Power data quality feature extraction method and system based on pre-trained language model

By building a multi-level metadata gene regulation network and generating pre-training language models, the problem of low feature extraction efficiency caused by the variety of power data and the huge number of power data is solved, and the accurate and efficient extraction of power data quality characteristics is achieved, and the power system is improved. Stability and service level.

CN118503687BActive Publication Date: 2024-11-08INFORMATION COMM COMPANY STATE GRID SHANDONG ELECTRIC POWER +1
View PDF 5 Cites 0 Cited by

Patent Information

Application Number
CN202410969384.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-07-19
Publication Date
2024-11-08
Estimated Expiration
2044-07-19

AI Technical Summary

Technical Problem

In the prior art, due to the diverse types and large quantities of power data, the extraction of features is large, the accuracy and efficiency are low, and it is difficult to meet application needs.

Method used

The power data quality feature extraction method based on the pre-trained language model is adopted. By constructing a multi-level metadata gene regulation network, a pre-trained language model is generated, and the power data quality features are automatically extracted, so as to achieve rapid processing and adaptability of massive heterogeneous data.

Benefits of technology

It realizes accurate and efficient extraction of power data quality characteristics, improves data processing efficiency and adaptability, adapts to dynamic changes in the power system, and improves power supply reliability and service level.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118503687B_ABST
    Figure CN118503687B_ABST
Patent Text Reader

Abstract

The present invention discloses a method and system for extracting power data quality features based on a pre-trained language model, which belongs to the field of intelligent operation and maintenance of power systems, wherein the method includes: collecting historical execution records according to the target task type; performing execution process splitting on the historical execution records to obtain multi-level metadata; performing relevance analysis and correlation analysis on metadata at each level to obtain relevance analysis results and correlation analysis results; generating a pre-trained language model; collecting power monitoring data, classifying it, and obtaining gene node matching results; when the number of gene nodes is equal to 0, cold start marking the power monitoring data. The present invention solves the technical problem that the workload of extracting features in the prior art is large, resulting in low accuracy and efficiency of power data quality extraction, and achieves the technical effect of accurately and efficiently extracting power data quality features by constructing a multi-level metadata gene regulatory network and using it to generate a pre-trained language model.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of intelligent operation and maintenance of power systems, and in particular to a method and system for extracting power data quality features based on a pre-trained language model. Background Art

[0002] During the operation of the power system, a massive amount of power data will be generated, covering various links such as power generation, transmission, transformation, distribution, and power consumption, including telemetry data, status data, alarm data, energy consumption data, etc. The data scale is large and the time span is long. Accurately and efficiently extracting the quality characteristics of power data is of great significance for ensuring the safe and stable operation of the power system and improving power supply reliability and service level. However, the existing technology needs to design independent feature indicator systems and calculation rules for different data types. Due to the wide variety of power data and the huge amount of data, the characteristics and laws of different types of data are significantly different, resulting in a very large workload for feature extraction; at the same time, the manually designed feature extraction rules have poor versatility and insufficient interpretability, and are difficult to adapt to the dynamic changes of the power system. Therefore, these problems make it difficult for the accuracy and efficiency of power data quality feature extraction to meet the growing application needs. Summary of the invention

[0003] The present invention provides a method and system for extracting power data quality features based on a pre-trained language model, aiming to solve the technical problem in the prior art that the power data quality extraction has low accuracy and efficiency due to the large amount of power data and the large workload of feature extraction.

[0004] The technical effect of accurately and efficiently extracting the quality characteristics of power data was achieved by constructing a multi-level metadata gene regulatory network and using it to generate a pre-trained language model.

[0005] In view of the above problems, the present invention provides a method and system for extracting power data quality features based on a pre-trained language model.

