimensional disaster prediction results are converted into structured natural language prompts, which are then input into a pre-trained large language model to generate pre-disaster prevention and post-disaster recovery plans.

[0007] Furthermore, the step of performing feature binary decomposition on the visual features and the temporal features, decoupling the features of each modality into alignment features for capturing cross-modal shared semantics and contextual features for preserving modality-private information, specifically involves: The feature vector output by the encoder is divided into two parts along the feature dimension: the first part is used as the alignment feature and the second part is used as the context feature. Both the visual features and the temporal features are decomposed into alignment features and contextual features.

[0008] Furthermore, the step of optimizing the alignment features and the context features through an unsupervised multi-constraint alignment mechanism to obtain quality-enhanced fusion features specifically includes: Calculate the alignment feature consistency constraint loss, which is used to bring the alignment features of different modalities of the same sample closer together and push the alignment features of different samples further apart; The statistical dependency between aligned features and context features is calculated using the Hilbert-Schmidt independence criterion, and the statistical dependency between aligned features and context features is minimized by feature independence constraint loss. The context feature diversity constraint loss is calculated using Jensen-Shannon divergence to constrain the distribution differences of context features between different nodes and prevent feature collapse. The weighted sum of the alignment feature consistency constraint loss, feature independence constraint loss, and context feature diversity constraint loss is taken as the total loss of unsupervised modality alignment; The alignment features and context features of all nodes are stacked and merged in the channel dimension to construct a fused feature tensor. A multi-head attention mechanism is used for deep interaction, and the node fused features after layer normalization and flattening are output.

[0009] Furthermore, the formula for calculating the alignment feature consistency constraint loss is as follows:

[0010] In the formula, The loss value is the alignment feature consistency constraint; B is the number of samples in the training batch, and the summation index indicates that all samples in the batch are traversed; i is the index of the current anchor sample; m and n are the indices of different modal types, where m≠n indicates comparison between different modalities; k is the index of negative samples in the batch other than the anchor sample; log is the natural logarithm function; exp is the exponential function with the natural constant e as the base; sim is the cosine similarity function, used to calculate the similarity between two feature vectors. Let m be the alignment feature of the m-th mode of the i-th sample; Let be the alignment feature of the i-th sample and the n-th modality, and Forming positive sample pairs; The alignment feature of the k-th sample and the n-th modality, with Forming negative sample pairs; Temperature is a parameter used to control the smoothness of the similarity distribution; Among them, cosine similarity The calculation formula is:

[0011] In the formula, u and v are two feature vectors for which the similarity is to be calculated; It is the transpose of u; Let u be the L2 norm; Let v be the L2 norm.

[0012] Furthermore, the formula for calculating the feature independence constraint loss is as follows:

[0013] In the formula, B is the feature independence constraint loss value; B is the number of samples in the training batch; i is the current sample index; m is the modality type index; This indicates whether the currently processed features originate from a visual modality or a temporal modality; Let Hilbert-Schmidt independence criterion function be used. Let m be the alignment feature of the m-th mode of the i-th sample; The context features of the m-th modality for the i-th sample; in, The calculation is based on the Gaussian kernel function, and the specific formula is as follows:

[0014] In the formula, P and Q are the characteristic matrices of the two random variables whose independence is to be calculated; d is the dimension of the eigenvector; This is the matrix trace operation, which is the sum of the elements on the main diagonal of the matrix; The Gaussian kernel matrix is ​​calculated based on the characteristic matrix P; The Gaussian kernel matrix is ​​calculated based on the characteristic matrix Q. Centered matrix; is a d×d identity matrix; 1 is a column vector of length d consisting entirely of 1s; It is a row vector of length d consisting entirely of 1s.

[0015] Furthermore, the formula for calculating the context feature diversity constraint loss is as follows:

[0016] In the formula, is the context feature diversity constraint loss value; j and k are the indices of different samples, j≠k indicates that diversity constraints are applied between different samples; m is the modality type index; The context features of the m-th modality for the j-th sample; represents the contextual features of the m-th mode of the k-th sample; JSD is the Jensen-Shannon divergence function; Among them, JSD is based on The divergence is calculated using the following formula:

[0017] In the formula, p is the probability density function of the first feature distribution, assuming that the context features follow a Gaussian distribution; q is the probability density function of the second feature distribution; for The divergence function measures the difference between two probability distributions.

[0018] Furthermore, based on the fused features, a recursive time compression network is used to extract high-order abrupt change features, and a time decompression network is used to gradually restore the spatiotemporal resolution to extrapolate the extreme weather state sequence at future moments. Specifically: During the compression stage, the temporal dimension is gradually compressed through multiple recursive time compression operations to extract high-order mutation features. Each compression operation includes average pooling, linear transformation and layer normalization. During the decompression stage, the spatiotemporal resolution is gradually restored through multiple recursive time decompression operations. Each decompression operation includes linear transformation, layer normalization, and GELU activation processing. After multiple decompression operations, the output sequence of extreme weather predictions for future times is generated.

