Prospective power technology prediction method based on topic extraction and time modeling

Abstract

The invention provides a prospective power technology prediction method based on topic extraction and time modeling. According to the method, power related text data is preprocessed, topic information of a text is extracted, and a power technology topic set is generated based on UMAP algorithm dimensionality reduction and HDBSCAN clustering analysis. Then, keywords are extracted through a TF-IDF method, and a theme and keyword description set is constructed; and in combination with a time label, constructing a time sequence of the power technology theme, eliminating short-term fluctuation and calculating a change rate by utilizing smoothing processing, and further modeling an evolution trend of the power technology theme. And according to a trend modeling result, screening out a prospective power technology theme and an evolution trend thereof, thereby realizing prediction of a future development direction of the power technology. The method can help to predict and identify the technical innovation trend in the power field, and improves the scientificity and foresight of technical decision support.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of text topic recognition and trend prediction, and particularly relates to a forward-looking power technology prediction method based on topic extraction and time modeling. BACKGROUND

[0002] With the accelerated development of new power systems and energy transformation, power technology is showing an increasingly diversified and intelligent development trend. In order to effectively support technology innovation management and strategy formulation, a method is urgently needed to timely mine and predict potential frontier technologies from massive technical literature. At present, the analysis of power technology evolution trends mainly relies on artificial expert judgment or retrospective analysis based on traditional statistical methods, which has the problems of low efficiency, strong subjectivity, and difficulty in covering a wide range of technical fields.

[0003] In recent years, the development of natural language processing (NLP) technology and machine learning methods has provided new possibilities for intelligent analysis of technical literature. Among them, topic modeling, as a method that can identify implicit topics from unstructured text, has been widely applied in the field of scientific and technical intelligence analysis. However, existing topic modeling methods mainly focus on clustering or labeling of static text, and lack the ability to systematically model and dynamically predict the evolution trend of technical topics over time.

[0004] In the prior art, Chinese patent CN113988087A discloses a method and device for calculating and predicting the trend of technical topics, which includes: obtaining a plurality of paper text data, preprocessing the 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, extracting technical topics using an LDA topic model on the weighted multi-feature fields; based on the extracted technical topics, designing a plurality of index measurement methods to measure the advancement of technical topics, and calculating the index values of the technical topic indexes, wherein the technical topic indexes include: intensity, stability, emerging degree and frontier degree; based on the index values, using a Logistic model to predict the development trend of the technical topics, and performing comparative analysis between the technical topics. This method extracts multiple features of the text, clearly divides the topics, and forms a general method for calculating and predicting the trend of technical topics.

[0005] However, the method still has the following shortcomings: first, the LDA topic model used by it is based on the bag-of-words assumption, which is difficult to effectively capture the context semantic association between words, resulting in limited precision and expression ability of topic extraction, and it is difficult to accurately depict the semantic complex power technology topic; second, the method indirectly reflects the evolution state of the topic through the preset evaluation index (such as intensity, emerging degree, etc.), which lacks direct modeling of the frequency change of the topic itself, and it is difficult to truly depict the evolution process of the technology topic in different time periods; third, the time modeling process mainly relies on the Logistic fitting of a single numerical index, and fails to fully mine the time sequence structure information implied in the text data, making it difficult to flexibly respond to the rapid changes of emerging technologies.

[0006] Therefore, there is an urgent need for a method that combines topic extraction and time modeling, which can automatically mine evolving topics from multi-source heterogeneous power technology literature and identify forward-looking power technologies based on their dynamic change trends, providing data support for technology prediction, policy making and enterprise R&D decision making. SUMMARY

[0007] The purpose of the present application is to overcome the shortcomings of the prior art and provide a forward-looking power technology prediction method based on topic extraction and time modeling.

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

[0009] The present application provides a forward-looking power technology prediction method based on topic extraction and time modeling, comprising the following steps:

[0010] Obtain text data related to power technology;

[0011] Perform data preprocessing on the text data, which includes text cleaning, denoising and extraction of time labels;

[0012] Extract topics from the text data using the BERTopic method to obtain multiple power technology topics;

[0013] According to the time labels extracted from the text data, combine the power technology topics to construct a topic time series, which is used to describe the change of the frequency of each power technology topic in different time windows of text data;

[0014] Trend modeling of the topic time series, including smoothing and rate of change analysis;

[0015] Based on the results of the trend modeling, determine the forward-looking power technology topic and output the forward-looking power technology topic and its corresponding evolution trend.

[0016] Further, the text data includes patent data and paper data.

