Power grid science and technology hotspot evolution identification and trend prediction method based on dynamic atlas and multi-dimensional time sequence fusion

Abstract

The invention belongs to the technical field of artificial intelligence, and particularly relates to a power grid science and technology hotspot evolution recognition and trend prediction method based on dynamic atlas and multi-dimensional time sequence fusion, which comprises the following steps of: firstly, converting three types of heterogeneous original data of papers, patents and project reports in the field of a power grid into a unified available format, and cleaning texts; meanwhile, extracting a high-weight keyword by using a TF-IDF algorithm; secondly, abstracting keywords into topic nodes, constructing a binary relation graph and a higher-order ternary hypergraph based on document co-occurrence, and describing direct and collaborative association between topics; on the basis, the binary co-occurrence graph and the ternary hypergraph are constructed at the same time, the high-order cooperation relation of'theme-project-classification ', 'mechanism-theme-patent' and the like can be captured, the inscription limitation that only a two-dimensional co-occurrence network is relied on traditionally is broken through, and the perception depth of the scientific research achievement multi-element cooperation mode is improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a method for identifying and predicting the evolution of power grid technology hotspots based on dynamic graphs and multi-dimensional time series fusion. Background Technology

[0002] As the power system continues its transformation towards intelligence, deep integration of renewable energy, and digitalization, research findings in the field of power grid technology are characterized by frequent updates, diverse sources, and high correlation. A large number of papers, patents, and project reports exist in parallel as text, structured data, and time-series information. Traditional power grid knowledge graphs often focus on the organization of static entities and relationships, or are only fully updated at low frequencies such as monthly or quarterly, lacking the ability to capture the rapid iteration of research topics at the micro-temporal scale. At the same time, existing trend prediction studies often rely solely on the time series of patents / documents, ignoring the dynamic network structure between results, high-order multi-faceted collaboration between topics, and deep semantic evolution.

[0003] In recent years, while there are entity embedding and completion models based on dynamic knowledge graphs in general domains, as well as cases predicting enterprise technology evolution, these models are difficult to directly transfer to power grid technology scenarios due to differences in data scale, domain terminology, and collaboration patterns. Furthermore, they lack interpretable causal reasoning and online adaptive capabilities, failing to meet the comprehensive needs of scientific research management and technological decision-making. A few industry-specific cases (such as patent evolution prediction in the semiconductor and automotive industries) use co-occurrence networks combined with LSTM to predict growth trends, but these models only focus on the time series of single patent quantities, ignoring graph structure, semantic information, and higher-order collaborations. In power grid load forecasting and equipment failure prediction, convolutional neural networks (CNNs) and long short-term memory networks (LSTMs) have been applied to process time-series signals, but these methods focus on the time series of physical quantities such as current and voltage, lacking research on the time series of scientific research outputs and technological themes.

[0004] In summary, existing work suffers from the following shortcomings: Static or infrequent updates; most patents and papers focus on static graph construction or periodic (e.g., monthly) full reconstruction, failing to respond in real-time to sudden hot topics. Single-indicator-driven analysis; often relying solely on document / patent quantity or co-occurrence frequency, neglecting network structure features, higher-order collaborative relationships, and semantic temporal components. Lack of trend prediction capabilities; existing solutions are mostly used for knowledge retrieval, reasoning, or fault diagnosis, lacking quantitative prediction models for the future trends of scientific and technological topics. Insufficient interpretability and adaptability; lacking causal reasoning and online drift detection mechanisms, unable to provide experts with interpretable hot topic evolution mechanisms, and struggling to automatically correct for changes in data distribution.