[0019] Furthermore, the step of dynamically constructing a node association graph by calculating the feature similarity between nodes based on the fused features, and using a graph convolutional network for feature propagation and aggregation to generate a power grid state vector containing global spatial coupling information, specifically involves: The cosine similarity between any two power grid nodes is calculated as the node association weight. The Softmax function is used to normalize all association weights to construct a dynamic adjacency matrix. Based on the dynamic adjacency matrix, a graph convolutional network is used for multi-layer feature propagation and aggregation. The features output by the graph convolutional network are then processed by a feedforward network to obtain a power grid state vector containing global spatial correlation information.

[0020] Furthermore, the process of fusing the extreme weather state sequence with the power grid state vector, and outputting a four-dimensional disaster prediction result containing the disaster location, time, category, and level through a prediction head network, specifically involves: The extreme weather state sequence is concatenated and fused with the power grid state vector to form a joint feature; The joint features are input into the prediction head network, which consists of convolutional layers and a feedforward neural network, and outputs a four-dimensional prediction result, which includes information on the location, time, category, and level of the disaster.

[0021] Another technical solution adopted by the present invention is: A four-dimensional, three-dimensional, collaborative power grid composite disaster linkage early warning system, employing the aforementioned method, comprises: The multimodal data acquisition module is used to acquire visual monitoring data and time-series monitoring data based on the distribution of power grid nodes; The feature encoding module is used to map multimodal monitoring data to a unified-dimensional feature space through a visual encoder and a temporal encoder to obtain visual features and temporal features; The feature decoupling and alignment module is used to perform feature binary decomposition on the visual features and the temporal features, decouple the features of each modality into alignment features and context features, and obtain fused features through constraint optimization by an unsupervised multi-constraint alignment mechanism. The extreme weather projection module is used to extract high-order mutation features based on the fused features using a recursive time compression network, and gradually restore the spatiotemporal resolution through a time decompression network to project the extreme weather state sequence at future moments. The spatial coupling analysis module is used to calculate the feature similarity between nodes based on the fused features to dynamically construct a node association graph, and to use a graph convolutional network to perform feature propagation and aggregation to generate a power grid state vector containing global spatial coupling information. The disaster prediction module is used to fuse the extreme weather state sequence with the power grid state vector, and output a four-dimensional disaster prediction result containing the disaster location, time, category and level through the prediction head network; The intelligent decision generation module is used to convert the four-dimensional disaster prediction results into structured natural language prompts, input them into a pre-trained large language model, and generate pre-disaster prevention and post-disaster recovery plans.

[0022] Compared with the prior art, the present invention has the following advantages: 1) This invention proposes an unsupervised heterogeneous feature decoupling and alignment mechanism based on multiple constraints: In response to the problems of scarce disaster labels and high heterogeneity of multi-source data in the power scenario, it innovatively achieves accurate decoupling of visual and temporal features under unsupervised conditions through triple constraint loss of consistency, independence and diversity; It not only solves the problem of spatial misalignment of heterogeneous features, but also ensures that the model can still extract robust disaster core semantics when large-scale labeled data is lacking, significantly improving the utilization rate of multimodal data; 2) This invention constructs a dynamic graph topology learning method based on feature similarity: In view of the problem that existing technologies rely on static power grid physical topology for analysis and are difficult to capture cross-regional meteorological coupling risks, this invention abandons static topology and uses feature similarity to dynamically construct a node association graph, so that the system can automatically identify and aggregate node information that is geographically distant but has similar meteorological risk patterns, accurately quantify the spatial coupling impact of compound disasters on the power grid, and fill the blind spots of traditional topology analysis. 3) This invention designs an extreme weather evolution inference model based on a compression-decompression architecture: In view of the complex and sudden nature of the evolution of complex extreme weather, a high-order mutation feature is distilled by a recursive time compressor and spatiotemporal details are reconstructed by a decompressor, which effectively balances the problem of information forgetting and noise interference in long sequence modeling and significantly improves the prediction accuracy of the spatiotemporal evolution trend of extreme weather. 4) This invention realizes an intelligent closed-loop process of monitoring-prediction-decision: It innovatively transforms quantitative numerical prediction results into natural language prompts, drives a large language model to generate highly readable and logically rigorous operation and maintenance decision-making schemes, breaks the limitation of traditional early warning systems that only output cold data, and directly provides dispatchers with practical guidance for pre-disaster prevention and post-disaster recovery, greatly improving the power grid's response speed and decision-making efficiency in dealing with complex disasters. Attached Figure Description

[0023] Figure 1 This is a flowchart of a method according to an embodiment of the present invention.