[0017] Further, the BERTopic method is used to extract topics from the text data to obtain a plurality of power technology topics, specifically including:

[0018] The pre-trained BERT model is used for text embedding of the text data to generate high-dimensional semantic vector representation of the text data;

[0019] The UMAP algorithm is used for dimension reduction processing of the high-dimensional semantic vector representation to obtain low-dimensional representation data;

[0020] The HDBSCAN clustering algorithm is used for clustering analysis of the low-dimensional representation data to generate a preliminary topic set;

[0021] The TF-IDF method is used to extract keywords from the text included in each topic in the preliminary topic set to generate the plurality of power technology topics and corresponding keyword descriptions;

[0022] The plurality of power technology topics are mapped with the original text data to form topic attribution labels, which are used to indicate the corresponding power technology topic of each text data.

[0023] Further, the UMAP algorithm is used for dimension reduction processing of the semantic vector representation to obtain low-dimensional representation data, specifically including:

[0024] The high-dimensional semantic vector representation corresponding to each text data is obtained, denoted as where D is the dimension of the high-dimensional semantic space, and i represents the text data number;

[0025] Based on the high-dimensional semantic vector, an adjacency graph G=(V,E,P) is constructed, where: V is a node set, corresponding to all text semantic vectors; E is an edge set, representing a pair of adjacent texts; P={p ij} is the conditional probability distribution between node pairs in the high-dimensional space, defined as follows:

[0026]

[0027] where p ij represents the conditional similarity probability of text data i and text data j in the high-dimensional space, p i represents the minimum distance of text data i and its nearest neighbor, used to control the local connection strength, and s i is a local scale factor;

[0028] The UMAP algorithm is used to optimize the mapping of the adjacency graph G to construct the embedding distribution Q={q ijThe target low-dimensional space is a Euclidean space with dimension d where d << D, for preserving the local adjacency relationship of the text semantic structure.

[0029] By minimizing the cross-entropy loss function between the high-dimensional distribution P and the low-dimensional distribution Q:

[0030]

[0031] where q ij is the similarity between the text data i and the text data j in the target low-dimensional space;

[0032] The low-dimensional representation data {y i} is back-propagated and gradient-optimized, and finally the low-dimensional representation data of each text data in the target low-dimensional space is obtained

[0033] Further, the similarity q ij between the text data i and the text data j in the target low-dimensional space is:

[0034]

[0035] where, is the low-dimensional representation data of the text data i in the low-dimensional space, and a and b are hyperparameters in the UMAP model for adjusting the distribution curvature.

[0036] Further, the low-dimensional representation data is clustered and analyzed by using the HDBSCAN clustering algorithm to generate a preliminary theme set, specifically including:

[0037] Obtain the representation vector of each text data in the low-dimensional space;

[0038] Set the minimum cluster size and minimum sample number of the HDBSCAN clustering algorithm and other parameters;

[0039] Based on the low-dimensional representation data, the distance relationship between samples is calculated, and a density map between samples is constructed;

[0040] Using the density map, identify the areas with higher local density to form a plurality of clustering clusters as a preliminary theme set;

[0041] Each text data is assigned a corresponding theme label according to the clustering cluster to which it belongs, and the data not belonging to any clustering cluster is marked as a noise point;

[0042] Obtain a preliminary theme set T = {T1, T2,..., T k}, where each T jrepresents an electric power technology topic identified by the density clustering algorithm, and k is the total number of identified topics.

[0043] Furthermore, the TF-IDF method is used to extract keywords from the text contained in each topic in the preliminary topic set to generate the multiple power technology topics and their corresponding keyword descriptions, specifically including:

[0044] For the text data set corresponding to each topic in the preliminary topic set, perform word segmentation, remove stop words and low-frequency words, and obtain a standardized word set;

[0045] Based on the term set and the text set, construct a term-document matrix M, where the matrix M is a sparse matrix, where each row corresponds to a term, each column corresponds to a text, and the matrix element M(i, j) represents the number of occurrences of term i in text data j;

[0046] Based on the term-document matrix, the following TF-IDF calculation formula is used to perform weighted calculation on each term in each topic:

[0047] TF-IDF(t i ,d j )=TF(t i ,d j )×IDF(t i )

[0048] in, Represents the term t i In text data d j The word frequency in f i,j For term t i In text data d j The number of times it appears in k f k,j For text data d j The total number of occurrences of all terms in ; Represents the term t i The inverse document frequency, N is the total number of all text data, n i To contain the term t i The amount of text data;

[0049] Sort all terms according to the TF-IDF value, and select the terms with the highest scores as representative keywords of the current topic;

[0050] Each topic is combined with its corresponding representative keyword to generate multiple power technology topics and their keyword description sets for subsequent time modeling and trend analysis.

