Electric power prospective technology topic evaluation method based on neural topic modeling

Abstract

The invention provides an electric power prospective technology topic evaluation method based on neural topic modeling, which comprises the following steps: acquiring text data related to an electric power technology, and preprocessing the text data; performing topic extraction on the preprocessed text data by adopting a neural topic modeling method, and identifying a plurality of prospective power technology topics; constructing a multi-dimensional index system for evaluating the power technology themes, and calculating scores of all the technology themes under the index system; and calculating a comprehensive evaluation score of each prospective power technology subject based on an entropy weight method. Through automatic topic extraction and multi-dimensional evaluation, the electric power technology topic with perspectiveness can be effectively identified, the performance of the electric power technology topic in the aspects of strength, stability, emerging degree, leading-edge degree and the like can be accurately evaluated, and then a scientific basis is provided for technology planning and decision making in the electric power field. The method can help technology research and development and industrial development directions to be accurately mastered, and improves the strategic planning capability of the electric power technology.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of power technology evaluation, and particularly relates to a power forward-looking technology theme evaluation method based on neural topic modeling. BACKGROUND

[0002] With the continuous development of power technology and the acceleration of energy transformation process, emerging technologies in the power field are constantly emerging. How to identify, evaluate and track forward-looking power technology themes in a timely manner has become a key problem in scientific and technological strategic planning, technology research and development layout, and industrial development decision-making. Existing technology theme identification methods mostly rely on manual classification or keyword-based rule matching, which have low intelligence, limited processing scale, and difficulty in dealing with complex semantic information, making it difficult to effectively handle large-scale, multi-source and heterogeneous text data processing requirements.

[0003] In addition, most existing technology theme evaluation methods are based on single-dimensional indicators such as the number of patents, the number of papers, or the number of citations, lacking systematic, comprehensive and multi-dimensional quantitative evaluation, making it difficult to fully depict the feature evolution and value potential of technology themes in terms of novelty, sustainability, influence and diffusion, and easily causing misjudgment or delayed identification of technology prospects. At the same time, some methods fail to fully consider the implicit semantic structure and potential theme relationship between documents, limiting the accuracy and forward-looking nature of technology theme identification.

[0004] In the prior art, Chinese patent CN113988087A discloses a method and device for technology theme multi-index calculation and trend prediction, wherein the method comprises: acquiring a plurality of paper text data, and performing preprocessing operations on key fields of the plurality of paper text data, wherein the key fields include: paper title, abstract and keywords; performing multi-feature weight calculation based on the preprocessed key fields, and extracting technology themes from the weighted multi-feature fields using an LDA topic model; based on the extracted technology themes, designing a plurality of index measurement methods for measuring the advancement of technology themes, and statistically calculating the index values of the technology theme indexes, wherein the technology theme indexes include: intensity, stability, emerging degree and front degree; based on the index values, using a Logistic model to predict the development trend of the technology themes, and performing comparative analysis between the technology themes. This method extracts multiple features of the text, clearly defines the theme division, and forms a general technology theme index calculation and trend prediction method.

[0005] However, the method still has the following shortcomings: the topic modeling still uses the traditional LDA topic model, which often has difficulty accurately identifying implicit semantic features when facing text data with strong professional characteristics, complex terminology, and diverse semantics in the power field, and the topic granularity is relatively coarse, making it difficult to reflect the real association between complex technologies. Secondly, its index system focuses on statistical quantitative analysis, although it covers intensity, emerging degree and other dimensions, but fails to introduce indicators reflecting the evolution characteristics of technology structure, such as classification diffusion degree, citation level, cross-border integration degree, etc., lacking a systematic description of the multi-dimensional evolution path of technology topics. In addition, this method does not fully combine the literature network structure and deep semantic analysis means, and does not construct a complete technology transmission path map, limiting the accuracy and interpretability of trend prediction. Finally, as a general technology topic evaluation tool, it lacks specific adaptation and optimization for the actual technology development characteristics of the power field, and it is difficult to meet the actual needs of the power industry for high-precision identification and dynamic tracking of forward-looking technology topics.

