Power prospective technology theme prediction method based on dynamic graph neural network

Abstract

The invention provides a power prospective technology theme prediction method based on a dynamic graph neural network. According to the method, multi-source data related to a power technology is obtained, and a natural language processing technology is utilized to extract a technical subject and a word vector representation corresponding to the technical subject. And classifying the technical topics through a clustering algorithm, and constructing a power technical topic map under a plurality of time nodes. A node of each power technology topic map represents a power technology topic. Based on the power technology theme maps, a dynamic graph neural network is adopted to carry out structural modeling on nodes and connection relations thereof, and a long-short-term memory network is combined to carry out time sequence modeling on dynamic information of node characteristics evolving along with time, so that prediction characteristic representation of each node is obtained. And calculating a potential score according to the node features, and identifying a power technology theme with high prospective potential. The method can effectively identify technical themes with high perspectiveness and potential in the electric power field, and provides powerful support for scientific and technological innovation and industrial upgrading of the electric power industry.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of power technology prediction, and particularly relates to a power forward-looking technical subject prediction method based on a dynamic graph neural network. BACKGROUND

[0002] With the rapid development of the power industry technology and the continuous transformation of the energy structure, how to identify the power technology subject with forward-looking and development potential in a large amount of technical information in a timely manner has become a key task to promote scientific and technological innovation and industrial upgrading. At present, the commonly used forward-looking technology identification methods mainly rely on expert evaluation or analysis methods based on static indicators such as keyword frequency and patent quantity, and lack in-depth modeling of the evolution process of technical subjects, making it difficult to reflect the complex semantic relationship and structural dynamics between technologies.

[0003] At present, the technical subject identification method mainly relies on artificial expert evaluation or statistical analysis method based on static indicators, such as keyword frequency, literature quantity, patent quantity, etc. This kind of method lacks in-depth modeling of technical semantic connotation, and cannot effectively reflect the semantic similarity, citation relationship and potential evolution trend between different technical subjects. At the same time, this kind of method generally does not consider the dynamic process of technical subjects changing over time, and it is difficult to fully reveal the evolution path and potential influence of technical development.

[0004] In the prior art, Chinese patent CN118863041A discloses a knowledge graph evolution method based on a dynamics propagation model, including the following steps: step one: obtaining a multi-modal industrial knowledge graph of a product object, clustering the industrial knowledge nodes in the knowledge graph; converting the clustering results into a knowledge event group, and defining the association relationship between the knowledge event groups; step two: modeling each knowledge event group using a dynamics propagation model, using a differential equation to describe the change of its state over time; converting all knowledge event groups in the knowledge graph into a differential equation, and adding an evolution factor to the knowledge event group association relationship; step three: dynamically modeling the knowledge graph using a dynamic graph neural network, and embedding the differential equation into the dynamic graph neural network; step four: biased training of the dynamic graph neural network, and summarizing the change of the knowledge event group in the time sequence according to the dynamic change of the knowledge event group state and the evolution factor. However, this method still has the following problems:

[0005] (1) It is mainly aimed at the construction and evolution modeling of the knowledge graph in general industrial field, and has limited adaptability for some technical fields with strong specialization and complex terminology system;

[0006] (2) It mainly takes structural propagation modeling as the core, and lacks fine mining of the implicit semantic relationship in multi-source text;

[0007] (3) Failure to systematically introduce quantifiable technology evolution evaluation indicators lacks a reliable basis for evaluating the forward-looking potential of technology topics;

[0008] (4) The description of the changes in node characteristics in the time series is relatively rough, which makes it difficult to accurately reveal the dynamic evolution characteristics of the technical theme;

[0009] (5) It focuses on structural modeling and fails to fully combine semantic expression with information fusion in the time dimension, resulting in limited predictive capabilities.

[0010] Therefore, when faced with professional technical topic identification tasks with high semantic complexity and strong dynamic evolution, existing methods still have problems such as insufficient accuracy, insufficient granularity, and weak predictability, making it difficult to meet the needs of actual technology forward-looking analysis. Summary of the Invention

[0011] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide a method for predicting electric power forward-looking technology topics based on dynamic graph neural network.