[0051] Further, the time label extracted from the text data is combined with the power technology subject to construct a subject time sequence, specifically including:

[0052] The power technology subject corresponding to each piece of text data is associated with its time label to form a subject-time pair;

[0053] The time granularity is set to year or quarter, and a unified time window set {t1, t2,..., t n} is constructed in chronological order;

[0054] For each power technology subject T i , the number of text data appearing in each time window t j is counted to construct a subject time frequency sequence, defined as follows:

[0055]

[0056] Where f i,j represents the frequency of power technology subject T i in time window t j , N is the total number of text data, L k represents the power technology subject label of the kth text data, Y k represents the time label of the kth text data, and δ is a Boolean indicator function, 1 if the condition is true, otherwise 0;

[0057] The time frequency sequence of each power technology subject is represented as:

[0058] S i ={f i,1 ,f i,2 ,...,f i,n}

[0059] Where S i is the time sequence of power technology subject T i ,

[0060] The subject time sequence of all power technology subjects {S1, S2,..., S k} is constructed.

[0061] Further, the subject time sequence is trend modeled, specifically including:

[0062] The time sequence S i ={f i,1 ,f i,2 ,...,f i,n} of each power technology subject is smoothed to eliminate the interference of short-term fluctuations on trend identification, and a moving average method is used to calculate the smoothed time sequence The formula is:

[0063]

[0064] in, Represents the power technology topic T in the jth time window i The smoothed frequency value of , w is the time window width;

[0065] Based on the smoothed series, calculate the power technology theme T i The rate of change r in each time window i,j , which is used to reflect the growth trend of technology topics, and its calculation formula is:

[0066]

[0067] Among them, ∈ is a small constant, r i,j Indicates the power technology theme T i The relative rate of change at time j compared to the previous time;

[0068] Calculate T for each power technology topic i The average rate of change over the most recent m consecutive time windows and the standard deviation of the rate of change σ i , the formula is as follows:

[0069]

[0070] Among them, m is the preset parameter, n is the total number of time windows, Indicates the power technology theme T i The average evolution rate in the last m time windows, σ i Indicates the power technology theme T i The standard deviation of the evolution rate over the last m time windows.

[0071] Furthermore, determining a forward-looking power technology theme based on the result of the trend modeling and outputting the forward-looking power technology theme and its corresponding evolution trend specifically includes:

[0072] According to the preset forward-looking judgment threshold θ r ,θ σ , according to each power technology theme T i The average evolution rate in the last m time windows and standard deviation σ i Topics that meet the following criteria are selected as forward-looking power technology topics:

[0073] And σ i <θ σ

[0074] Output all subjects that meet the conditions, their time series curves and keyword descriptions as prediction results.

[0075] Compared with the prior art, the present application has the following advantages:

[0076] (1) The present application extracts the topics of the text data in the power technology field by introducing the BERTopic method. Compared with the traditional LDA topic model, BERTopic can effectively combine the pre-trained BERT model for text embedding, generate high-dimensional semantic vector representation, and process dimension reduction through UMAP algorithm and identify power technology topics through HDBSCAN clustering algorithm. This technical means significantly improves the semantic precision and accuracy of topic extraction, which can more finely capture the potential information in the power technology topics, thereby avoiding the limitation of the bag-of-words assumption in traditional methods, which cannot reflect the contextual semantic relationship. This accurate topic modeling provides a more reliable data basis for subsequent power technology prediction and trend analysis.

[0077] (2) The present application combines the time labels extracted from the text data with the power technology topics to construct a topic time series, which describes the change of the frequency of the power technology topics in the text data in different periods. Compared with the technical index analysis in the prior art which relies on fixed time periods, the present application can dynamically construct the evolution trajectory of the technical topics according to the time information of the text. This technical means can effectively capture the temporal characteristics of power technology, providing more realistic and flexible data support for subsequent trend modeling and forward-looking technology prediction, and enhancing the accurate grasp of the evolution trend of power technology.

[0078] (3) When modeling the trend of the topic time series, the present application not only uses smoothing method, but also introduces change rate analysis to further finely depict the evolution trend of the technical topics. Unlike the prior art which only uses Logistic model to predict the trend of a single technical index, the present application can comprehensively and meticulously evaluate the evolution speed and potential of the power technology topics by modeling the frequency change of each technical topic. This innovative trend modeling method effectively improves the accuracy and reliability of the forward-looking power technology prediction, ensuring that potential emerging technologies and frontier technologies can be discovered in a timely manner.

[0079] (4)The present application can not only identify the historical trend of power technology theme, but also effectively predict future technology trends by introducing dynamic modeling technology combined with time series change rate analysis. Unlike the existing technology based on static text theme extraction and single technology index prediction method, the technical solution of the present application can dynamically adapt to different stages of technology development cycle by flexibly responding to the rapid changes in the field of power technology, providing more forward-looking technology prediction support for decision makers in the field of power. This ability enables the present application to quickly and accurately grasp and predict the future direction of technology in the face of rapidly changing power technology innovation.