[0006] The first aspect disclosed by the present invention provides a method for extracting power data quality features based on a pre-trained language model, the method comprising: collecting historical execution records according to the target task type; performing execution process splitting on the historical execution records to obtain first-level metadata, second-level metadata, and up to the N-th level metadata, wherein the semantic levels of the first-level metadata, the second-level metadata, and up to the N-th level metadata are sequentially enhanced, and the k+1-th level metadata is generated dependent on the k-th level metadata, where N is an integer and k is an integer; performing correlation analysis on the k-th level metadata according to the k+1-th level metadata to obtain a correlation analysis Results; According to the k+1th level metadata, the correlation analysis is performed on the kth level metadata to obtain the correlation analysis results; According to the association analysis results and the correlation analysis results, the local gene regulatory network of the kth level metadata to the k+1th level metadata is constructed; Repeat the cyclic analysis, fuse several local gene regulatory networks, generate a gene regulatory knowledge base, perform unsupervised configuration, and generate a pre-trained language model; Collect power monitoring data, classify it through the pre-trained language model, and obtain the gene node matching results; When the number of gene nodes in the gene node matching result is equal to 0, the power monitoring data is marked as cold start.

[0007] Another aspect disclosed by the present invention provides a power data quality feature extraction system based on a pre-trained language model, the system comprising: a record collection module, used to collect historical execution records according to the target task type; a process splitting module, used to perform execution process splitting on the historical execution records, and obtain first-level metadata, second-level metadata, and up to the N-th level metadata, the semantic levels of the first-level metadata, the second-level metadata, and up to the N-th level metadata are enhanced in sequence, and the k+1-th level metadata depends on the k-th level metadata to generate, N is an integer, and k is an integer; an association analysis module, used to perform association analysis on the k-th level metadata according to the k+1-th level metadata to obtain the association analysis result; a similarity analysis module , used to perform correlation analysis on the k-th level metadata according to the k+1-th level metadata to obtain the correlation analysis results; the network construction module is used to construct the local gene regulation network of the k-th level metadata to the k+1-th level metadata according to the association analysis results and the correlation analysis results; the model generation module is used to repeat the cycle analysis, fuse several local gene regulation networks, generate a gene regulation knowledge base, perform unsupervised configuration, and generate a pre-trained language model; the data classification module is used to collect power monitoring data, classify it through the pre-trained language model, and obtain the gene node matching result; the cold start identification module is used to perform cold start identification on the power monitoring data when the number of gene nodes in the gene node matching result is equal to 0.

[0008] One or more technical solutions provided in the present invention have at least the following technical effects or advantages:

[0009] The historical execution records are collected according to the target task type to provide raw materials for subsequent analysis and lay the data foundation; the historical execution records are split into execution processes to obtain the first-level metadata, the second-level metadata, and the N-th-level metadata, and the historical data is deconstructed into multiple metadata levels with successively deeper semantic levels, so as to realize the data organization form from shallow to deep and from surface to inside, and prepare for metadata association and fusion. Compared with directly processing the original data, the hierarchical organization reveals the internal structure of the data, which is convenient for subsequent analysis; the metadata of the kth level is analyzed for correlation according to the metadata of the k+1th level to obtain relevant Based on the correlation analysis results, the intrinsic connections between adjacent metadata levels are mined through bottom-up correlation analysis to establish semantic connections. Compared with the independent analysis of each layer of metadata, the correlation between levels is considered to more comprehensively characterize the relationship between data. The correlation analysis results are obtained by performing correlation analysis on the metadata of the kth level according to the metadata of the k+1th level. The similarity between adjacent metadata levels is measured through bottom-up correlation analysis to further quantify the correlation between metadata. Based on the correlation analysis results and the correlation analysis results, a local gene regulatory network of the metadata of the kth level to the metadata of the k+1th level is constructed to construct the metadata. According to the regulatory relationship network between levels, the internal structure and rules of the data are revealed. Compared with traditional feature engineering, the complex relationship of metadata can be more accurately characterized through biologically inspired network construction; repeated cyclic analysis is performed to integrate several local gene regulatory networks, generate a gene regulatory knowledge base, perform unsupervised configuration, and generate a pre-trained language model so that it can learn to represent the quality characteristics of power data; power monitoring data is collected, classified through the pre-trained language model, and the gene node matching results are obtained, and the quality characteristics of the power monitoring data are automatically extracted to achieve rapid processing of massive heterogeneous data. Compared with the traditional extraction rule matching, the deep learning model can adapt to different data types and extract more comprehensive and in-depth features; when the number of gene nodes in the gene node matching result is equal to 0, the power monitoring data is cold-started. When there is data that cannot be classified, it is regarded as low-quality data and the technical solution of cold-starting identification solves the technical problem of low accuracy and efficiency of power data quality extraction in the prior art due to the diverse types and huge number of power data and the large workload of feature extraction, and achieves the technical effect of accurately and efficiently extracting power data quality characteristics by constructing a multi-level metadata gene regulatory network and using it to generate a pre-trained language model.