[0006] Therefore, it is urgent to propose a power technology field-oriented, neural network topic modeling and multi-dimensional index system integrated, semantic understanding ability strong, structure perception ability optimal, and comprehensive evaluation system of power forward-looking technology topic evaluation method to more scientifically and accurately support the power industry's technology strategy decision-making and innovation layout. SUMMARY

[0007] The purpose of the present application is to overcome the defects of the prior art and provide a neural topic modeling-based evaluation method for power forward-looking technology topics.

[0008] The purpose of the present application can be achieved by the following technical solutions:

[0009] In one aspect, the present application provides a neural topic modeling-based evaluation method for power forward-looking technology topics, comprising the following steps:

[0010] Obtain text data related to power technology within a predetermined time period;

[0011] Preprocess the text data, which includes text cleaning, denoising and time label extraction;

[0012] Use a neural topic modeling method to extract topics from the preprocessed text data to obtain a plurality of forward-looking power technology topics;

[0013] Construct an index system for identifying power forward-looking technology topics;

[0014] Obtain literature data corresponding to each forward-looking power technology topic, and calculate the scores of each index of each forward-looking power technology topic under the index system according to the literature data;

[0015] Based on the index scores, the entropy weight method is used to calculate the comprehensive evaluation score of each forward-looking power technology topic.

[0016] Furthermore, the text data includes a plurality of text samples, and the text samples include patent data or paper data related to electric power technology published within a preset time period.

[0017] Furthermore, the neural topic modeling method is used to extract topics from the pre-processed text data to obtain multiple forward-looking power technology topics, specifically including:

[0018] The pre-processed text data is vectorized and encoded, and a bag-of-words model is constructed to obtain a text representation matrix X = [x1, x2, ..., x n ], where x i represents the word frequency vector of the i-th text sample, and n is the total number of text samples;

[0019] The text representation matrix X is input into the pre-trained variational autoencoder, which includes an encoder and a decoder. The encoder is responsible for extracting the potential latent variable z from the input data. i , the decoder is responsible for the latent variable z i Reconstruct and generate predicted word distribution The encoder part learns the potential variable distribution of each sample through a multi-layer neural network. Each latent variable z i Normal distribution:

[0020]

[0021] Among them, μ i and σ i are the mean and standard deviation of the encoder output respectively;

[0022] Through cluster analysis, the latent variable z i Clustering is performed to obtain multiple forward-looking power technology themes.

[0023] Furthermore, the loss function of the variational autoencoder is:

[0024]

[0025] in, is the loss function, q(z i |x i ) is the variational distribution of the encoder output, indicating that given input x i Latent variable z i The conditional probability distribution of p(z) is the prior distribution of the latent variable z, and p(x i |z i) is a reconstruction distribution generated by the decoder, representing the probability distribution of the input x i when given the latent variable z i .

[0026] Further, the index system of the power foresight technology subject includes a plurality of first-level indicators, including technology novelty, technology sustainability, technology impact, and technology diffusion.

[0027] Further, the technology novelty includes a plurality of second-level indicators, including an average non-patent citation amount indicator, a citation amount proportion indicator, a technology keyword novelty indicator, and a cross-domain application evaluation indicator.

[0028] The average non-patent citation amount indicator calculation formula is:

[0029]

[0030] where R avg is the average non-patent citation amount, C i is the number of non-patent citations of the i-th document, and N is the total number of foresight power technology subject related documents.

[0031] The citation amount proportion indicator calculation formula is:

[0032]

[0033] where R ratio is the citation amount proportion, C science,i is the number of citations of the i-th document in the well-known literature database, and C total,i is the total number of citations of the i-th document.

[0034] The technology keyword novelty indicator calculation formula is:

[0035]

[0036] where R kw-new is the technology keyword novelty indicator, KW is the keyword set extracted from the current foresight power technology subject literature data, the extraction method is to select keywords from the literature data based on the TF-IDF algorithm, KW hist is the keyword set in the historical foresight power technology subject, the historical foresight power technology subject refers to a technology set consisting of all patents belonging to the classification number as the current foresight power technology subject, and the publication date being within 5 years before the first publication date of the current foresight power technology subject, the keyword extraction method is the same as the current subject, and |KW\KW hist | represents the number of novel keywords in the current keyword set that do not belong to the historical set.