[0024] Figure 2 This is a diagram showing the overall framework of the disaster early warning system in this invention. Detailed Implementation

[0025] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0026] This invention proposes a method and system for early warning of complex power grid disasters based on four-dimensional three-dimensional collaborative mechanisms. Its overall logical flow is as follows: Figure 1 As shown, the overall system framework is as follows: Figure 2As shown. Four-dimensional synergy refers to the collaborative integration of data from four dimensions: airborne remote sensing monitoring, space-based satellite monitoring, ground-based sensor monitoring, and power grid self-operation monitoring, forming an integrated three-dimensional monitoring system encompassing air, space, ground, and power.

[0027] like Figure 1 As shown, a power grid composite disaster linkage early warning method based on four-dimensional three-dimensional collaborative mechanism includes the following steps: Step S1: Multimodal data acquisition and feature encoding The system first acquires multi-source monitoring data based on the distribution of power grid nodes; these nodes include the locations of power infrastructure such as substations and power poles; for any power grid node i, the system acquires visual monitoring data. With time series monitoring data Visual monitoring data This includes image-based data such as remote sensing cloud images and drone inspection images; and time-series monitoring data. This includes sequence data such as meteorological sensor readings and power operation data.

[0028] To address the issue of inconsistent feature spaces in heterogeneous data, the system employs a dedicated encoder to map the data to a unified feature space. Visual features are extracted using a visual encoder comprised of multi-layer convolutional networks and stacked layer normalization layers; temporal features are extracted using a temporal encoder based on a long short-term memory network and linear projection layers. This step, by processing heterogeneous data separately using dedicated modal encoders, fully extracts the unique information of each modality while mapping different modalities to a unified feature space, laying the foundation for dimensional alignment in subsequent multimodal fusion.

[0029] The mathematical expression for feature encoding is as follows:

[0030]

[0031] In the formula, Let be the visual feature vector of the i-th power grid node, with a dimension of 2d; For visual encoder functions; This represents the visual monitoring data for the i-th power grid node; Let be the time-series feature vector of the i-th power grid node; Represents a 2d-dimensional real vector space; This is a timing encoder function; d represents the time-series monitoring data of the i-th power grid node; d is the dimension of a single feature subspace; 2d is the dimension of the complete feature vector.

[0032] Step S2: Unsupervised multi-constraint heterogeneous feature decoupling and alignment

[0033] In order to separate the shared disaster core semantics from the modality-specific environmental noise, the system performs a binary decomposition strategy on the extracted features to obtain alignment features for capturing cross-modal shared semantics and context features for preserving modality-private information.

[0034] The mathematical expression for the characteristic binary decomposition is as follows:

[0035]

[0036] In the formula, Let m be the alignment feature of the m-th mode of the i-th power grid node; The complete feature vector output by the encoder; Indicates taking the eigenvector The elements from dimension 0 to dimension (d-1) have a total of d dimensions; The context features of the m-th mode of the i-th power grid node; This indicates taking the elements from the dth dimension to the (2d-1)th dimension of the feature vector z, for a total of d dimensions; d is the dimension of a single feature subspace; 2d is the dimension of the complete feature vector.

[0037] In this step, the 2D features output by the encoder are divided into two parts along the feature dimension: the first d dimensions are used as alignment features, and the last d dimensions are used as context features. This fixed-dimensional splitting method is simple to implement, does not require the introduction of additional learnable splitting parameters, can reduce the training complexity of the model under unsupervised conditions, and at the same time ensures a balance in the dimensional capacity of the alignment features and context features. It can fully carry the core disaster semantics shared across modalities, and is also sufficient to retain the unique differentiated information of each modality.

[0038] Subsequently, the system stacks and merges the two types of features from all nodes along the channel dimension to construct a fused feature tensor, and then performs deep interaction through a multi-head attention mechanism. The features output by the multi-head attention mechanism are then normalized and flattened at each layer to generate a unified node feature representation. The mathematical expression of this process is as follows:

[0039]

[0040] In the formula, The input fusion feature tensor has dimensions B×N×2×d, where B is the number of samples in the training batch; N is the total number of power grid nodes; 2 is the number of modalities, i.e., visual modalities and temporal modalities; d is the dimension of a single feature subspace; Stack is the stacking operation along the specified dimension; dim=2 indicates that stacking is performed in the second dimension. Let represent a real tensor space of dimension B×N×2×d; MHA represents the multi-head attention mechanism function; Query, Key, and Value are the three input matrices of the multi-head attention mechanism, all set here as Flatten is the flattening operation, which flattens the last two dimensions of the tensor; H is the output node fusion feature matrix, with dimensions B×N×2d; This represents a real matrix space of dimension B×N×2d.