[0080] (5)The present application extracts keywords for each power technology theme based on the TF-IDF method, and then generates a set of power technology themes and their keyword descriptions. This technical means can effectively extract the core ideas and key content of each technology theme, providing clear technology theme map for power technology researchers and decision makers. This fine-grained theme description not only helps technology researchers better understand the technology content, but also provides solid semantic support for technology trend prediction. BRIEF DESCRIPTION OF DRAWINGS

[0081] Figure 1 The prediction method flowchart of the embodiment of the present application is shown in

[0082] Figure 2 The theme extraction method flowchart of the embodiment of the present application is shown in

[0083] Figure 3 The prediction system module diagram of the embodiment of the present application is shown in DETAILED DESCRIPTION

[0084] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor shall fall within the scope of protection of the present application.

[0085] Embodiment 1:

[0086] The present embodiment provides a forward-looking power technology prediction method based on theme extraction and time modeling, as shown in Figure 1 The steps are as follows:

[0087] Step S1: Obtain text data related to power technology;

[0088] Step S2: Data preprocessing of text data, including text cleaning, denoising and time label extraction;

[0089] Step S3: extracting topics from the text data using the BERTopic method to obtain a plurality of power technology topics;

[0090] Step S4: constructing topic time series according to the time labels extracted from the text data in combination with the power technology topics, the topic time series being used to describe the change in the frequency of occurrence of each power technology topic in the text data in different time windows;

[0091] Step S5: trend modeling of the topic time series, including smoothing processing and rate of change analysis;

[0092] Step S6: determining the forward-looking power technology topics based on the results of the trend modeling and outputting the forward-looking power technology topics and their corresponding evolution trends.

[0093] The text data includes patent data and paper data.

[0094] The BERTopic method is used to extract topics from the text data to obtain a plurality of power technology topics, such as Figure 2 as shown, specifically including:

[0095] Step S301: text embedding of the text data using a pre-trained BERT model to generate high-dimensional semantic vector representation of the text data;

[0096] Step S302: dimensionality reduction processing of the high-dimensional semantic vector representation based on the UMAP algorithm to obtain low-dimensional representation data;

[0097] Step S303: clustering analysis of the low-dimensional representation data using the HDBSCAN clustering algorithm to generate a preliminary topic set;

[0098] Step S304: keyword extraction of the text contained in each topic in the preliminary topic set based on the TF-IDF method to generate a plurality of power technology topics and their corresponding keyword descriptions;

[0099] Step S305: mapping the plurality of power technology topics with the original text data to form topic attribution labels, the topic attribution labels being used to indicate the power technology topics corresponding to each piece of text data.

[0100] In this embodiment, the pre-trained BERT model is used for text embedding, aiming to convert text data into high-dimensional semantic vector representation, so as to capture information in the text from a more deep semantic dimension. The BERT model can effectively capture the complex relationship between words by modeling the context through deep learning, and provide more accurate semantic understanding, which can avoid the understanding error caused by the traditional word frequency or simple rule processing method, especially for the professional power technology field, which can extract more accurate information. The UMAP algorithm is used to reduce the dimension of the high-dimensional semantic vector, which is to reduce the complexity of data processing and speed up the subsequent calculation process, while trying to preserve the semantic information of the text. Dimension reduction can reduce the computational burden and make the subsequent clustering analysis more efficient, which helps to improve the response speed and processing capacity of the whole system. The low-dimensional representation after dimension reduction helps to more clearly show the structural characteristics of the text data, especially for the topic mining in large-scale data sets. Through the HDBSCAN algorithm for clustering analysis, multiple power technology topics can be automatically identified and classified according to the similarity of the text. Compared with traditional topic modeling methods, HDBSCAN does not rely on the number of pre-set topics, and can more flexibly extract natural clusters from data, which helps to discover potential emerging technology fields or trends, thereby improving the accuracy and innovation of topic analysis. Combined with the TF-IDF method, the keywords of each topic are extracted, which can help to further refine the core features of each topic, making the extracted topics more representative. Through these keywords, the key content and technical development direction of each power technology topic can be more clearly understood, which helps the subsequent technology trend prediction.