[0012] The purpose of the present invention can be achieved by the following technical solutions:

[0013] The present invention provides a method for predicting electric power forward-looking technology topics based on a dynamic graph neural network, comprising the following steps:

[0014] Acquiring multi-source data related to power technology, including technical literature, patents, academic papers, and technical reports;

[0015] Processing the multi-source data using natural language processing techniques to extract technical topics and their corresponding word vector representations;

[0016] A clustering algorithm is used to classify the technical topics, and the technical topics are divided into several topic categories according to the similarity of the word vectors of each technical topic, so as to obtain a set of power technology topics;

[0017] Based on the electric power technology topic set, construct an electric power technology topic graph at multiple time nodes, wherein the nodes of the electric power technology topic graph are the electric power technology topics, the edges are the citation relationships between the electric power technology topics, and the edge weights are determined by the citation frequency and relevance;

[0018] Based on the power technology theme diagram at multiple time nodes, a dynamic graph neural network is used to structurally model each node and its connection relationship. The long short-term memory network (LSTM) is then combined to perform time series modeling on the dynamic information of the node features evolving over time, thereby obtaining the predicted node feature representation of each node in the power technology theme diagram.

[0019] Calculate the potential score of each power technology theme based on the predicted node feature representation of each node in the power technology theme graph;

[0020] According to the potential score, an electric power technology subject with high forward-looking potential is identified.

[0021] Further, the multi-source data is processed by natural language technology to extract technical subjects and corresponding word vector representations thereof, specifically including:

[0022] A plurality of time nodes are obtained from multi-source text data in the electric power field to construct an original text data set Wherein t represents a time node, and n is the number of texts.

[0023] The original text data set is preprocessed, including removing stop words, word segmentation, part-of-speech tagging, and named entity recognition, to obtain a structured text representation.

[0024] Based on an electric power technology vocabulary and a domain ontology, the entities and keywords of the structured text representation are mapped and screened in the domain to extract a representative set of technical terms T = {T1, T2, …, T m}, wherein T i represents the i-th technical subject candidate term, and m is the total number of extracted terms.

[0025] Each technical term T i extracted is word vectorized using a pre-trained BERT language model to obtain a corresponding word vector representation X i = Encoder(T i ), and the word vector set X = {X1, X2, …, X m} of all technical terms is normalized.

[0026] Further, the word vector similarity calculation formula is:

[0027]

[0028] Wherein, sin(X i ,X j ) is the similarity of the word vector X i of the technical term and the word vector X j of the technical term,

[0029] Further, based on the set of electric power technology subjects, an electric power technology subject graph under a plurality of time nodes is constructed, specifically including:

[0030] At each time node T t (t = 1, 2, …, n), based on the electric power technology subjects appearing in the corresponding time period and their reference relationship, an electric power technology subject graph Gt =(V t ,E t ,W t ,F t ), where V t Indicates that at time node T t A collection of electrical technology topics that appear in E t W represents the set of directed edges established based on the reference relationship between topics; t represents the weight set of each edge, F t Represents the node feature matrix, where each power technology theme node v i Corresponding to a set of structured features.

[0031] Furthermore, the method for determining the reference relationship includes: if the original text corresponding to a certain technical topic explicitly quotes or semantically associates the text corresponding to another topic, it is considered that there is a reference relationship between the two, and a directed edge is established from the referenced topic to the citing topic.

[0032] Furthermore, the edge weight calculation formula is:

[0033] w ij =α1·C ij +α2·sin(v j ,v i )

[0034] Among them, w ij represents the weight of the edge connecting node i and node j, α1 and α2 are preset weights, sim(v j ,v i ) is the word vector v of node i i and the word vector v of node j j The similarity, C ij It represents the citation batch of power technology topic j cited from power technology topic i. The calculation formula is:

[0035]

[0036] Among them, N is the number of original texts in the time window, γ is the weighting coefficient, and RefMatch k (v j ,v i ) is the v in the kth text j Is v explicitly cited in the references? i , defined as follows:

[0037]

[0038] CoOccur k (v j ,vi ) is the semantic co-occurrence score of v j in the kth document, calculated as: i

[0039]

[0040] where cosine(v j ,v i ) is the cosine similarity of the word vector v i of node i and the word vector v j of node j, d k is the number of power technology topics appearing in the kth document.