[0005] Therefore, to meet the needs of the power grid technology field for real-time hotspot identification, accurate trend prediction, and interpretable causal insights from massive heterogeneous dynamic results, this invention proposes a method for identifying and predicting the evolution of power grid technology hotspots based on dynamic graphs and multi-dimensional temporal series fusion. This method constructs a dynamic knowledge graph containing binary co-occurrence graphs and ternary hypergraphs by standardizing papers, patents, and project reports and extracting high-frequency keywords. Furthermore, it extracts multi-dimensional temporal features such as normalized activity, evolution rate, network centrality changes, and eigenvector components. Self-supervised pre-training is used to enhance node and high-order collaborative representations. A dual-stream spatiotemporal deep network, fusing spatial channel (GNN), temporal channel (Transformer), and cross-attention, is designed to spatiotemporally model node representations. Finally, a multilayer perceptron is used to predict future topic activity, and causal correction driven by prediction error gradients generates interpretable hotspot evolution paths, achieving accurate and traceable prediction of power grid technology development trends. Summary of the Invention

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: A method for identifying and predicting the evolution and trends of power grid technology hotspots based on dynamic graph fusion and multi-dimensional time series analysis includes the following specific steps: S1, Standardization and Preprocessing of Multi-Source Heterogeneous Data: Transform three types of heterogeneous raw data in the power grid field—papers, patents, and project reports—into a unified and usable format, clean the text, and extract high-weight keywords using the TF-IDF algorithm; S2, Dynamic Graph and Hypergraph Construction: Keywords are abstracted into topic nodes, and binary relation graphs and higher-order ternary hypergraphs are constructed based on document co-occurrence to characterize the direct and collaborative relationships between topics; S3, Multidimensional Temporal Feature Extraction: It is difficult to capture the activity and trend changes of the topic by graph structure alone. Therefore, the behavior of nodes in each time period is quantified into multidimensional indicators for spatiotemporal model fusion to capture the activity, growth rate and influence changes of nodes in the network. S4, Self-Supervised Pre-training: Graph neural networks and hypergraph autoencoders are pre-trained through unlabeled self-supervised tasks to enhance the model's ability to represent nodes and higher-order patterns. Node contrastive learning (InfoNCE) and hypergraph reconstruction (HyperGAE) are employed to learn effective low-order vector representations of nodes and graphs, which serve as inputs to subsequent spatiotemporal models. Furthermore, after self-supervised pre-training, the model obtains general representations of node and hyperedge patterns, and the obtained pre-trained parameters serve as initial parameters for subsequent spatiotemporal fusion models, improving the model's convergence speed and generalization ability. S5, Spatiotemporal Dual-Stream Fusion Model: Designs parallel spatial and temporal stream networks, combining the advantages of graph neural networks (GNN) in spatial modeling with the advantages of transformers in temporal modeling, and achieves efficient information fusion through cross-attention; S6, Trend Prediction and Causal Correction: First, future hot spot indicators are predicted based on multilayer perceptron (MLP), and the causal relationship between hot spots is further improved through error-driven causal correction; then, Granger test is used to screen significant causal edges, and the prediction error gradient is backpropagated to the causal graph.

[0007] As a preferred embodiment of the method for identifying and predicting the evolution of power grid technology hotspots based on dynamic graphs and multi-dimensional time series fusion described in this invention, the specific steps of S1 are as follows: S11 collects raw data in the power grid field from three dimensions: papers, patents, and research projects. Papers extract titles, abstracts, authors, publication dates, and keywords; patents extract application numbers, authorization numbers, IPC / CPC classification numbers, abstracts, and application dates; and project reports extract project names, project types, abstracts, start and end dates, and amounts. S12 uses a unified data dictionary and transformation rules for formatting, where papers, patents, and project reports are converted into JSON-LD, XML, and CSV formats, respectively. S13, clean the text content, use regular expressions to extract time and category tags, remove stop words and punctuation marks, and perform stemming and word segmentation; S14, Finally, based on the TF-IDF method, high-weight keywords for each document are extracted:

[0008] in, Indicates terms In the document The number of times it appears in Indicates terms The number of documents, This indicates the total number of documents in the corpus.