[0101] Based on the UMAP algorithm, the semantic vector representation is reduced in dimension to obtain low-dimensional representation data, specifically including:

[0102] Get the high-dimensional semantic vector representation corresponding to each piece of text data, denoted as Where D is the dimension of the high-dimensional semantic space, and i represents the text data number;

[0103] Based on the high-dimensional semantic vector, an adjacency graph G=(V,E,P) is constructed, where: V is the node set, corresponding to all text semantic vectors; E is the edge set, representing the near neighbor text pair; P={p ij} is the conditional probability distribution between node pairs in the high-dimensional space, defined as follows:

[0104]

[0105] Where, p ij represents the conditional similarity probability of text data i and text data j in the high-dimensional space, and p iRepresents the minimum distance between text data i and its nearest neighbor, used to control the local connection strength, σ i is the local scale factor;

[0106] The UMAP algorithm is used to optimize the mapping of the adjacency graph G and construct the embedding distribution Q in the target low-dimensional space = {q ij}, the target low-dimensional space is the Euclidean space of dimension d Where d<<D is used to preserve the local adjacency relationship of the text semantic structure;

[0107] By minimizing the cross entropy loss function between the high-dimensional distribution P and the low-dimensional distribution Q:

[0108]

[0109] Among them, q ij is the similarity between text data i and text data j in the target low-dimensional space;

[0110] For low-dimensional representation data {y i} Perform back propagation and gradient optimization to finally obtain the low-dimensional representation data of each text data in the target low-dimensional space

[0111] The similarity q between text data i and text data j in the target low-dimensional space ij for:

[0112]

[0113] in, is the low-dimensional representation data of text data i in the low-dimensional space, and a and b are the hyperparameters used to adjust the distribution curvature in the UMAP model.

[0114] In this paper, the UMAP algorithm is used to reduce the dimensionality of high-dimensional semantic vector representations. The purpose is to map high-dimensional data to a low-dimensional space by preserving the local adjacency relationship between text data, so as to perform subsequent clustering and analysis more efficiently. First, an adjacency graph is constructed based on the high-dimensional semantic vector to ensure that the similarity between text data can be effectively preserved during the dimensionality reduction process. Specifically, the conditional probability distribution p ij Parameter ρ is used to quantify the similarity between texts and control the strength of local connections. i and σ i The adjacency graph can be adjusted based on the relationships between texts, ensuring that similar text data maintains relatively close relationships after dimensionality reduction. This approach effectively ensures that the complex semantic information of texts in high-dimensional space is not lost during the dimensionality reduction process, thereby preserving the structural characteristics of the original data.

[0115] The UMAP algorithm optimizes the dimensionality reduction process by minimizing the cross-entropy loss function between high-dimensional and low-dimensional spaces, enabling precise mapping to the target low-dimensional space. This process reduces errors in the low-dimensional representation while maintaining similarity between texts. The similarity calculation for text data in the low-dimensional space uses hyperparameters a and b, which adjust the distribution curvature, further enhancing the geometric consistency of the data representation after dimensionality reduction. By optimizing these hyperparameters, similar text data in the low-dimensional space is clustered more closely while maintaining a sufficient distance from dissimilar text. Overall, UMAP dimensionality reduction not only reduces computational complexity but also effectively preserves the semantic information of the text data, providing more accurate and efficient input data for subsequent clustering and topic extraction. This approach can provide improved efficiency and accuracy in processing large-scale text data.

[0116] The HDBSCAN clustering algorithm is used to perform cluster analysis on the low-dimensional representation data to generate a preliminary topic set, including:

[0117] Get the representation vector of each text data in low-dimensional space;

[0118] Set parameters such as the minimum cluster size and minimum number of samples for the HDBSCAN clustering algorithm;

[0119] Based on the low-dimensional representation data, the distance relationship between samples is calculated and the density map between samples is constructed;

[0120] Using the density map, we identify areas with high local density and form several clusters as preliminary topic sets.

[0121] Each piece of text data is assigned a corresponding topic label according to the cluster to which it belongs, and the data that does not belong to any cluster is marked as noise points;

[0122] Get a preliminary topic set T = {T1, T2, ..., T k}, where each T j represents an electric power technology topic identified by the density clustering algorithm, and k is the total number of identified topics.

[0123] Based on the TF-IDF method, keywords are extracted from the text contained in each topic in the preliminary topic set to generate multiple power technology topics and their corresponding keyword descriptions, including:

[0124] For the text data set corresponding to each topic in the preliminary topic set, perform word segmentation, remove stop words and low-frequency words, and obtain a standardized word set;

[0125] Based on the term set and the text set, construct a term-document matrix M. The matrix M is a sparse matrix, where each row corresponds to a term and each column corresponds to a text. The matrix element M(i, j) represents the number of occurrences of term i in text data j.

[0126] Based on the term-document matrix, the following TF-IDF calculation formula is used to perform weighted calculation on each term in each topic:

[0127] TF-IDF(t i ,d j )=TF(t i ,d j )×IDF(t i )

[0128] in, Represents the term t i In text data d j The word frequency in f i,j For term t i In text data d j The number of times it appears in k f k,j For text data d j The total number of occurrences of all terms in ; Represents the term t i The inverse document frequency, N is the total number of all text data, n i To contain the term t i The amount of text data;

[0129] Sort all terms according to TF-IDF values ​​and select the terms with the highest scores as representative keywords for the current topic;

[0130] Each topic is combined with its corresponding representative keyword to generate multiple power technology topics and their keyword description sets for subsequent time modeling and trend analysis.