[0041] To achieve high-quality feature decoupling under unsupervised conditions, this invention designs a total alignment loss function that includes three constraints; the details are as follows: 1) Alignment Feature Consistency Constraint Loss: This loss, through contrastive learning, brings the aligned features of different modalities of the same sample closer together and pushes the features of different samples further apart, thereby enhancing cross-modal semantic consistency. The contrastive learning mechanism enables the model to automatically learn shared semantic representations across modalities without relying on labeled data. Its calculation formula is as follows:

[0042] In the formula, The loss value is the alignment feature consistency constraint; B is the number of samples in the training batch, and the summation index indicates that all samples in the batch are traversed; i is the index of the current anchor sample; m and n are the indices of different modal types, where m≠n indicates comparison between different modalities; k is the index of negative samples in the batch other than the anchor sample; log is the natural logarithm function; exp is the exponential function with the natural constant e as the base; sim is the cosine similarity function, used to calculate the similarity between two feature vectors. Let m be the alignment feature of the m-th mode of the i-th sample; Let be the alignment feature of the i-th sample and the n-th modality, and Forming positive sample pairs; The alignment feature of the k-th sample and the n-th modality, with Forming negative sample pairs; Temperature is a parameter used to control the smoothness of the similarity distribution; in this embodiment... =0.1.

[0043] Among them, cosine similarity The calculation formula is:

[0044] In the formula, u and v are two feature vectors for which the similarity is to be calculated; It is the transpose of u; Let u be the L2 norm; Let v be the L2 norm.

[0045] 2) Feature Independence Constraint Loss: The system uses the Hilbert-Schmidt independence criterion to calculate the statistical dependency between aligned features and context features, minimizing this dependency to eliminate feature redundancy and ensure that aligned features and context features encode mutually independent information. The calculation formula is as follows:

[0046] In the formula, B is the feature independence constraint loss value; B is the number of samples in the training batch; i is the current sample index; m is the modality type index; This indicates whether the currently processed features originate from a visual modality or a temporal modality; Let Hilbert-Schmidt independence criterion function be used. Let m be the alignment feature of the m-th mode of the i-th sample; The context features of the m-th modality for the i-th sample; in, The calculation is based on the Gaussian kernel function, and the specific formula is as follows:

[0047] In the formula, P and Q are the characteristic matrices of the two random variables whose independence is to be calculated; d is the dimension of the eigenvector; This is the matrix trace operation, which is the sum of the elements on the main diagonal of the matrix; The Gaussian kernel matrix is ​​calculated based on the characteristic matrix P; The Gaussian kernel matrix is ​​calculated based on the characteristic matrix Q. Centered matrix; is a d×d identity matrix; 1 is a column vector of length d consisting entirely of 1s; It is a row vector of length d consisting entirely of 1s.

[0048] 3) Context Feature Diversity Constraint Loss: The system uses Jensen-Shannon divergence to calculate the context feature diversity constraint loss, constraining the distribution differences of context features between different nodes to prevent feature collapse and ensure sufficient diversity of context features across different nodes; the calculation formula is as follows:

[0049] In the formula, is the context feature diversity constraint loss value; j and k are the indices of different samples, j≠k indicates that diversity constraints are applied between different samples; m is the modality type index; The context features of the m-th modality for the j-th sample; represents the contextual features of the m-th mode of the k-th sample; JSD is the Jensen-Shannon divergence function; Among them, JSD is based on The divergence is calculated using the following formula:

[0050] In the formula, p is the probability density function of the first feature distribution, assuming that the context features follow a Gaussian distribution; q is the probability density function of the second feature distribution; for The divergence function measures the difference between two probability distributions.

[0051] Combining the above three losses, the total loss of unsupervised modal alignment is defined as:

[0052] In the formula, represents the total loss value for unsupervised modal alignment; The loss value is the consistency constraint value for alignment features; These are the weighting coefficients for the feature independence constraint loss; The loss value is the feature independence constraint. These are the weighting coefficients for the context feature diversity constraint loss; The loss value is constrained by the diversity of contextual features.

[0053] This step achieves precise decoupling and quality enhancement of heterogeneous features under unsupervised conditions through triple-constraint loss, solving the problems of scarce disaster labels and high heterogeneity of multi-source data in the power sector, and significantly improving the utilization efficiency of multimodal data.

[0054] Step S3: Extreme weather evolution simulation based on compression-decompression architecture

[0055] Based on extracted fusion features The system employs a compression-decompression architecture to predict future weather conditions. This architecture is specifically designed to address the complex and unpredictable nature of complex extreme weather evolution, effectively balancing information loss and noise interference issues in long-sequence modeling.

[0056] 1) During the compression stage, high-order mutation features are extracted using a recursive time compressor (CR). The time compressor progressively compresses the temporal dimension through average pooling, and extracts high-order mutation features after linear transformation and layer normalization. The recursive definition of the i-th compression output and the internal structure formula of the time compressor CR block are as follows:

[0057]

[0058] In the formula, Let be the feature tensor after the i-th compression; Let i be the time compressor function for the i-th layer; Let i be the feature tensor after the (i-1)th compression, when i=1 = ; This is a function for a single-layer time compressor. This is a layer normalization operation used to stabilize the training process; This is a linear transformation operation that maps input features to a specified dimension; This is an average pooling operation used to compress time-series dimensions; The formula for calculating the temporal dimension of the output of the average pooling layer is:

[0059] In the formula, Let be the temporal dimension after the i-th pooling; The temporal dimension before the (i-1)th pooling; For fill size; The pooling kernel size; The step size.