[0037] The cross-field application evaluation index calculation formula is:

[0038]

[0039] Wherein, R cross is a cross-field application evaluation index, N sub represents the number of patents under the prospective power technology subject whose IPC main classification number is different from the main classification number of the current subject, N total represents the total number of patents under the prospective power technology subject.

[0040] Further, the technical persistence includes a heat persistence indicator and a time span indicator.

[0041] The heat persistence indicator calculation formula is:

[0042]

[0043] Wherein, S continuity is a heat persistence indicator, T is the total number of time windows divided in the life cycle of the technology subject, I t is a heat indicator variable of the tth time window, and the formula is:

[0044]

[0045] Wherein, H t is the heat value of the technology subject in the time window t, which is calculated based on the number of patents and papers related to the technology subject in the time period, and θ is a preset threshold value.

[0046] The time span indicator calculation formula is:

[0047] S time = T last -T first

[0048] Wherein, S time is a time span indicator, T first is the time when the technology subject is first recorded in the literature or patent, and T last is the time when the technology subject is last recorded in the literature or patent.

[0049] Further, the technical influence includes an academic influence indicator.

[0050] The academic influence indicator calculation formula is:

[0051]

[0052] Wherein, I academicN is the total number of papers related to the technical theme, and W p is the weight of the pth paper, W p = 1 if the paper is published in a core journal, otherwise W p = 0.

[0053] Further, the technology diffusion includes a citation propagation path indicator and a professional diffusion indicator.

[0054] The calculation formula of the citation propagation path indicator is:

[0055]

[0056] Wherein, R path is the citation propagation path indicator, N papers is the total number of papers related to the current technical theme, C p,q indicates whether the pth paper is cited by the qth paper, 1 if cited, otherwise 0, P p is the propagation level of the pth paper in the citation network, which is obtained by constructing a citation graph and calculating the longest propagation path of each paper.

[0057] The calculation formula of the professional diffusion indicator is:

[0058]

[0059] Wherein, PE is the professional diffusion indicator, N is the total number of patents related to the technical theme, and C i is the number of IPC classifications of the ith patent.

[0060] Further, the literature data corresponding to the forward-looking power technology theme includes patent data and paper data, the patent data includes patent title, abstract, applicant, inventor, application date, publication date, IPC classification number, citation information and priority information, etc. The paper data includes paper title, author, author institution, abstract, keyword, publication journal, publication time, citation information, fund support information and subject classification, etc.

[0061] Compared with the prior art, the present application has the following advantages:

[0062] (1) The present application introduces a neural network topic modeling method, especially uses a variational autoencoder (VAE) to model the hidden variables of the preprocessed power technology literature data, and combines clustering analysis to realize the identification of the forward-looking power technology theme. It overcomes the defects of the LDA model in the prior art in processing professional, complex terminology and semantic power technology text, such as coarse recognition granularity and weak semantic understanding ability, significantly improving the semantic expression ability and accuracy of technology topic modeling.

[0063] (2) The present invention constructs a multi-dimensional evaluation index system for electric power forward-looking technology themes, covering four dimensions: technology novelty, sustainability, influence and diffusion, and refines it into multiple secondary indicators including keyword novelty, cross-domain diffusion, citation path structure, time span, etc., and adopts the entropy weight method to weight and integrate various indicators, thereby achieving a quantitative characterization of the full life cycle and multi-level evolution characteristics of technology themes, effectively avoiding the problems of single evaluation dimension, isolated indicators, and difficulty in fully reflecting the evolution of technology value in existing methods.

[0064] (3) By combining time tag processing, deep semantic analysis and literature network structure modeling, the present invention enhances the structural perception and temporal tracking capabilities of the propagation paths and evolution trends of technical topics, which is conducive to the accurate identification and dynamic prediction of potential technical hotspots, and improves the scientificity and foresight of trend judgment.