[0131] By using the HDBSCAN clustering algorithm to perform cluster analysis on low-dimensional representations of data, we can automatically identify potential technical topics within a large volume of power technology text data and adaptively form multiple topic clusters based on the data's natural structure, while effectively processing noisy data. Next, we use the TF-IDF method to extract keywords from the text within each topic, accurately identifying the core content of each topic and eliminating irrelevant information, thereby ensuring the representativeness and accuracy of each technical topic. This step ensures the extraction of high-value topic information from complex power technology literature. Keyword extraction makes the topics clearer and more understandable, facilitating subsequent trend analysis and technology forecasting.

[0132] According to the time label extracted from the text data, combined with the power technology theme, a theme time sequence is constructed, specifically including:

[0133] The power technology theme corresponding to each piece of text data is associated with its time label to form a theme-time pair;

[0134] The time granularity is set to year or quarter, and a unified time window set {t1, t2,..., t n} is constructed in chronological order;

[0135] For each power technology theme T i , the number of text data appearing in each time window t j is counted, and a theme time frequency sequence is constructed, defined as follows:

[0136]

[0137] Where f i,j represents the frequency of the power technology theme T i in the time window t j , N is the total number of text data, L k represents the power technology theme label of the kth text data, Y k represents the time label of the kth text data, and δ is a Boolean indicator function, 1 if the condition is true, otherwise 0;

[0138] The time frequency sequence of each power technology theme is represented as:

[0139] S i ={f i,1 ,f i,2 ,...,f i,n}

[0140] Where S i is the time sequence of the power technology theme T i ,

[0141] The theme time sequence of all power technology themes {S1, S2,..., S k} is constructed.

[0142] Trend modeling is performed on the theme time sequence, specifically including:

[0143] The time sequence S i ={f i,1 ,f i,2 ,...,f i,n} of each power technology theme is smoothed to eliminate the interference of short-term fluctuations on trend identification, and the moving average method is used to calculate the smoothed time sequence The formula is:

[0144]

[0145] wherein, represents the smoothed frequency value of the power technology theme T i in the jth time window, and w is the time window width;

[0146] On the basis of the smoothed sequence, the change rate r i of the power technology theme T i,j in each time window is calculated, which is used to reflect the growth trend of the technology theme, and the calculation formula is:

[0147]

[0148] wherein, ∈ is a small constant, r i,j represents the relative change rate of the power technology theme T i at time j compared with the previous time;

[0149] The average change rate and the change rate standard deviation σ i of each power technology theme T i in the last m continuous time windows are calculated, and the formula is as follows:

[0150]

[0151] wherein, m is a preset parameter, n is the total number of time windows, represents the average evolution rate of the power technology theme T i in the last m time windows, and σ i represents the standard deviation of the evolution rate of the power technology theme T i in the last m time windows.

[0152] Based on the results of the trend modeling, the prospective power technology theme is determined, and the prospective power technology theme and the corresponding evolution trend are output, specifically including:

[0153] According to the preset prospective determination thresholds θ r , θ σ , according to the average evolution rate and the standard deviation σ i of each power technology theme T i in the last m time windows, the theme satisfying the following conditions is screened as the prospective power technology theme:

[0154] and σ i < θ σ

[0155] Output all subjects that meet the conditions, their time series curves and keyword descriptions as the prediction results.

[0156] By smoothing the time series of power technology subjects and calculating the rate of change, short-term fluctuations can be effectively removed, and the long-term trends and evolution patterns of power technology subjects can be accurately identified. Moving average method is used to smooth the time series, and then the rate of change of each subject is calculated, which can clearly reflect the growth or decline trend of the technology subject. At the same time, by calculating the average value and standard deviation of the rate of change, the stability and development potential of the technology subject can be further determined, and then those power technology subjects with foresight and growth are screened out. This process enables the prediction model to more accurately identify future technology trends that may affect the power field, providing valuable foresight guidance for decision-makers and ensuring that technology research and resource investment are more scientific and efficient.

[0157] Embodiment 2:

[0158] The embodiment provides a foresight power technology prediction system based on subject extraction and time modeling, as shown in Figure 3 , comprising:

[0159] A text data acquisition module is configured to collect patent data and paper data related to power technology.

[0160] A data preprocessing module is configured to clean, denoise and extract time labels from the collected text data.

[0161] A subject extraction module is configured to model the preprocessed text data based on the BERTopic method to obtain a plurality of power technology subjects and keyword descriptions thereof.