[0009] As a preferred embodiment of the method for identifying and predicting the evolution of power grid technology hotspots based on dynamic graphs and multi-dimensional time series fusion described in this invention, the specific steps of S2 are as follows: S21, generate topic nodes, use the K-means algorithm to cluster all keywords, merge synonyms and near-synonyms to obtain a standard topic set. ; S22, traverse all data to construct binary and ternary hypergraphs as input for subsequent models:

[0010] in, Indicates time period The adjacency matrix, Represents a node During the period The number of times they appear together Represents a node During the period The probability of occurrence Represents a node During the period The probability of them occurring simultaneously It is the fusion coefficient; the ternary hypergraph is:

[0011] in, Let R represent the 3D hypergraph tensor, and let R denote the decomposition rank. Indicates the weighting factor. A vector representing the decomposition of the hypergraph structure. Represents the cross product of vectors; S23, When adding new data, use an exponential smoothing update method with a fixed time period as the unit:

[0012] in, This represents the adjacency matrix and tensor calculated based on window data for the new time period. This represents the smoothing coefficient between historical and newly added data.

[0013] As a preferred embodiment of the method for identifying and predicting the evolution of power grid technology hotspots based on dynamic graphs and multi-dimensional time series fusion described in this invention, wherein: the combined feature vector constructed in S3 is:

[0014] in, Indicates normalized activity. Indicates the rate of fireworks. Indicates the change in degree center. Principal eigenvector components, Indicates time period Number of internal documents and This represents the mean and standard deviation of the node within the sliding window. For a small constant, Indicates time period Degree centrality of internal nodes Representation matrix The eigenvalues ​​corresponding to the principal eigenvector.

[0015] As a preferred embodiment of the method for identifying and predicting the evolution of power grid technology hotspots based on dynamic graphs and multi-dimensional time series fusion described in this invention, wherein: the loss function in the pre-training process of S4 is:

[0016] in, Represents the set of positive examples. Denotes the set of negative examples. It is a control factor. Represents a binary adjacency matrix. This represents a ternary hypergraph tensor.

[0017] As a preferred embodiment of the method for identifying and predicting the evolution of power grid technology hotspots based on dynamic graph and multi-dimensional time series fusion described in this invention, wherein: the graph neural network (GNN) in S5 implements the input of the spatial channel, including the side channel and the hypergraph channel:

[0018] in, Indicates an edge channel. Indicates a hypergraph channel. Indicates the edge channel weight. , Denotes the decomposition factor of the hypergraph. This represents the ReLU activation function.

[0019] As a preferred embodiment of the method for identifying and predicting the evolution of power grid technology hotspots based on dynamic graphs and multi-dimensional time series fusion described in this invention, wherein: the time channel in S5 adopts the Transformer model, and the input is a sequence of spatial feature matrices for T time periods: .

[0020] As a preferred embodiment of the method for identifying and predicting the evolution of power grid technology hotspots based on dynamic graphs and multi-dimensional time series fusion described in this invention, wherein: in step S5, a multi-head attention mechanism is used to fuse the output features of the spatiotemporal model.

[0021] in, Represents the mapping matrix, Representational spatial model, The spatiotemporal model is represented by d, which represents the attention dimension; the final node representation is obtained after fusion. It integrates spatial structure and temporal dynamic information as input for subsequent predictions.

[0022] As a preferred embodiment of the power grid technology hotspot evolution identification and trend prediction method based on dynamic graph and multi-dimensional time series fusion described in this invention, wherein: the formulas for trend prediction and causal correction in S6 are:

[0023] in, The final fused representation of the nodes, L represents the difference between the predicted value and the expected value, and L represents the loss function of the multilayer perceptron.

[0024] As a preferred embodiment of the power grid technology hotspot evolution identification and trend prediction method based on dynamic graph and multi-dimensional time series fusion described in this invention, wherein: the formula for backpropagating the prediction error gradient to the causal graph in step S6 is:

[0025] in, Representing the subject in a cause-and-effect diagram and theme The edge weights are determined, and the edge weights are dynamically increased or decreased based on the sign and magnitude of the gradient to update the causal relationship.