[0060] 2) During the decompression stage, the spatiotemporal resolution is gradually restored through the time decompressor (DCR). The time decompressor gradually restores the spatiotemporal resolution through linear transformation, layer normalization, and GELU activation. The recursive definition of the j-th decompression output and the internal structure formula of the time decompressor (DCR) block are as follows:

[0061]

[0062] In the formula, Let J be the characteristic tensor after the j-th decompression. This is the function for the j-th layer time decompressor; Let J be the characteristic tensor after the (j-1)th decompression, when j=1. This is the final output of the compression stage; This is a function for a single-layer time decompressor. The activation function for the Gaussian error linear unit; For layer normalization operation; This is a linear transformation operation; After J rounds of decompression, the final output is the extreme weather prediction sequence for future times:

[0063] In the formula, This is an extreme weather forecast sequence; J represents the total number of decompression rounds. Let J be the characteristic tensor after the Jth decompression.

[0064] This step effectively distills high-order mutation features in long sequences through recursive compression operations, and gradually restores spatiotemporal details through decompression operations. The combination of these two methods significantly improves the prediction accuracy of the spatiotemporal evolution trend of extreme weather.

[0065] Step S4: Dynamic graph topology learning and spatial coupling analysis based on feature similarity

[0066] To address the spatial correlation of power grid disasters, the system does not rely on static power grid physical topology, but instead constructs a dynamic correlation graph based on feature similarity. Existing technologies mostly rely on static power grid physical topology for analysis, making it difficult to capture cross-regional meteorological coupling risks. This invention utilizes feature similarity to dynamically construct a node correlation graph, enabling the system to automatically identify and aggregate node information that is geographically distant but has similar meteorological risk patterns.

[0067] 1) The system calculates the association weight between node i and node j using Softmax-normalized cosine similarity, as shown in the following formula:

[0068] in,

[0069] In the formula, Let be the association weight between node i and node j, representing the degree of influence of node j on node i; Let be the fused feature vector of the i-th node; Let be the fused feature vector of the j-th node; is the fused feature vector of the k-th node; sim is the cosine similarity function; Temperature is a parameter used to control the smoothness of attention distribution; N is the total number of power grid nodes. for Transpose of; for The L2 norm; for The L2 norm.

[0070] 2) Based on the generated dynamic adjacency matrix, the system uses a graph convolutional network (GCN) for feature propagation and aggregation, as shown in the formula:

[0071] In the formula, This represents the output feature matrix of the l-th layer graph convolutional network. Use the Sigmoid activation function; It is a dynamic adjacency matrix. The dimension is N×N; Let be the input feature matrix of the l-th layer graph convolutional network; Let be the learnable weight matrix of the l-th layer graph convolutional network; 3) The features output by the graph convolutional network are passed through a feedforward network. After processing, a power grid state vector containing global spatial correlation information is obtained. :

[0072] In the formula, This is the power grid state vector; It is a feedforward neural network function, consisting of alternating layers of linear layers and nonlinear activation functions; Let L be the output feature matrix of the Lth layer of the graph convolutional network, where L is the total number of layers in the graph convolutional network.

[0073] This step dynamically constructs an association graph based on feature similarity, enabling the system to automatically identify and aggregate node information that is geographically distant but has similar meteorological risk patterns, accurately quantify the spatial coupling impact of complex disasters on the power grid, and fill the blind spots of traditional topology analysis.

[0074] Step S5: Four-dimensional disaster prediction and intelligent decision generation

[0075] The system will include extreme weather forecast features With power grid state characteristics The data is then concatenated and fused before being input into the prediction head network. The prediction head consists of convolutional layers and a feedforward neural network. Composition, outputting the four-dimensional prediction result Y:

[0076] In the formula, The result matrix for four-dimensional disaster prediction has dimensions B×N×4. This is a feedforward neural network function; This is a convolutional layer function used to further extract local features; This is an extreme weather forecast sequence; is the power grid state vector; ⊕ is the concatenation operator, which concatenates two feature tensors along their feature dimensions; B is the number of samples in the training batch; N is the total number of power grid nodes; 4 are the four dimensions of the prediction output, corresponding to disaster location, time, category, and level, respectively; Among them, the four-dimensional prediction results The specific meanings are as follows: L is the disaster location vector, representing the geographical location information of each power grid node where a disaster occurs; T is the disaster occurrence time vector, representing the predicted time of the disaster at each power grid node; C is the disaster category vector, representing the type of disaster at each power grid node, including categories such as rainstorm, blizzard, strong wind, and freezing rain; R is the disaster level vector, representing the severity level of the disaster at each power grid node.