[0162] A time modeling module is configured to construct a subject time series according to the time labels in the text data, and to smooth and analyze the rate of change of the time series.

[0163] A foresight determination module is configured to filter out power technology subjects with foresight according to the trend modeling results, in combination with a preset evolution rate threshold and stability standard.

[0164] A result output module is configured to output the predicted foresight power technology subjects and their corresponding evolution trend curves and keyword information.

[0165] Another aspect of the embodiment provides an electronic device comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the foresight power technology prediction method based on subject extraction and time modeling according to any one of the above embodiments when executing the program.

[0166] The embodiment also provides a computer readable storage medium, which stores a computer program, and the program is executed by a processor to implement the subject extraction and time modeling based prospective power technology prediction method according to any one of the above.

[0167] If the above functions are realized in the form of software function units and sold or used as independent products, the functions can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application or parts of the present application that essentially contribute to the prior art can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in the embodiments of the present application. The aforementioned storage medium includes a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0168] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of various equivalent modifications or replacements within the technical range disclosed by the present application, and these modifications or replacements should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A forward-looking power technology forecasting method based on topic extraction and time modeling, characterized in that: The following steps are involved: Acquire text data related to power technology; Performing data preprocessing on the text data, wherein the data preprocessing includes text cleaning, noise removal, and time tag extraction; Using the BERTopic method to extract topics from the text data, multiple power technology topics are obtained; Based on the time tags extracted from the text data and in combination with the power technology topics, a topic time series is constructed, wherein the topic time series is used to describe the change in the frequency of occurrence of each power technology topic in the text data in different time windows; Performing trend modeling on the subject time series, wherein the trend modeling includes smoothing and rate of change analysis; Based on the results of the trend modeling, a forward-looking power technology theme is determined, and the forward-looking power technology theme and its corresponding evolution trend are output.

2. A forward-looking power technology forecasting method based on topic extraction and time modeling according to claim 1, characterized in that: The text data includes patent data and paper data.

3. The forward-looking power technology forecasting method based on topic extraction and time modeling according to claim 1 is characterized in that: The BERTopic method is used to extract topics from the text data to obtain multiple power technology topics, specifically including: Performing text embedding on the text data using a pre-trained BERT model to generate a high-dimensional semantic vector representation of the text data; Performing dimensionality reduction processing on the high-dimensional semantic vector representation based on the UMAP algorithm to obtain low-dimensional representation data; Performing cluster analysis on the low-dimensional representation data using the HDBSCAN clustering algorithm to generate a preliminary topic set; Extract keywords from the text contained in each topic in the preliminary topic set based on the TF-IDF method to generate the multiple power technology topics and their corresponding keyword descriptions; The multiple electric power technology topics are mapped to the original text data to form topic attribution labels, where the topic attribution labels are used to indicate the electric power technology topic corresponding to each piece of text data.

4. A forward-looking power technology forecasting method based on topic extraction and time modeling according to claim 3, characterized in that: The dimensionality reduction processing of the semantic vector representation based on the UMAP algorithm to obtain low-dimensional representation data specifically includes: Get the high-dimensional semantic vector representation corresponding to each text data, recorded as Where D is the dimension of the high-dimensional semantic space, and i represents the text data number; Based on the high-dimensional semantic vector, an adjacency graph G = (V, E, P) is constructed, where: V is a node set corresponding to all text semantic vectors; E is an edge set representing neighbor text pairs; P = {p ij } is the conditional probability distribution between node pairs in high-dimensional space, defined as follows: Among them, p ij Represents the conditional similarity probability between text data i and text data j in high-dimensional space, ρ i Represents the minimum distance between text data i and its nearest neighbor, used to control the local connection strength, σ i is the local scale factor; The UMAP algorithm is used to optimize the mapping of the adjacency graph G and construct the embedding distribution Q in the target low-dimensional space = {q ij }, the target low-dimensional space is a Euclidean space with dimension d Where d<<D is used to preserve the local adjacency relationship of the text semantic structure; By minimizing the cross entropy loss function between the high-dimensional distribution P and the low-dimensional distribution Q: Among them, q ij is the similarity between text data i and text data j in the target low-dimensional space; For low-dimensional representation data {y i } Perform back propagation and gradient optimization to finally obtain the low-dimensional representation data of each text data in the target low-dimensional space 5. A forward-looking power technology forecasting method based on topic extraction and time modeling according to claim 4, characterized in that: The similarity q between text data i and text data j in the target low-dimensional space ij for: in, is the low-dimensional representation data of text data i in the low-dimensional space, and a and b are the hyperparameters used to adjust the distribution curvature in the UMAP model.