[0026] Compared with existing technologies: This invention offers the following advantages: Multi-granularity structural representation: Simultaneously constructing binary co-occurrence graphs and ternary hypergraphs enables the capture of high-order collaborative relationships such as "topic-project-classification" and "institution-topic-patent," breaking through the rigid limitations of traditional two-dimensional co-occurrence networks and enhancing the depth of perception of diverse collaborative modes of scientific research results. Spatiotemporal deep fusion: Integrating multi-dimensional indicators such as normalized activity, evolution rate, degree centrality change, and principal feature vector components at the node level; through the parallel dual-stream of the spatial channel (GNN) and the temporal channel (Transformer), and achieving seamless information exchange through cross-attention, ensuring that the model can learn high-order features of the network structure while capturing temporal dynamics and sudden changes. Self-supervised pre-training and few-shot generalization: Employing self-supervised tasks of InfoNCE contrastive learning and hypergraph reconstruction, the model can learn robust node and hyperedge representations even in unlabeled or few-shot scenarios, significantly improving the ability to identify emerging hot topics in "zero-shot" or "few-shot" scenarios and accelerating downstream convergence speed. Explainable causal correction mechanism: The prediction error gradient is applied inversely to the causal edge weights to dynamically enhance or weaken the causal relationship between topics and generate a visualized causal path graph; compared with black box prediction, it provides a logical explanation and enhances decision-makers' confidence in the model output.

[0027] In summary, this invention constructs a dynamic knowledge graph containing binary co-occurrence graphs and ternary hypergraphs by standardizing the processing of papers, patents, and project reports and extracting high-frequency keywords. It further extracts multi-dimensional temporal features such as normalized activity, evolution rate, network centrality changes, and feature vector components. Self-supervised pre-training is used to enhance node and high-order collaborative representations. A dual-stream spatiotemporal deep network, fusing spatial channel (GNN), temporal channel (Transformer), and cross-attention, is designed to spatiotemporally model node representations. Finally, a multilayer perceptron is used to predict future topic activity, and causal correction driven by prediction error gradients generates interpretable hotspot evolution paths, achieving accurate and traceable prediction of power grid technology development trends. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below.

[0029] This invention provides a method for identifying and predicting the evolution of power grid technology hotspots based on dynamic graphs and multi-dimensional time series fusion, including the following specific steps: S1, Standardization and Preprocessing of Multi-Source Heterogeneous Data: Transform three types of heterogeneous raw data in the power grid field—papers, patents, and project reports—into a unified and usable format, clean the text, and extract high-weight keywords using the TF-IDF algorithm; The specific steps of S1 are as follows: S11 collects raw data in the power grid field from three dimensions: papers, patents, and research projects. Papers extract titles, abstracts, authors, publication dates, and keywords; patents extract application numbers, authorization numbers, IPC / CPC classification numbers, abstracts, and application dates; and project reports extract project names, project types, abstracts, start and end dates, and amounts. S12 uses a unified data dictionary and transformation rules for formatting, where papers, patents, and project reports are converted into JSON-LD, XML, and CSV formats, respectively. S13, clean the text content, use regular expressions to extract time and category tags, remove stop words and punctuation marks, and perform stemming and word segmentation; S14, Finally, based on the TF-IDF method, high-weight keywords for each document are extracted:

[0030] in, Indicates terms In the document The number of times it appears in Indicates terms The number of documents, This indicates the total number of documents in the corpus.

[0031] S2, Dynamic Graph and Hypergraph Construction: Keywords are abstracted into topic nodes, and binary relation graphs and higher-order ternary hypergraphs are constructed based on document co-occurrence to characterize the direct and collaborative relationships between topics; The specific steps of S2 are as follows: S21, generate topic nodes, use the K-means algorithm to cluster all keywords, merge synonyms and near-synonyms to obtain a standard topic set. ; S22, traverse all data to construct binary and ternary hypergraphs as input for subsequent models:

[0032] in, Indicates time period The adjacency matrix, Represents a node During the period The number of times they appear together Represents a node During the period The probability of occurrence Represents a node During the period The probability of them occurring simultaneously It is the fusion coefficient; the ternary hypergraph is:

[0033] in, Let R represent the 3D hypergraph tensor, and let R denote the decomposition rank. Indicates the weighting factor. A vector representing the decomposition of the hypergraph structure. Represents the cross product of vectors; S23, When adding new data, use an exponential smoothing update method with a fixed time period as the unit:

[0034] in, This represents the adjacency matrix and tensor calculated based on window data for the new time period. This represents the smoothing coefficient between historical and newly added data.