[0077] The numerical results are then converted into natural language prompts, which are input into a pre-trained large language model (LLM) to generate specific pre-disaster prevention and post-disaster recovery plans. The large language model, combined with a power grid operation and maintenance knowledge base, transforms quantitative predictions into highly readable and logically rigorous operation and maintenance decision-making schemes.

[0078] This step transforms quantitative numerical prediction results into natural language prompts that drive a large language model to generate decision-making solutions. This breaks the limitations of traditional early warning systems that only output cold data, and directly provides dispatchers with practical guidance for pre-disaster prevention and post-disaster recovery.

[0079] To achieve collaborative optimization across multiple tasks, the model is jointly trained. The total system loss is a weighted average of the disaster classification loss, the meteorological regression loss, and the aforementioned unsupervised modal alignment total loss, as shown in the following formula:

[0080] in,

[0081]

[0082] In the formula, This represents the total loss value during joint training of the model; Weighting coefficients for the loss items in the disaster classification; Classify the loss value for the disaster; The weighting coefficients for the meteorological regression loss term; This represents the meteorological regression loss value. These are the weighting coefficients for the unsupervised modal alignment loss term; The total loss for unsupervised modal alignment; N represents the total number of samples in the training batch; The total number of categories representing disaster types or levels; The true label represents the i-th sample; This represents the predicted probability that the i-th sample output by the prediction head network belongs to the k-th class; Represents the true value of the i-th sample; Represents the model's predicted value; It represents the L2 norm of a vector.

[0083] By minimizing this objective function, the system can fully utilize multi-source heterogeneous data and improve the accuracy of complex disaster early warning even when labels are scarce. Multi-task joint training enables the three tasks of disaster classification, meteorological regression, and modality alignment to promote each other. The classification task provides semantic guidance for alignment, the regression task provides numerical constraints for prediction, and the alignment task provides high-quality feature representations for the former two.

[0084] The method of this invention acquires visual monitoring data and time-series monitoring data of power grid nodes, maps them to a unified feature space using an encoder, and performs feature binary decomposition to obtain cross-modal shared alignment features and modality-private context features. It employs an unsupervised multi-constraint alignment mechanism to achieve deep fusion of heterogeneous features, generating node fusion features. It recursively infers the evolution of extreme weather conditions based on a compression-decompression architecture. It dynamically constructs a node association graph based on feature similarity and uses a graph convolutional network to aggregate node features with similar risk patterns, generating a power grid state vector containing global spatial coupling information. After fusing the inferred meteorological features and power grid state features, it outputs a four-dimensional disaster prediction result, which is then transformed into a natural language prompt to drive a large language model to generate a response plan, achieving accurate early warning and intelligent decision-making for complex disasters.

[0085] Example 2

[0086] like Figure 2 As shown, a power grid composite disaster linkage early warning system based on four-dimensional three-dimensional collaborative system includes the following modules: The multimodal data acquisition module is used to acquire visual monitoring data and time-series monitoring data based on the distribution of power grid nodes. This module connects to various data sources such as remote sensing satellites, drones, meteorological sensors, and power monitoring equipment through interfaces to achieve unified acquisition of four-dimensional data from air, space, ground, and power.

[0087] The feature encoding module maps multimodal monitoring data to a unified-dimensional feature space through a visual encoder and a temporal encoder, obtaining visual and temporal features. The visual encoder consists of a multi-layer convolutional network and layer normalization stacks, while the temporal encoder is built based on a long short-term memory network and linear projection layers.

[0088] The feature decoupling and alignment module performs feature binary decomposition on visual and temporal features, decoupling the features of each modality into aligned features and contextual features. It then uses an unsupervised multi-constraint alignment mechanism for constraint optimization to obtain fused features. This module includes consistency constraint submodules, independence constraint submodules, and diversity constraint submodules, each calculating its corresponding constraint loss.

[0089] The extreme weather projection module is used to extract high-order abrupt change features based on fused features using a recursive time compression network, and then gradually restore the spatiotemporal resolution through a time decompression network to project extreme weather state sequences for future moments. This module contains multiple cascaded time compression and time decompression units.

[0090] The spatial coupling analysis module is used to calculate the feature similarity between nodes based on fused features to dynamically construct a node association graph. It uses graph convolutional networks for feature propagation and aggregation to generate a power grid state vector containing global spatial coupling information.

[0091] The disaster prediction module is used to fuse extreme weather state sequences with power grid state vectors and output four-dimensional disaster prediction results containing disaster location, time, category and level through the prediction head network.

[0092] The intelligent decision generation module is used to convert the four-dimensional disaster prediction results into structured natural language prompts, input them into a pre-trained large language model, and generate pre-disaster prevention and post-disaster recovery plans.