[0041] Further, the node feature matrix includes time novelty, concept drift rate, and cross-domain fusion degree, wherein:

[0042] The time novelty represents the first occurrence time of the technology topic and its growth rate over time, and the formula is:

[0043]

[0044] where TNI i is the time novelty of the ith power technology topic node, τ i is the time interval from the first occurrence of the power technology topic v i to the current time node T t , f i (t) is the occurrence frequency of the power technology topic v i within the time node T t , f i (t -1) is the occurrence frequency of the power technology topic v i within the previous time node T t-1 , and ∈ is a very small positive number.

[0045] The concept drift rate represents the rate of change of the topic semantic vector over time, and the formula is:

[0046]

[0047] where CDI i is the concept drift rate of the ith power technology topic node, is the word vector representation of the power technology topic v i at the time node T t , and is the word vector representation of the power technology topic v i at the previous time node T t-1 ​word vector representation, cosine is cosine similarity;

[0048] The cross-domain fusion degree represents the frequency degree of technical topics across different subfields, and the formula is:

[0049]

[0050] CFI i is the cross-domain fusion degree of the i-th power technology topic node, represents the power technology topic v i involved in the current time node T t , and D is the total number of predefined power subfields.

[0051] Further, the power technology topic graph under multiple time nodes is used to model the structure of each node and its connection relationship by using a dynamic graph neural network, and the dynamic information of the evolution of node features over time is modeled by using a long short-term memory network (LSTM) to obtain the predicted node feature representation of each node in the power technology topic graph. Specifically, it includes:

[0052] Obtain power technology topic graph data {G1, G2,..., G T} under multiple time nodes, and for each time node t, use a dynamic graph neural network EvolveGCN to model the connection relationship between nodes in the power technology topic graph G t , specifically including:

[0053] For each node v i , the node feature matrix F i (t) at time t is obtained by aggregating the information of its adjacent nodes through graph convolution operation to obtain the structural representation of the node

[0054] Use a long short-term memory network (LSTM) to model the node feature matrix, and for each node v i , collect its feature representation time series [F i (1) ,F i (2) ,..., F i (T) ] at different time nodes t, and input it into the LSTM network for processing s i =LSTM([F i (1) ,F i (2) ,..., F i (T) ]) to obtain the node v iTemporal dynamic representation s in the whole time sequence i ;

[0055] Node structure representation obtained by EvolveGCN And the temporal dynamic representation s obtained by LSTM i Fusion, get the predicted node feature representation z of each node i :

[0056]

[0057] Where [·][·] represents the feature splicing operation, MLP is a multi-layer perceptron, used to fuse and map the spliced feature vector to generate a unified dimensional dynamic representation z i .

[0058] Further, the potential score of each power technology theme is calculated according to the predicted node feature representation of each node in the power technology theme graph, and the formula is:

[0059]

[0060] Where P i is the potential score of node v i , β1, β2 are weight coefficients, z i is the predicted node feature representation of node v i , Decoder is a feature decoding function, N(i) represents the neighbor node set of node v i , w ij represents the weight of the connection edge between node i and node j.

[0061] Further, the power technology theme with high forward-looking potential is identified according to the potential score, specifically including: sorting the potential scores of each node power technology theme, and selecting the highest score of the preset number of power technology themes as the power technology theme with high forward-looking potential.

[0062] Compared with the prior art, the present application has the following advantages:

[0063] (1) The present application extracts technical themes and word vector representation through multi-source heterogeneous data fusion and natural language processing technology, solves the problem of insufficient semantic understanding of terms in existing methods, realizes the consistency of semantic expression of power technology concepts, and helps to accurately extract domain technical themes.

[0064] (2) The present invention constructs a time-evolving electric power technology topic graph and introduces reference relationships and semantic similarity to jointly determine edge weights, thereby making up for the defect that traditional static analysis methods based on co-occurrence or frequency cannot capture the technology association structure, achieving more fine-grained topic relationship modeling and improving the expressive power of the graph structure.