[0035] S3, Multidimensional Temporal Feature Extraction: Graph structures alone are insufficient to capture the activity and trend changes of a topic. Therefore, the behavior of nodes in each time period is quantified into multidimensional indicators for fusion in the spatiotemporal model to capture the activity, growth rate, and influence changes of nodes in the network. The constructed combined feature vector is as follows:

[0036] in, Indicates normalized activity. Indicates the rate of fireworks. Indicates the change in degree center. Principal eigenvector components, Indicates time period Number of internal documents and This represents the mean and standard deviation of the node within the sliding window. For a small constant, Indicates time period Degree centrality of internal nodes Representation matrix The eigenvalues ​​corresponding to the principal eigenvector.

[0037] S4, Self-Supervised Pre-training: A graph neural network and a hypergraph autoencoder are pre-trained through an unlabeled self-supervised task to enhance the model's ability to represent nodes and higher-order patterns. Node contrastive learning (InfoNCE) and hypergraph reconstruction (HyperGAE) are employed to learn effective low-order vector representations of nodes and the graph, which serve as inputs to the subsequent spatiotemporal model. The loss function during pre-training is:

[0038] in, Represents the set of positive examples. Denotes the set of negative examples. It is a control factor. Represents a binary adjacency matrix. This represents a ternary hypergraph tensor. After self-supervised pre-training, the model obtains a general representation of node and hyperedge patterns. The obtained pre-trained parameters serve as the initial parameters for the subsequent spatiotemporal fusion model, improving the model's convergence speed and generalization ability.

[0039] S5, Spatiotemporal Dual-Stream Fusion Model: This model designs parallel spatial and temporal stream networks, combining the advantages of Graph Neural Networks (GNNs) in spatial modeling with the advantages of Transformers in temporal modeling. Efficient information fusion is achieved through cross-attention. The GNNs handle the input to the spatial channels, including edge channels and hypergraph channels.

[0040] in, Indicates an edge channel. Indicates a hypergraph channel. Indicates the edge channel weight. , Denotes the decomposition factor of the hypergraph. This represents the ReLU activation function. The time channel uses a Transformer model, with the input being a sequence of spatial feature matrices over T time intervals:

[0041] A multi-head attention mechanism is used to fuse the output features of the spatiotemporal model:

[0042] in, Represents the mapping matrix, Representational spatial model, The spatiotemporal model is represented by d, which represents the attention dimension; the final node representation is obtained after fusion. It integrates spatial structure and temporal dynamic information as input for subsequent predictions.

[0043] S6, Trend Prediction and Causal Correction: Predict future hot spot indicators based on multilayer perceptron (MLP), and further improve the causal relationship between hot spots through error-driven causal correction;

[0044] in, The final fused representation of the nodes, Let L represent the difference between the predicted and expected values, and L represent the loss function of the multilayer perceptron. A Granger test is used to screen for significant causal edges, and the prediction error gradient is backpropagated to the causal graph.

[0045] in, Representing the subject in a cause-and-effect diagram and theme The edge weights are determined, and the edge weights are dynamically increased or decreased based on the sign and magnitude of the gradient to update the causal relationship.

[0046] Based on the above, the present invention provides, but is not limited to, the following embodiments: Example 1: The dataset used is the publicly available Microsoft Academic Graph (MAG) dataset, which contains metadata for papers in the field of computer science, covering a period of 60 months from January 2016 to December 2020. The dataset contains 150,000 papers; the number of topic nodes after keyword clustering is n=1200; and the number of item and category nodes is p=300 and c=50, respectively.