[0065] (4) This invention is no longer limited to a single quantitative indicator (such as the number of patents, citation frequency, etc.) as the basis for judgment. Instead, it constructs an indicator system covering four dimensions: novelty, sustainability, influence, and diffusion. It is further refined into multiple sub-indicators such as "keyword novelty", "cross-domain integration", "citation path level", "time span", and "professional diffusion". This systematic quantitative evaluation system can reflect the evolution status and value potential of a technology theme from multiple perspectives and throughout the entire cycle, and is more in line with the actual needs of the complex technology ecology of the power industry.

[0066] (5) This paper introduces the structural modeling of literature citation networks into the calculation of indicators, such as the calculation of "citation propagation path hierarchy" and "professional diffusion index", so that the development process of each technical theme can be tracked not only in terms of quantity, but also in terms of structure, with clear propagation trajectory and cross-domain integration. This processing approach greatly enhances the structural interpretability of technology evaluation and provides a solid data foundation for subsequent trend prediction, technology early warning and strategic planning. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 is a flow chart of the method of the present invention;

[0068] Figure 2 This is the evaluation index model diagram of the present invention. DETAILED DESCRIPTION

[0069] 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 any creative efforts should fall within the scope of protection of the present invention.

[0070] Example 1:

[0071] This embodiment provides a method for evaluating power forward-looking technology topics based on neural topic modeling. Figure 1 As shown, the following steps are included:

[0072] Step S1: Acquire text data related to power technology within a preset time period;

[0073] Step S2: preprocessing the text data, including text cleaning, denoising and time tag extraction;

[0074] Step S3: Using a neural topic modeling method to extract topics from the preprocessed text data, and obtaining multiple forward-looking power technology topics;

[0075] Step S4: construct an indicator system for identifying forward-looking technology topics in the power industry;

[0076] Step S5: Obtain literature data corresponding to each forward-looking power technology topic, and calculate the index scores of each forward-looking power technology topic under the index system based on the literature data;

[0077] Step S6: Based on the indicator scores, the entropy weight method is used to calculate the comprehensive evaluation score of each forward-looking power technology topic.

[0078] The text data includes a plurality of text samples, and the text samples include patent data or paper data related to electric power technology published within a preset time period.

[0079] The neural topic modeling method is used to extract topics from the preprocessed text data, and multiple forward-looking power technology topics are obtained, including:

[0080] The preprocessed text data is vectorized and encoded, and a bag-of-words model is constructed to obtain the text representation matrix X = [x1, x2, ..., x n ], where x i represents the word frequency vector of the i-th text sample, and n is the total number of text samples;

[0081] The text representation matrix X is input into the pre-trained variational autoencoder, which includes an encoder and a decoder. The encoder is responsible for extracting the potential latent variable z from the input data. i, the decoder is responsible for the latent variable z i Reconstruct and generate predicted word distribution The encoder part learns the potential variable distribution of each sample through a multi-layer neural network. Each latent variable z i Normal distribution:

[0082]

[0083] Among them, μ i and σ i are the mean and standard deviation of the encoder output respectively;

[0084] Through cluster analysis, the latent variable z i Clustering is performed to obtain multiple forward-looking power technology themes.

[0085] This embodiment significantly improves the accuracy and flexibility of identifying forward-looking power technology topics by employing a neural topic modeling approach for text topic extraction, specifically introducing a variational autoencoder and cluster analysis. Specifically, the variational autoencoder extracts latent variables from the input text through the encoder, mapping the text into a low-dimensional latent space, preserving the text's core information without oversimplifying or losing important details. By performing cluster analysis in this low-dimensional space, this embodiment can identify potential forward-looking power technology topics that may span multiple technical fields. This process not only enhances the ability to capture the deep semantic structure of the text but also effectively avoids the reliance on specific keywords in traditional methods, improving the ability to identify complex power technology topics. Furthermore, the normal distribution assumption of the variational autoencoder ensures that the model's representation of latent variables more closely matches the distribution characteristics of the data, thereby improving the model's stability and accuracy. The implementation of this technical solution enables more comprehensive and accurate extraction of power technology topics, providing a solid foundation for subsequent multi-dimensional evaluation and ensuring the scientific and reliable nature of forward-looking power technology analysis.