[0065] (3) The present invention adopts a dynamic graph neural network (such as EvolveGCN) to model the time-varying relationship between nodes, and combines LSTM to capture the temporal variation of node features. Compared with the existing technology that only uses structural modeling or static graph methods, it more accurately depicts the evolution of power technology topics in the time dimension, thereby improving the prediction effect of future technology hotspots.

[0066] (4) The present invention constructs indicators such as temporal novelty, concept drift rate, and cross-domain integration and integrates the influence of neighboring nodes to perform potential scoring, effectively solving the problem of lack of quantitative evaluation basis in traditional methods, achieving a comprehensive evaluation of the potential value of power technology topics, and improving the scientificity and reliability of high-value technology identification.

[0067] (5) The present invention ultimately selects power technology topics with forward-looking potential by ranking the potential scores, which not only enhances the pertinence of technology forecasts but also provides strong data support for scientific research layout, patent mining and policy formulation in the power industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 is a flow chart of the method of the present invention;

[0069] Figure 2 This is the evaluation index model diagram of the present invention. DETAILED DESCRIPTION

[0070] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0071] Example 1:

[0072] This embodiment provides a method for predicting power technology topics based on dynamic graph neural network. Figure 1 、 Figure 2 As shown, the following steps are included:

[0073] Step S1: Acquire multi-source data related to power technology, including technical literature, patents, academic papers, and technical reports;

[0074] Step S2: processing the multi-source data through natural language technology to extract technical topics and their corresponding word vector representations;

[0075] Processing the multi-source data through natural language technology to extract technical topics and their corresponding word vector representations, specifically including:

[0076] Obtaining multi-source text data from the power field at multiple time nodes to construct an original text data set Where t represents the time node, and n is the number of texts.

[0077] Processing the original text data set Preprocessing, including removing stop words, tokenization, part-of-speech tagging, and named entity recognition, to obtain structured text representations.

[0078] Based on the power technology vocabulary and domain ontology, the entities and keywords of the structured text representations are mapped and filtered in the domain to extract a representative set of technical terms T = {T1, T2, …, T m} where T i represents the i-th technical topic candidate term, and m is the total number of extracted terms.

[0079] Using a pre-trained BERT language model to perform word vectorization on each technical term T i extracted, obtaining the corresponding word vector representation X i = Encoder(T i ), and normalizing the word vector set X = {X1, X2, …, X m}.

[0080] Step S2 processes multi-source text data in the power field through natural language processing. First, the original text is cleaned and structured to ensure information standardization and semantic integrity. Then, representative technical terms are extracted in combination with the power field ontology and vocabulary, and a pre-trained BERT language model is used to represent the terms in word vectors to capture deep semantic associations between terms. Finally, the word vectors are normalized to improve the accuracy and stability of similarity calculation. Through this step, a consistent representation of the power technology semantic space is constructed, and high-quality, structured input data is provided for subsequent technical topic clustering and graph modeling, thereby enhancing the model's ability to capture the evolution trend of power technology topics and improving the accuracy and scientificity of forward-looking technology identification.

[0081] Step S3: classifying technical topics using a clustering algorithm, dividing them into several topic categories based on the word vector similarity of each technical topic, and obtaining a set of power technology topics; the word vector similarity calculation formula is:

[0082]

[0083] wherein, sin(X i ,X j ) is the similarity of the word vector X i of the technical term with the word vector X j of the technical term.

[0084] Step S4: based on the set of power technology topics, constructing power technology topic graphs under multiple time nodes, the nodes of the power technology topic graph being each power technology topic, the edges being the citation relationships between the power technology topics, and the edge weights being determined by the citation frequencies and the correlation degrees;

[0085] Based on the set of power technology topics, constructing power technology topic graphs under multiple time nodes, specifically comprising:

[0086] At each time node T t (t = 1, 2, …, n), based on the power technology topics appearing in the corresponding time period and their citation relationships, constructing a power technology topic graph G t = (V t , E t , W t , F t ), wherein V t represents the set of power technology topics appearing in the time node T t ; E t represents the set of directed edges established according to the citation relationships between the topics; W t represents the set of weights of each edge, and F t represents the node feature matrix, wherein each power technology topic node v i corresponds to a set of structured features.