[0093] In this embodiment, the modules are sequentially connected to form a complete closed-loop chain from data acquisition, feature processing, disaster scenario simulation to decision generation. Data is transferred between modules through standardized interfaces to ensure the system's scalability and maintainability.

[0094] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the principles and essence of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A power grid composite disaster linkage early warning method based on four-dimensional three-dimensional collaborative mechanism, characterized in that, Includes the following steps: Multimodal monitoring data is acquired based on the distribution of power grid nodes, including visual monitoring data and time-series monitoring data. The visual monitoring data and the temporal monitoring data are mapped to a feature space of a unified dimension by a visual encoder and a temporal encoder, respectively, to obtain visual features and temporal features. The visual features and the temporal features are subjected to feature binary decomposition, and the features of each modality are decoupled into alignment features for capturing cross-modal shared semantics and context features for preserving modality-private information. The alignment features and the context features are constrained and optimized by an unsupervised multi-constraint alignment mechanism to obtain quality-enhanced fusion features; Based on the fusion features, a recursive time compression network is used to extract high-order mutation features, and a time decompression network is used to gradually restore the spatiotemporal resolution to deduce the extreme weather state sequence at future moments. Based on the fused features, the feature similarity between nodes is calculated to dynamically construct a node association graph. A graph convolutional network is used for feature propagation and aggregation to generate a power grid state vector containing global spatial coupling information. The extreme weather state sequence is fused with the power grid state vector, and a four-dimensional disaster prediction result containing the disaster location, time, category and level is output through the prediction head network; The four-dimensional disaster prediction results are converted into structured natural language prompts, which are then input into a pre-trained large language model to generate pre-disaster prevention and post-disaster recovery plans.

2. The method for early warning of complex power grid disasters based on four-dimensional three-dimensional collaborative coordination according to claim 1, characterized in that, The step of performing feature binary decomposition on the visual features and the temporal features decouples the features of each modality into alignment features for capturing cross-modal shared semantics and context features for preserving modality-private information, specifically: The feature vector output by the encoder is divided into two parts along the feature dimension: the first part is used as the alignment feature and the second part is used as the context feature. Both the visual features and the temporal features are decomposed into alignment features and contextual features.

3. The method for early warning of complex power grid disasters based on four-dimensional three-dimensional collaborative coordination according to claim 1, characterized in that, The step of optimizing the alignment features and context features through an unsupervised multi-constraint alignment mechanism to obtain quality-enhanced fusion features specifically includes: Calculate the alignment feature consistency constraint loss, which is used to bring the alignment features of different modalities of the same sample closer together and push the alignment features of different samples further apart; The statistical dependency between aligned features and context features is calculated using the Hilbert-Schmidt independence criterion, and the statistical dependency between aligned features and context features is minimized by feature independence constraint loss. The context feature diversity constraint loss is calculated using Jensen-Shannon divergence to constrain the distribution differences of context features between different nodes and prevent feature collapse. The weighted sum of the alignment feature consistency constraint loss, feature independence constraint loss, and context feature diversity constraint loss is taken as the total loss of unsupervised modality alignment; The alignment features and context features of all nodes are stacked and merged in the channel dimension to construct a fused feature tensor. A multi-head attention mechanism is used for deep interaction, and the node fused features after layer normalization and flattening are output.

4. The method for early warning of complex power grid disasters based on four-dimensional three-dimensional collaborative coordination according to claim 3, characterized in that, The formula for calculating the alignment feature consistency constraint loss is as follows: In the formula, The loss value is the consistency constraint value for alignment features; B represents the number of samples in the training batch, and the summation index indicates that all samples in the batch are traversed; i is the index of the current anchor sample; m and n are the indices of different modal types, where m≠n indicates comparison between different modalities; k is the index of negative samples in the batch other than the anchor sample; log is the natural logarithm function; exp is the exponential function with the natural constant e as the base; sim is the cosine similarity function, used to calculate the similarity between two feature vectors. Let m be the alignment feature of the m-th mode of the i-th sample; Let be the alignment feature of the i-th sample and the n-th modality, and Forming positive sample pairs; The alignment feature of the k-th sample and the n-th modality, with Forming negative sample pairs; Temperature is a parameter used to control the smoothness of the similarity distribution; Among them, cosine similarity The calculation formula is: In the formula, u and v are two feature vectors for which the similarity is to be calculated; It is the transpose of u; Let u be the L2 norm; Let v be the L2 norm.