[0086] The loss function of the variational autoencoder is:

[0087]

[0088] in, is the loss function, q(z i |x i ) is the variational distribution of the encoder output, indicating that given input x i Latent variable z i The conditional probability distribution of p(z) is the prior distribution of the latent variable z, and p(x i |z i ) is the reconstructed distribution generated by the decoder, which represents the given latent variable z i When generating input xi The probability distribution of .

[0089] The indicator system for the theme of forward-looking power technologies includes multiple first-level indicators, including technological novelty, technological sustainability, technological influence, and technological diffusion.

[0090] Technological novelty includes multiple secondary indicators, including the average number of non-patent citations per article, the citation ratio, the technical keyword novelty, and the cross-domain application evaluation index.

[0091] The formula for calculating the average number of non-patent citations per article is:

[0092]

[0093] Among them, R avg is the average number of non-patent citations per article, C i is the number of non-patent citations of the i-th document, and N is the total number of documents related to the forward-looking power technology topic;

[0094] The calculation formula for the citation ratio index is:

[0095]

[0096] Among them, R ratio is the citation ratio, C science,i is the number of citations of the i-th document published in a well-known literature database, C total,i is the total number of citations of the i-th document;

[0097] The calculation formula for the technical keyword novelty index is:

[0098]

[0099] Among them, R kw-new is the technical keyword novelty index, KW is the keyword set extracted from the literature data corresponding to the current forward-looking power technology theme, and the extraction method is to select keywords from the literature data based on the TF-IDF algorithm. KW hist The keyword collection in the historical prospective power technology theme. The historical prospective power technology theme refers to the technology collection consisting of all patents that belong to the same classification number as the current prospective power technology theme and whose publication date is within 5 years before the first publication of the current prospective power technology theme. The keyword extraction method is the same as the current theme. |KW\KW hist | represents the number of novel keywords in the current keyword set that do not belong to the historical set;

[0100] The calculation formula for cross-domain application evaluation indicators is:

[0101]

[0102] wherein R cross is the cross-domain application evaluation indicator, N sub represents the number of patents under the prospective power technology subject matter whose IPC main classification number is different from the main classification number of the current subject matter, N total represents the total number of patents under the prospective power technology subject matter.

[0103] The technical persistence includes a heat persistence indicator and a time span indicator.

[0104] The heat persistence indicator calculation formula is:

[0105]

[0106] wherein S continuity is the heat persistence indicator, T is the total number of time windows divided in the life cycle of the technology subject matter, I t is the heat indicator variable of the tth time window, and the formula is:

[0107]

[0108] wherein H t is the heat value of the technology subject matter in the time window t, which is calculated based on the number of patents and papers related to the technology subject matter in the time period, and θ is a preset threshold value.

[0109] The time span indicator calculation formula is:

[0110] S time = T last - T first

[0111] wherein S time is the time span indicator, T first is the time when the technology subject matter is first recorded in the literature or patent, and T last is the time when the technology subject matter is last recorded in the literature or patent.

[0112] The technical influence includes an academic influence indicator.

[0113] The academic influence indicator calculation formula is:

[0114]

[0115] wherein I academic is the academic influence indicator, N is the total number of papers related to the technology subject matter, W p is the weight of the pth paper, W p = 1 if the paper is published in a core journal, otherwise W p = 0.

[0116] The technical diffusion includes a citation propagation path indicator and a professional diffusion indicator.

[0117] The calculation formula of the citation propagation path indicator is:

[0118]

[0119] Wherein, R path is the citation propagation path indicator, N papers is the total number of papers related to the current technical subject, C p,q represents whether the pth paper is cited by the qth paper, 1 if cited, otherwise 0, P p is the propagation level of the pth paper in the citation network, which is obtained by constructing the citation graph and calculating the longest propagation path of each paper;

[0120] The calculation formula of the professional diffusion indicator is:

[0121]

[0122] Wherein, PE is the professional diffusion indicator, N is the total number of patents related to the technical subject, C i is the number of IPC classifications of the ith patent.