[0087] The judgment method of the citation relationship includes: if a certain technical topic corresponds to the original text that explicitly cites or semantically associates another topic corresponding to the text, it is considered that the two exist a citation relationship, and a directed edge from the cited topic to the cited topic is established.

[0088] The weight calculation formula of the edge is:

[0089] w ij = α1·C ij + α2·sin(v j ,v i )

[0090] wherein, w ij represents the weight of the edge connecting node i and node j, α1 and α2 are preset weights, sim(v j ,v i) is the word vector of node i i The similarity of word vector v j ij C k represents the reference batch from power technology subject i to power technology subject j, and the calculation formula is:

[0091]

[0092] where N is the number of original texts in the time window, γ is the weighting coefficient, RefMatch j (v i ) is the word vector v j whether v i is explicitly referenced in the reference, which is defined as follows:

[0093]

[0094] CoOccur k (v j ,v i ) is the semantic co-occurrence score of v j and v i in the kth text, and the calculation method is:

[0095]

[0096] where cosine(v j ,v i ) is the cosine similarity of word vector v i of node i and word vector v j of node j, and d k is the number of power technology subjects appearing in the kth literature.

[0097] The node feature matrix includes time novelty, concept drift rate, and cross-domain fusion degree, where:

[0098] Time novelty represents the first appearance time of the technology subject and its growth rate over time, and the formula is:

[0099]

[0100] where TNI i is the time novelty of the ith power technology subject node, τ i is the time interval from the first appearance of power technology subject v i to the current time node T t , f i (t) is the growth rate of power technology subject v i at time node T t ​The frequency of occurrence within, f i (t -1) For the power technology theme v i At the previous time node T t-1 The frequency of occurrence in ,∈ is a very small positive number;

[0101] The concept drift rate indicates the rate at which the topic semantic vector changes over time, and the formula is:

[0102]

[0103] Among them, CDI i is the concept drift rate of the i-th power technology theme node, For the power technology theme v i At time node T t The word vector representation of For the power technology theme v i At the previous time node T t-1 The word vector representation of , cosine is the cosine similarity;

[0104] The degree of cross-domain integration indicates the frequency with which a technical topic spans different sub-domains. The formula is:

[0105]

[0106] Among them, CFI i is the cross-domain integration degree of the i-th power technology theme node, Indicates the power technology theme v i At the current time node T t The number of power sub-sectors involved in the project, and D is the total number of predefined power sub-sectors.

[0107] By constructing a power technology topic graph at multiple time points, the technical topics are modeled as nodes in the graph structure, and edges are constructed through explicit references and semantic associations. This not only truly reflects the inheritance and development of technologies, but also leverages deep semantic information to compensate for the information missing in traditional structural graphs. Furthermore, the edge weights, composed of reference relationships and word vector similarity, quantify the strength of associations between different topics, making the graph structure more aligned with the actual logic of technological evolution and helping to improve the discriminative and generalization capabilities of subsequent graph neural network learning. Temporal novelty, concept drift rate, and cross-domain integration are introduced as node features to comprehensively capture the dynamic evolution of technical topics. Temporal novelty measures the time of emergence and development speed of a technology, helping to identify emerging technologies in rapid growth. Concept drift rate reflects the changing trends in technical semantics and can reveal whether the technical content has undergone significant changes. Cross-domain integration characterizes the diffusion of a technology across multiple sub-domains, reflecting its technological influence and universality. The node feature matrix composed of these features not only improves the representation quality of nodes in the technical topic graph but also provides rich, dynamic, and predictive input information for dynamic graph neural networks and time series modeling.

[0108] In summary, this series of designs not only makes the structure of the technology theme map more reasonable and timely, but also enhances the model's insight into the development potential of technology themes, thereby achieving effective prediction of power technology themes with high forward-looking potential, and ultimately improving the accuracy and practicality of the entire prediction scheme.

[0109] Step S5: Based on the power technology theme diagram at multiple time nodes, a dynamic graph neural network is used to perform structural modeling on each node and its connection relationship. The long short-term memory network (LSTM) is then used to perform time series modeling on the dynamic information of the node features evolving over time, thereby obtaining a predicted node feature representation for each node in the power technology theme diagram.