5. The method for early warning of complex power grid disasters based on four-dimensional three-dimensional collaborative coordination according to claim 3, characterized in that, The formula for calculating the feature independence constraint loss is as follows: In the formula, B is the feature independence constraint loss value; B is the number of samples in the training batch; i is the current sample index; m is the modality type index; This indicates whether the currently processed features originate from a visual modality or a temporal modality; Let Hilbert-Schmidt independence criterion function be used. Let m be the alignment feature of the m-th mode of the i-th sample; The context features of the m-th modality for the i-th sample; in, The calculation is based on the Gaussian kernel function, and the specific formula is as follows: In the formula, P and Q are the characteristic matrices of the two random variables whose independence is to be calculated; d is the dimension of the eigenvector; This is the matrix trace operation, which is the sum of the elements on the main diagonal of the matrix; The Gaussian kernel matrix is ​​calculated based on the characteristic matrix P; The Gaussian kernel matrix is ​​calculated based on the characteristic matrix Q. Centered matrix; is a d×d identity matrix; 1 is a column vector of length d consisting entirely of 1s; It is a row vector of length d consisting entirely of 1s.

6. The method for early warning of complex power grid disasters based on four-dimensional three-dimensional collaborative coordination according to claim 3, characterized in that, The formula for calculating the context feature diversity constraint loss is as follows: In the formula, is the context feature diversity constraint loss value; j and k are the indices of different samples, j≠k indicates that diversity constraints are applied between different samples; m is the modality type index; The context features of the m-th modality for the j-th sample; represents the contextual features of the m-th mode of the k-th sample; JSD is the Jensen-Shannon divergence function; Among them, JSD is based on The divergence is calculated using the following formula: In the formula, p is the probability density function of the first feature distribution, assuming that the context features follow a Gaussian distribution; q is the probability density function of the second feature distribution; for The divergence function measures the difference between two probability distributions.

7. The method for early warning of complex power grid disasters based on four-dimensional three-dimensional collaborative coordination according to claim 1, characterized in that, Based on the fused features, a recursive time compression network is used to extract high-order mutation features, and a time decompression network is used to gradually restore the spatiotemporal resolution to extrapolate the extreme weather state sequence at future moments. Specifically: During the compression stage, the temporal dimension is gradually compressed through multiple recursive time compression operations to extract high-order mutation features. Each compression operation includes average pooling, linear transformation and layer normalization. During the decompression stage, the spatiotemporal resolution is gradually restored through multiple recursive time decompression operations. Each decompression operation includes linear transformation, layer normalization, and GELU activation processing. After multiple decompression operations, the output sequence of extreme weather predictions for future times is generated.

8. The method for early warning of complex power grid disasters based on four-dimensional three-dimensional collaborative coordination according to claim 1, characterized in that, The process of calculating the feature similarity between nodes based on the fused features to dynamically construct a node association graph, and using a graph convolutional network for feature propagation and aggregation to generate a power grid state vector containing global spatial coupling information, specifically involves: The cosine similarity between any two power grid nodes is calculated as the node association weight. The Softmax function is used to normalize all association weights to construct a dynamic adjacency matrix. Based on the dynamic adjacency matrix, a graph convolutional network is used for multi-layer feature propagation and aggregation. The features output by the graph convolutional network are then processed by a feedforward network to obtain a power grid state vector containing global spatial correlation information.

9. The method for early warning of complex power grid disasters based on four-dimensional three-dimensional collaborative coordination according to claim 1, characterized in that, The process of fusing the extreme weather state sequence with the power grid state vector and outputting a four-dimensional disaster prediction result containing disaster location, time, category, and level through a prediction head network is as follows: The extreme weather state sequence is concatenated and fused with the power grid state vector to form a joint feature; The joint features are input into the prediction head network, which consists of convolutional layers and a feedforward neural network, and outputs a four-dimensional prediction result, which includes information on the location, time, category, and level of the disaster.

10. A power grid composite disaster linkage early warning system based on four-dimensional three-dimensional collaborative operation, characterized in that, The system comprising the method of any one of claims 1 to 9 includes: The multimodal data acquisition module is used to acquire visual monitoring data and time-series monitoring data based on the distribution of power grid nodes; The feature encoding module is used to map multimodal monitoring data to a unified-dimensional feature space through a visual encoder and a temporal encoder to obtain visual features and temporal features; The feature decoupling and alignment module is used to perform feature binary decomposition on the visual features and the temporal features, decouple the features of each modality into alignment features and context features, and obtain fused features through constraint optimization by an unsupervised multi-constraint alignment mechanism. The extreme weather projection module is used to extract high-order mutation features based on the fused features using a recursive time compression network, and gradually restore the spatiotemporal resolution through a time decompression network to project the extreme weather state sequence at future moments. The spatial coupling analysis module is used to calculate the feature similarity between nodes based on the fused features to dynamically construct a node association graph, and to use a graph convolutional network to perform feature propagation and aggregation to generate a power grid state vector containing global spatial coupling information. The disaster prediction module is used to fuse the extreme weather state sequence with the power grid state vector, and output a four-dimensional disaster prediction result containing the disaster location, time, category and level through the prediction head network; The intelligent decision generation module is used to convert the four-dimensional disaster prediction results into structured natural language prompts, input them into a pre-trained large language model, and generate pre-disaster prevention and post-disaster recovery plans.

Images