[0123] The literature data corresponding to the forward-looking power technology subject includes patent data and paper data. The patent data includes patent title, abstract, applicant, embodiment person, application date, publication date, IPC classification number, citation information and priority information, etc. The paper data includes paper title, author, author institution, abstract, keyword, publication journal, publication time, citation information, fund support information and subject classification, etc.

[0124] Example 2:

[0125] The embodiment provides a power forward-looking technology subject evaluation system based on neural topic modeling, as shown in Figure 2 , which comprises:

[0126] A text data acquisition module is used to collect patent data, paper data and other technical literature data related to power technology;

[0127] A data preprocessing module is used to clean and denoise the collected text data, and extract time labels to ensure high-quality input of data;

[0128] A topic extraction module uses a neural topic modeling method (such as variational autoencoder and clustering analysis) to model the preprocessed text data, automatically identifies multiple forward-looking power technology subjects and their corresponding keyword descriptions;

[0129] The indicator system construction module is used to build a comprehensive indicator system based on the characteristics of the technology theme, covering evaluation criteria in multiple dimensions such as the strength, stability, emerging degree and cutting-edge degree of the technology theme;

[0130] The indicator score calculation module calculates the scores of each forward-looking power technology topic under various indicators based on information extracted from literature data, providing a quantitative basis for subsequent evaluation;

[0131] Comprehensive evaluation module, which comprehensively evaluates each forward-looking power technology topic based on the entropy weight method, and ultimately obtains a comprehensive score for each technology topic to assist in decision-making;

[0132] The result output module is used to output the evaluation results, including the comprehensive evaluation score of each technical topic and its performance in different dimensions.

[0133] On the other hand, this embodiment provides an electronic device including a memory and a processor, wherein a computer program is stored in the memory, and when the processor executes the program, it implements a method for evaluating electric power forward-looking technology topics based on neural topic modeling as described above.

[0134] This embodiment also provides a computer-readable storage medium having a computer program stored thereon. When the program is executed by a processor, it implements a method for evaluating electric power forward-looking technology topics based on neural topic modeling as described above.

[0135] 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.

[0136] 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 evaluating power forward-looking technology topics based on neural topic modeling, characterized in that: The following steps are involved: Obtain text data related to power technology within a preset time period; Preprocessing the text data, including text cleaning, denoising, and time tag extraction; Using a neural topic modeling method to extract topics from the preprocessed text data, and obtaining multiple forward-looking power technology topics; Build an indicator system for identifying forward-looking technology topics in the power industry; Obtaining literature data corresponding to each forward-looking power technology topic, and calculating the scores of various indicators of each forward-looking power technology topic under the indicator system based on the literature data; Based on the index scores, the entropy weight method is used to calculate the comprehensive evaluation score of each forward-looking power technology topic.

2. The method for evaluating electric power forward-looking technology topics based on neural topic modeling according to claim 1 is characterized in that: The text data includes a plurality of text samples, and the text samples include patent data or paper data related to electric power technology published within a preset time period.

3. The method for evaluating electric power prospective technology topics based on neural topic modeling according to claim 1 is characterized in that: The neural topic modeling method is used to extract topics from the pre-processed text data to obtain multiple forward-looking power technology topics, including: The pre-processed text data is vectorized and encoded, and a bag-of-words model is constructed to obtain a text representation matrix X = [x1, x2, ..., x n ], where x i represents the word frequency vector of the i-th text sample, and n is the total number of text samples; The text representation matrix X is input into the pre-trained variational autoencoder, which includes an encoder and a decoder. The encoder is responsible for extracting the potential latent variable z from the input data. i , the decoder is responsible for the latent variable z i Reconstruct and generate predicted word distribution The encoder part learns the potential variable distribution of each sample through a multi-layer neural network. Each latent variable z i Normal distribution: Among them, μ i and σ i are the mean and standard deviation of the encoder output respectively; Through cluster analysis, the latent variable z i Clustering is performed to obtain multiple forward-looking power technology themes.