[0110] Based on the power technology theme diagram at multiple time nodes, a dynamic graph neural network is used to structurally model each node and its connection relationship. The long short-term memory network (LSTM) is combined with time series modeling to dynamically model the dynamic information of node features evolving over time. The predicted node feature representation of each node in the power technology theme diagram is obtained, specifically including:

[0111] Get the power technology theme diagram data {G1, G2, ..., G T}, for each time node t, the power technology theme graph G t , using the dynamic graph neural network EvolveGCN to structurally model the connection relationship between nodes, specifically including:

[0112] For each node v i The node feature matrix F at time ti (t) , the structural representation of the node is obtained by aggregating the information of its adjacent nodes through the graph convolution operation

[0113] The node feature matrix is modeled using a long short-term memory network (LSTM), and the feature representation of each node v i at different time nodes t is collected to form a time sequence [F i (1) ,F i (2) ,...,F i (T) ] and input into the LSTM network for processing s i =LSTM([F i (1) ,F i (2) ,...,F i (T) ]) to obtain the time series dynamic representation s i of the node v i in the entire time sequence;

[0114] The node structural representation obtained by EvolveGCN and the time series dynamic representation s i obtained by LSTM are fused to obtain the predicted node feature representation z i of each node:

[0115]

[0116] where [·][·] represents a feature concatenation operation, and MLP is a multi-layer perceptron used to fuse and map the concatenated feature vectors to generate a unified dynamic representation z i .

[0117] This step constructs a power technology topic graph at multiple time nodes and introduces a dynamic graph neural network combined with LSTM modeling, aiming to comprehensively extract information on the structural relationship and time evolution of technology topics in two dimensions and improve the prediction ability of technology development trends. EvolveGCN can dynamically capture the semantic dependence of graph structure changes over time, enhance the correlation expression between nodes, and avoid the neglect of dynamic technology evolution by static graph methods. LSTM models the long-term dependence of node features over time and extracts the time development pattern of technology topics. Through the fusion of structural representation and time series representation, a unified predicted feature representation is generated, enabling the model to have both graph structure perception ability and time evolution perception ability. This setting effectively improves the understanding and prediction accuracy of the model for the evolution of power technology topics and enhances the adaptability and accuracy of the overall scheme in complex evolution scenarios.

[0118] Step S6: Calculate the potential score of each power technology theme according to the predicted node feature representation of each node in the power technology topic graph.

[0119] According to the predicted node feature representation of each node in the power technology topic graph, the potential score of each power technology theme is calculated, and the formula is:

[0120]

[0121] Where P is the potential score of node v, β1, β2 are weight coefficients, z is the predicted node feature representation of node v, Decoder is the feature decoding function, N(i) represents the neighbor node set of node v, w represents the weight of the connection edge between node i and node j. i i i i i ij

[0122] Step S7: Identify the power technology theme with high forward-looking potential according to the potential score.

[0123] According to the potential score, the power technology theme with high forward-looking potential is identified, which specifically includes: sorting the potential scores of each node power technology theme, and selecting the top pre-set number of power technology themes with the highest scores as the power technology themes with high forward-looking potential.

[0124] Embodiment 2:

[0125] The embodiment provides a power forward-looking technology theme prediction system based on a dynamic graph neural network, which comprises:

[0126] Data acquisition module: This module is used to acquire multi-source data from the power field, including technical literature, patents, academic papers and technical reports, etc., as the original data input of the system.

[0127] Data processing module: This module processes the acquired multi-source data through natural language processing technology, extracts the technology theme and its word vector representation. Specifically, it includes text preprocessing (such as removing stop words, word segmentation, part-of-speech tagging, etc.), entity recognition and domain mapping, and uses BERT model to word vectorize technical terms.

[0128] Theme classification and clustering module: This module classifies the technology theme using clustering algorithm, divides it into several theme categories based on word vector similarity, and generates a set of power technology themes.

[0129] ​​​​​​Power Technology Topic Graph Construction Module: Based on a collection of power technology topics, this module constructs a power technology topic graph at multiple time points. The nodes in the graph represent power technology topics, and the edges represent the citation relationships between technology topics. Edge weights are determined by citation frequency and relevance.