[0047] The specific steps for implementing this invention are as follows: (1) Data preprocessing Metadata for computer science papers was extracted from Microsoft Academic Graph (MAG), covering 150,000 papers from January 2016 to December 2020. The original JSON format was converted to JSON-LD, and fields including title, abstract, keywords, and publication date were extracted. Regular expressions were used to extract metadata (publication month / year, author, institution), and English stop words and common noise characters were removed. Stemming and word segmentation of the abstracts were performed based on a domain dictionary, TF-IDF was calculated per document, and the top 10 keywords with the highest weights were selected.

[0048] (2) Dynamic graph and hypergraph construction

[0049] Cluster the keywords mentioned above (K-means, k=1200) to generate a topic node set, and set a sliding window. Moon, sliding step size In one month, T=55 snapshots are obtained. Within each snapshot, if two topics appear in the same document, the number of times is... To establish the edges of the graph structure, add edges to the co-occurrence graph and assign values:

[0050] Simultaneously, a ternary tensor is constructed for the project / category nodes (p=300, c=50). High-order patterns were extracted using CP decomposition (R=20). Both the historical graph and the hypergraph underwent exponential smoothing. Incremental updates.

[0051] (3) Multidimensional temporal feature extraction

[0052] For each node In each snapshot Calculate the following characteristic indicators and normalize the activity level. ,in For document count, For in-window statistics; evolution rate Degree centrality change ; Main eigenvector components ,Pick The largest eigenvalue corresponds to the vector components, which are combined into .

[0053] (4) Self-supervised pre-training

[0054] Node-based comparative learning: Construct 200,000 positive and negative pairs, τ=0.5, and the InfoNCE loss converges to less than 0.1.

[0055] Hypergraph Reconstruction: HyperGAE was used to reconstruct binary and ternary structures separately, and the reconstruction error was... It is less than 0.01 after 100 epochs.

[0056] (5) Spatiotemporal dual-stream fusion model

[0057] Spatial flow contains 2 layers of GNN, 128 hidden units, and processes... With decomposition factors; the time flow contains 4 Transformer layers, 8 attention heads, and 256 hidden dimensions; the dimensions of cross-attention... ,pass Self-adaptive integration.

[0058] (6) Trend prediction and causal correction

[0059] MLP two-layer (64→1), input Learning rate On the test set collected in Q1–Q2 2021, RMSE=0.032. Initial causal edges were filtered by Granger and the causal graph was dynamically updated using the prediction error gradient to generate interpretable hotspot evolution paths.

[0060] The prediction results of this scheme are compared with those of several other models as follows:

[0061] Example 2: The specific steps for implementing this invention are as follows: (1) Data preprocessing This study retrieved 8,500 science and technology reports, 2,300 patents, and 950 projects from China Southern Power Grid from January 2018 to June 2024. These were converted to JSON-LD, XML, and CSV formats. The original JSON format was then converted to JSON-LD, and the extracted fields included: title, abstract, keywords, and publication date. Regular expressions were used to extract metadata (publication month / year, author, and institution), and English stop words and common noise characters were removed. The abstracts were stemmed and segmented based on a domain dictionary, and TF-IDF was calculated per document. The top 8 keywords with the highest weights were selected.

[0062] (2) Dynamic graph and hypergraph construction

[0063] The above keywords are clustered using K-means to generate a set of topic nodes. Project Nodes Classification nodes Set up a sliding window Moon, sliding step size In one month, T=22 snapshots are obtained. Within each snapshot, if two topics appear in the same document, the number of times is... To establish the edges of the graph structure, add edges to the co-occurrence graph and assign values:

[0064] Simultaneously, a ternary tensor is constructed for the project / category nodes (p=200, c=40). Higher-order patterns were extracted using CP decomposition (R=15). Both the historical graph and the hypergraph underwent exponential smoothing. Incremental updates.