4. The method for evaluating electric power prospective technology topics based on neural topic modeling according to claim 3 is characterized in that: The loss function of the variational autoencoder is: in, is the loss function, q(z i |x i ) is the variational distribution of the encoder output, indicating that given input x i Latent variable z i The conditional probability distribution of p(z) is the prior distribution of the latent variable z, and p(x i |z i ) is the reconstructed distribution generated by the decoder, which represents the given latent variable z i When generating input x i The probability distribution of .

5. The method for evaluating electric power prospective technology topics based on neural topic modeling according to claim 1 is characterized in that: The index system for the forward-looking technology theme of electricity includes multiple first-level indicators, including technological novelty, technological sustainability, technological influence, and technological diffusion.

6. The method for evaluating electric power prospective technology topics based on neural topic modeling according to claim 5, characterized in that: The technical novelty includes multiple secondary indicators, including the average non-patent citation index, the citation ratio index, the technical keyword novelty index and the cross-domain application evaluation index; The calculation formula for the non-patent citation index per article is: Among them, R avg is the average number of non-patent citations per article, C i is the number of non-patent citations of the i-th document, and N is the total number of documents related to the forward-looking power technology topic; The calculation formula for the citation ratio index is: Among them, R ratio is the citation ratio, C science,i is the number of citations of the i-th document published in a well-known literature database, C total,i is the total number of citations of the i-th document; The calculation formula for the novelty index of the technical keyword is: Among them, R kw-new is the technical keyword novelty index, KW is the keyword set extracted from the literature data corresponding to the current forward-looking power technology theme, and the extraction method is to select keywords from the literature data based on the TF-IDF algorithm. KW hist The keyword collection in the historical prospective electric power technology theme refers to the technology collection consisting of all patents that belong to the same classification number as the current prospective electric power technology theme and whose publication date is within 5 years before the first publication of the current prospective electric power technology theme. The keyword extraction method is the same as the current theme. |KW\KW hist | represents the number of novel keywords in the current keyword set that do not belong to the historical set; The calculation formula for the cross-domain application evaluation index is: Among them, R cross For cross-domain application evaluation indicators, N sub Indicates the number of patents under the forward-looking power technology theme whose IPC main classification number is different from the main classification number of the current theme, N total Indicates the total number of patents under the forward-looking power technology theme.

7. The method for evaluating electric power prospective technology topics based on neural topic modeling according to claim 5, characterized in that: The technical sustainability includes a heat persistence index and a time span index; The calculation formula of the heat persistence index is: Among them, S continuity is the heat persistence indicator, T is the total number of time windows divided in the life cycle of the technology topic, and I t is the heat indicator variable of the t-th time window, and the formula is: Among them, H t is the popularity value of the technical topic in the time window t, which is calculated based on the number of patents and papers related to the technical topic in this time period, and θ is the preset threshold; The calculation formula of the time span indicator is: S time =T last -T first Among them, S time is the time span indicator, T first T is the time when the technical subject is first recorded in literature or patents. last The time when the technical subject was last recorded in a document or patent.

8. The method for evaluating electric power forward-looking technology topics based on neural topic modeling according to claim 5, characterized in that: The technical influence includes academic influence indicators; The calculation formula for the academic influence index is: Among them, I academic is the academic influence indicator, N is the total number of papers related to the technical topic, and W p is the weight of the pth paper. If the paper is published in a core journal, W p =1, otherwise W p =0.

9. The method for evaluating electric power prospective technology topics based on neural topic modeling according to claim 5, characterized in that: The technology diffusion includes citation propagation path indicators and professional diffusion indicators; The calculation formula for the citation propagation path index is: Among them, R path is the citation propagation path indicator, N papers is the total number of papers related to the current technical topic, C p,q Indicates whether the p-th paper is cited by the q-th paper. If so, it is 1, otherwise 0. p is the diffusion level of the p-th paper in the citation network, obtained by constructing a citation graph and calculating the longest diffusion path of each paper; The calculation formula of the professional diffusion index is: Among them, PE is the professional diffusion index, N is the total number of patents related to the technical subject, C i is the number of IPC classifications of the i-th patent.

10. The method for evaluating electric power prospective technology topics based on neural topic modeling according to claim 1, characterized in that: The literature data corresponding to the forward-looking power technology topic includes patent data and paper data.

Images