[0130] Dynamic graph neural network modeling module: This module uses dynamic graph neural networks (such as EvolveGCN) to perform structural modeling on the power technology theme graph at multiple time nodes, aggregate node adjacent information, and combine LSTM to model the dynamic information of node features changing over time to obtain predicted node feature representation.

[0131] Potential score calculation module: This module calculates the potential score of each power technology topic based on the predicted node feature representation and identifies technology topics with forward-looking potential based on the score.

[0132] If the above functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0133] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and such modifications or substitutions are intended to be within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be subject to the scope of protection of the claims.

Claims

1. A method for predicting power forward-looking technology topics based on dynamic graph neural network, characterized in that: The following steps are involved: Acquiring multi-source data related to power technology, including technical literature, patents, academic papers, and technical reports; Processing the multi-source data using natural language processing techniques to extract technical topics and their corresponding word vector representations; A clustering algorithm is used to classify the technical topics, and the technical topics are divided into several topic categories according to the similarity of the word vectors of each technical topic, so as to obtain a set of power technology topics; Based on the electric power technology topic set, construct an electric power technology topic graph at multiple time nodes, wherein the nodes of the electric power technology topic graph are the electric power technology topics, the edges are the citation relationships between the electric power technology topics, and the edge weights are determined by the citation frequency and relevance; Based on the power technology theme diagram at multiple time nodes, a dynamic graph neural network is used to structurally model each node and its connection relationship. The long short-term memory network (LSTM) is then combined to perform time series modeling on the dynamic information of the node features evolving over time, thereby obtaining the predicted node feature representation of each node in the power technology theme diagram. Calculate the potential score of each power technology theme based on the predicted node feature representation of each node in the power technology theme graph; Based on the potential scores, power technology themes with high forward-looking potential are identified.

2. The method for predicting electric power forward-looking technology topics based on dynamic graph neural network according to claim 1 is characterized in that: The multi-source data is processed using natural language technology to extract technical topics and their corresponding word vector representations, specifically including: Obtain multi-source text data from the power sector at multiple time points and construct the original text dataset Where t represents the time node and n is the number of texts; For the original text dataset Perform preprocessing, including stop word removal, word segmentation, part-of-speech tagging, and named entity recognition, to obtain structured text representation; Based on the electric power technology vocabulary and domain ontology, the entities and keywords represented by the structured text are mapped and filtered to extract the representative technical term set T = {T1, T2, ..., T m }, where T i represents the i-th candidate technical topic term, and m is the total number of extracted terms; Use the pre-trained BERT language model to train each technical term T extracted i Perform word vectorization to obtain the corresponding word vector representation X i =Encoder(T i ), the word vector set X = {X1, X2, ..., X m } is normalized.

3. The method for predicting electric power forward-looking technology topics based on dynamic graph neural network according to claim 1 is characterized in that: The word vector similarity calculation formula is: Among them, sin(X i ,X j ) is the word vector X of technical terms i and word vectors X of technical terms j similarity.

4. The method for predicting electric power forward-looking technology topics based on dynamic graph neural network according to claim 1 is characterized in that: The constructing of a power technology theme map at multiple time nodes based on the power technology theme set specifically includes: At each time point T t (t=1,2,…,n), based on the power technology topics and their citation relationships that appear in the corresponding time period, the power technology topic graph G is constructed. t =(V t ,E t ,W t ,F t ), where V t Indicates that at time node T t A collection of electrical technology topics that appear in E t W represents the set of directed edges established based on the reference relationship between topics; t represents the weight set of each edge, F t Represents the node feature matrix, where each power technology theme node v i Corresponding to a set of structured features.

5. The method for predicting electric power forward-looking technology topics based on dynamic graph neural network according to claim 4 is characterized in that: The method for judging the reference relationship includes: if the original text corresponding to a certain technical topic explicitly quotes or semantically associates the text corresponding to another topic, it is considered that there is a reference relationship between the two, and a directed edge is established from the referenced topic to the citing topic.