[0065] (3) Multidimensional temporal feature extraction For each node In each snapshot Calculate the following characteristic indicators and normalize the activity level. ,in For document count, For in-window statistics; evolution rate Degree centrality change ; Main eigenvector components ,Pick The largest eigenvalue corresponds to the vector components, which are combined into .

[0066] (4) Self-supervised pre-training

[0067] Node-based comparative learning: Construct 120,000 positive and negative pairs, τ=0.7, and the InfoNCE loss is less than 0.08.

[0068] Hypergraph Reconstruction: HyperGAE was used to reconstruct binary and ternary structures separately, and the reconstruction error was... It is less than 0.01 after 100 epochs.

[0069] (5) Spatiotemporal dual-stream fusion model

[0070] Spatial flow contains 3 layers of GNN, 128 hidden units, and processes... With decomposition factors; the time flow contains 6 Transformer layers, 4 attention heads, and 128 hidden dimensions; the dimensions of cross-attention. ,pass Self-adaptive integration.

[0071] (6) Trend prediction and causal correction

[0072] MLP two-layer (64→1), input Learning rate On the test set collected in Q1–Q2 2021, RMSE=0.026. Initial causal edges were filtered by Granger, and the causal graph was dynamically updated using the prediction error gradient to generate interpretable hotspot evolution paths.

[0073] The prediction results of this scheme are compared with those of several other models as follows:

[0074] Although the present invention has been described above with reference to embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of the invention. In particular, as long as there is no structural conflict, the features in the disclosed embodiments can be combined with each other in any manner. The lack of an exhaustive description of these combinations in this specification is merely for the sake of brevity and resource conservation. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A method for identifying and predicting the evolution of power grid technology hotspots based on dynamic graph fusion and multi-dimensional time series fusion, characterized in that, The specific steps are as follows: S1, Standardization and Preprocessing of Multi-Source Heterogeneous Data: Transform three types of heterogeneous raw data in the power grid field—papers, patents, and project reports—into a unified and usable format, clean the text, and extract high-weight keywords using the TF-IDF algorithm; S2, Dynamic Graph and Hypergraph Construction: Keywords are abstracted into topic nodes, and binary relation graphs and higher-order ternary hypergraphs are constructed based on document co-occurrence to characterize the direct and collaborative relationships between topics; S3, Multidimensional Temporal Feature Extraction: Quantify the behavior of nodes in each time period into multidimensional indicators for spatiotemporal model fusion, so as to capture the activity, growth rate and influence changes of nodes in the network; S4, Self-supervised pre-training: Pre-train graph neural networks and hypergraph autoencoders through unlabeled self-supervised tasks to enhance the model's ability to represent nodes and high-order patterns; learn effective low-order vector representations of nodes and graphs using node contrastive learning and hypergraph reconstruction, which serve as inputs for subsequent spatiotemporal models; S5, Spatiotemporal Dual-Stream Fusion Model: Designs parallel spatial and temporal stream networks, combining the advantages of graph neural networks in spatial modeling with the advantages of Transformers in temporal modeling, and achieves efficient information fusion through cross-attention; S6, Trend Prediction and Causal Correction: First, future hot spot indicators are predicted based on multilayer perceptron, and the causal relationship between hot spots is further improved through error-driven causal correction; then, Granger test is used to screen significant causal edges, and the prediction error gradient is backpropagated to the causal graph.

2. The method for identifying and predicting the evolution of power grid technology hotspots based on dynamic graphs and multi-dimensional time series fusion as described in claim 1, is characterized in that, The specific steps of S1 are as follows: S11 collects raw data in the power grid field from three dimensions: papers, patents, and research projects. S12 uses a unified data dictionary and transformation rules for formatting, where papers, patents, and project reports are converted into JSON-LD, XML, and CSV formats, respectively. S13, clean the text content, use regular expressions to extract time and category tags, remove stop words and punctuation marks, and perform stemming and word segmentation; S14, Finally, based on the TF-IDF method, high-weight keywords for each document are extracted: in, Indicates terms In the document The number of times it appears in Indicates terms The number of documents, This indicates the total number of documents in the corpus.