6. The forward-looking power technology forecasting method based on topic extraction and time modeling according to claim 3 is characterized in that: The HDBSCAN clustering algorithm is used to perform cluster analysis on the low-dimensional representation data to generate a preliminary topic set, specifically including: Get the representation vector of each text data in low-dimensional space; Set parameters such as the minimum cluster size and minimum number of samples for the HDBSCAN clustering algorithm; Based on the low-dimensional representation data, the distance relationship between the samples is calculated to construct a density map between the samples; Using the density map, identify areas with higher local density and form several clusters as preliminary topic sets; Each piece of text data is assigned a corresponding topic label according to the cluster to which it belongs, and the data that does not belong to any cluster is marked as noise points; Get a preliminary topic set T = {T1, T2, ..., T k }, where each T j represents an electric power technology topic identified by the density clustering algorithm, and k is the total number of identified topics.

7. The forward-looking power technology forecasting method based on topic extraction and time modeling according to claim 3 is characterized in that: The keyword extraction of the text contained in each topic in the preliminary topic set based on the TF-IDF method to generate the multiple power technology topics and their corresponding keyword descriptions specifically includes: For the text data set corresponding to each topic in the preliminary topic set, perform word segmentation, remove stop words and low-frequency words, and obtain a standardized word set; Based on the term set and the text set, construct a term-document matrix M, where the matrix M is a sparse matrix, wherein each row corresponds to a term, each column corresponds to a text, and the matrix element M(i, j) represents the number of occurrences of term i in text data j; Based on the term-document matrix, the following TF-IDF calculation formula is used to perform weighted calculation on each term in each topic: TF-IDF(t i ,d j )=TF(t i ,d j )×IDF(t i ) in, Represents the term t i In text data d j The word frequency in f i,j For term t i In text data d j The number of times it appears in k f k,j For text data d j The total number of occurrences of all terms in ; Represents the term t i The inverse document frequency, N is the total number of all text data, n i To contain the term t i The amount of text data; Sort all terms according to the TF-IDF value, and select the terms with the highest scores as representative keywords of the current topic; Each topic is combined with its corresponding representative keyword to generate multiple power technology topics and their keyword description sets for subsequent time modeling and trend analysis.

8. The forward-looking power technology forecasting method based on topic extraction and time modeling according to claim 1 is characterized in that: The step of constructing a topic time series based on the time tags extracted from the text data and combining the power technology topic specifically includes: Associate the power technology topic corresponding to each text data with its time tag to form a topic-time pair; Set the time granularity to year or quarter, and build a unified time window set {t1, t2, ..., t n }; For each power technology topic T i , statistics of each time window t j The number of text data that appear in the topic is used to construct a topic time frequency series, which is defined as follows: Among them, f i,j Indicates the power technology theme T i In the time window t j The frequency of occurrence in, N is the total number of text data, L k The power technology theme label of the k-th text data, Y k represents the time tag of the k-th text data, δ is a Boolean indicator function, which is 1 if the condition is met and 0 otherwise; The time frequency series of each power technology topic is expressed as: S i ={f i,1 ,f i,2 ,...,f i,n } Among them, S i For the power technology theme T i time series, Construct the time series of all electric power technology topics {S1, S2, ..., S k }.

9. The forward-looking power technology forecasting method based on topic extraction and time modeling according to claim 1 is characterized in that: The trend modeling of the subject time series specifically includes: For each power technology theme, the time series S i ={f i,1 ,f i,2 ,...,f i,n Smoothing is performed to eliminate the interference of short-term fluctuations on trend identification, and the sliding average method is used to calculate the smoothed time series The formula is: in, Represents the power technology topic T in the jth time window i The smoothed frequency value of , w is the time window width; Based on the smoothed series, calculate the power technology theme T i The rate of change r in each time window i,j , which is used to reflect the growth trend of technology topics, and its calculation formula is: Among them, ∈ is a small constant, r i,j Indicates the power technology theme T i The relative rate of change at time j compared to the previous time; Calculate T for each power technology topic i The average rate of change over the most recent m consecutive time windows and the standard deviation of the rate of change σ i , the formula is as follows: Among them, m is the preset parameter, n is the total number of time windows, Indicates the power technology theme T i The average evolution rate in the last m time windows, σ i Indicates the power technology theme T i The standard deviation of the evolution rate over the last m time windows.

10. A forward-looking power technology forecasting method based on topic extraction and time modeling according to claim 1 or 9, characterized in that: Determining a forward-looking power technology theme based on the trend modeling results and outputting the forward-looking power technology theme and its corresponding evolution trend specifically includes: According to the preset forward-looking judgment threshold θ r ,θ σ , according to each power technology theme T i The average evolution rate in the last m time windows and standard deviation σ i Topics that meet the following criteria are selected as forward-looking power technology topics: And you i <θ σ Output all topics that meet the conditions, their time series curves and keyword descriptions as prediction results.

Images