6. The method for predicting electric power forward-looking technology topics based on dynamic graph neural network according to claim 4 is characterized in that: The edge weight calculation formula is: w ij =α1·C ij +α2·sin(v j ,v i ) Among them, w ij represents the weight of the edge connecting node i and node j, α1 and α2 are preset weights, sim(v j ,v i ) is the word vector v of node i i and the word vector v of node j j The similarity, C ij It represents the citation batch of power technology topic j cited from power technology topic i. The calculation formula is: Among them, N is the number of original texts in the time window, γ is the weighting coefficient, and RefMatch k (v j ,v i ) is the v in the kth text j Is v explicitly cited in the references? i , defined as follows: CoOccur k (v j ,v i ) is the v in the kth text j With v i The semantic co-occurrence score of is calculated as: Among them, cosine(v j ,v i ) is the word vector v of node i i and the word vector v of node j j The cosine similarity of d k is the number of power technology topics appearing in the k-th document.

7. The method for predicting electric power forward-looking technology topics based on dynamic graph neural network according to claim 4 is characterized in that: The node feature matrix includes temporal novelty, concept drift rate, and cross-domain integration degree, where: The temporal novelty indicates the time when a technical topic first appears and its growth rate over time, and the formula is: Among them, TNI i is the temporal novelty of the i-th power technology theme node, τ i For the power technology theme v i The first occurrence is T from the current time node t The time interval, f i (t) For the power technology theme v i At time node T t The frequency of occurrence within, f i (t-1) For the power technology theme v i At the previous time node T t-1 The frequency of occurrence in ,∈ is a very small positive number; The concept drift rate represents the rate at which the topic semantic vector changes over time, and the formula is: Among them, CDI i is the concept drift rate of the i-th power technology theme node, For the power technology theme v i At time node T t The word vector representation of For the power technology theme v i At the previous time node T t-1 The word vector representation of , cosine is the cosine similarity; The cross-domain integration degree indicates the frequency of technical topics spanning different sub-domains, and the formula is: Among them, CFI i is the cross-domain integration degree of the i-th power technology theme node, Indicates the power technology theme v i At the current time node T t The number of power sub-sectors involved in the project, and D is the total number of predefined power sub-sectors.

8. The method for predicting electric power forward-looking technology topics based on dynamic graph neural network according to claim 1 is characterized in that: Based on the power technology theme diagram at multiple time nodes, the dynamic graph neural network is used to perform structural modeling on each node and its connection relationship. The long short-term memory network (LSTM) is combined to perform time series modeling on the dynamic information of the node features evolving over time, and the predicted node feature representation of each node in the power technology theme diagram is obtained, specifically including: Get the power technology theme diagram data {G1, G2, ..., G T }, for each time node t, the power technology theme graph G t , using the dynamic graph neural network EvolveGCN to structurally model the connection relationship between nodes, specifically including: For each node v i The node feature matrix F at time t i (t) , aggregate the information of its adjacent nodes through graph convolution operations to obtain the structural representation of the node Use the long short-term memory network LSTM to model the node feature matrix, and for each node v i , collect its feature representation time series at different time nodes t [F i (1) ,F i (2) ,...,F i (T) ] and input into the LSTM network for processing s i =LSTM([F i (1) ,F i (2) ,...,F i (T) ]), get node v i Temporal dynamic representation of the entire time series i ; The node structure obtained by EvolveGCN is represented as And the temporal dynamic representation s obtained by LSTM i Fusion is performed to obtain the predicted node feature representation z of each node i : Among them, [·][·] represents the feature concatenation operation, and MLP is a multi-layer perceptron, which is used to fuse and map the concatenated feature vectors to generate a dynamic representation z of uniform dimension. i .

9. The method for predicting electric power forward-looking technology topics based on dynamic graph neural network according to claim 1, characterized in that: The potential score of each power technology theme is calculated based on the predicted node feature representation of each node in the power technology theme graph. The formula is: Among them, P i For node v i The potential score, β1 and β2 are weight coefficients, z i For node v i The predicted node feature representation, Decoder is the feature decoding function, N(i) represents the node v i The set of neighbor nodes, w ij Represents the weight of the edge connecting node i and node j.

10. The method for predicting electric power forward-looking technology topics based on dynamic graph neural network according to claim 1, characterized in that: The identifying of electric power technology topics with high foresight potential based on the potential scores specifically includes: sorting the potential scores of the electric power technology topics at each node, and selecting a preset number of electric power technology topics with the highest scores as electric power technology topics with high foresight potential.

Images