3. The method for identifying and predicting the evolution of power grid technology hotspots based on dynamic graphs and multi-dimensional time series fusion as described in claim 1, is characterized in that... The specific steps of S2 are as follows: S21, generate topic nodes, use the K-means algorithm to cluster all keywords, merge synonyms and near-synonyms to obtain a standard topic set. ; S22, traverse all data to construct binary and ternary hypergraphs as input for subsequent models: in, Indicates time period The adjacency matrix, Represents a node During the period The number of times they appear together Represents a node During the period The probability of occurrence Represents a node During the period The probability of them occurring simultaneously It is the fusion coefficient; the ternary hypergraph is: in, Let R represent the 3D hypergraph tensor, and let R denote the decomposition rank. Indicates the weighting factor. A vector representing the decomposition of the hypergraph structure. Represents the cross product of vectors; S23, When adding new data, use an exponential smoothing update method with a fixed time period as the unit: in, This represents the adjacency matrix and tensor calculated based on window data for the new time period. This represents the smoothing coefficient between historical and newly added data.

4. The method for identifying and predicting the evolution of power grid technology hotspots based on dynamic graphs and multi-dimensional time series fusion as described in claim 1, is characterized in that... The combined feature vector constructed in S3 is: in, Indicates normalized activity. Indicates the rate of fireworks. Indicates the change in degree center. Principal eigenvector components, Indicates time period Number of internal documents and This represents the mean and standard deviation of the node within the sliding window. For a small constant, Indicates time period Degree centrality of internal nodes Representation matrix The eigenvalues ​​corresponding to the principal eigenvector.

5. The method for identifying and predicting the evolution of power grid technology hotspots based on dynamic graphs and multi-dimensional time series fusion as described in claim 1, characterized in that, The loss function in the pre-training process of S4 is: in, Represents the set of positive examples. Denotes the set of negative examples. It is a control factor. Represents a binary adjacency matrix. This represents a ternary hypergraph tensor.

6. The method for identifying and predicting the evolution of power grid technology hotspots based on dynamic graphs and multi-dimensional time series fusion as described in claim 1, characterized in that, The graph neural network in S5 implements the input of spatial channels, including edge channels and hypergraph channels: in, Indicates an edge channel. Indicates a hypergraph channel. Indicates the edge channel weight. , Denotes the decomposition factor of the hypergraph. This represents the ReLU activation function.

7. The method for identifying and predicting the evolution of power grid technology hotspots based on dynamic graphs and multi-dimensional time series fusion as described in claim 6, is characterized in that... The time channel in S5 adopts the Transformer model, and the input is a sequence of spatial feature matrices for T time periods: 。 8. The method for identifying and predicting the evolution of power grid technology hotspots based on dynamic graphs and multi-dimensional time series fusion as described in claim 7, is characterized in that, The S5 uses a multi-head attention mechanism to fuse the output features of the spatiotemporal model: in, Represents the mapping matrix, Representational spatial model, The spatiotemporal model is represented by d, which represents the attention dimension; the final node representation is obtained after fusion. It integrates spatial structure and temporal dynamic information as input for subsequent predictions.

9. The method for identifying and predicting the evolution of power grid technology hotspots based on dynamic graphs and multi-dimensional time series fusion as described in claim 1, characterized in that, The formulas for trend prediction and causal correction in S6 are as follows: in, The final fused representation of the nodes, L represents the difference between the predicted value and the expected value, and L represents the loss function of the multilayer perceptron.

10. The method for identifying and predicting the evolution of power grid technology hotspots based on dynamic graphs and multi-dimensional time series fusion according to claim 9, characterized in that, The formula for backpropagating the prediction error gradient to the causal graph in S6 is as follows: in, Representing the subject in a cause-and-effect diagram and theme The edge weights are determined, and the edge weights are dynamically increased or decreased based on the sign and magnitude of the gradient to update the causal relationship.

Images