[0010] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it can be implemented according to the contents of the specification. In order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 A flow chart of a method for extracting power data quality features based on a pre-trained language model is provided for an embodiment of the present invention;

[0012] Figure 2 A structural schematic diagram of a power data quality feature extraction system based on a pre-trained language model is provided for an embodiment of the present invention.

[0013] Explanation of the accompanying drawings: record collection module 11, process splitting module 12, association analysis module 13, similarity analysis module 14, network construction module 15, model generation module 16, data classification module 17, cold start identification module 18. DETAILED DESCRIPTION

[0014] The overall idea of ​​the technical solution provided by the present invention is as follows:

[0015] The embodiments of the present invention provide a method and system for extracting power data quality features based on a pre-trained language model.

[0016] First, according to the target task type, relevant historical execution records are collected to provide a comprehensive and accurate data basis for subsequent analysis. Then, the collected historical data is subjected to multi-level, semantically progressive metadata splitting. In a bottom-up manner, the correlation and relevance between metadata are analyzed layer by layer to explore the inherent structure and regularity of the data. On this basis, a local gene regulation network of local metadata is constructed, and a complete gene regulation knowledge base is formed through loop analysis and fusion optimization. The gene regulation knowledge base is used for unsupervised configuration to generate a pre-trained language model of power data quality features. In the application stage, the pre-trained language model is directly used to classify and process the newly collected power monitoring data, quickly and accurately extract its quality features, and cold-start the low-quality data.

[0017] Compared with the existing technology, the present invention constructs a multi-level metadata gene regulatory network, mines the deep semantic associations of data, and generates a pre-trained language model. It grasps the representation of power data quality characteristics in a data-driven and autonomous learning manner, replaces the inefficient rule-based feature extraction in traditional methods, and accurately and efficiently extracts power data quality characteristics.

[0018] After introducing the basic principles of the present invention, various non-limiting implementation methods of the present invention will be specifically described below in conjunction with the accompanying drawings.

[0019] Embodiment 1

[0020] like Figure 1 As shown, an embodiment of the present invention provides a method for extracting power data quality features based on a pre-trained language model, the method comprising:

[0021] S1: Collect historical execution records according to the target task type;

[0022] Specifically, the target task type refers to the application task in the specific power field that needs to be completed. The types of power data are complex and diverse, including current, voltage, active power, reactive power and other types, and there is a certain correlation between the various types of data. For different target task types, the required historical execution records are different. Therefore, in order to improve the accuracy and efficiency of subsequent power data quality feature extraction, the relevant historical execution records are first collected in a targeted manner according to the specific target task type.

[0023] For example, if the target task type is power load forecasting, historical load data, weather data, holiday data, etc. are collected as historical execution records; if the target task type is power equipment fault diagnosis, historical equipment operating parameter data, alarm data, maintenance records, etc. are collected as historical execution records.

[0024] Through targeted historical execution record collection, a high-quality data foundation is provided for subsequent multi-level metadata splitting and gene regulatory network construction, thereby improving the effectiveness and reliability of the entire power data quality feature extraction.

[0025] S2: performing execution process splitting on the historical execution record to obtain first-level metadata, second-level metadata, and finally N-level metadata, wherein the semantic levels of the first-level metadata, the second-level metadata, and finally the N-level metadata are sequentially enhanced, and the k+1-th level metadata is generated depending on the k-th level metadata, where N is an integer and k is an integer;

[0026] Specifically, the extraction of power data quality features requires full mining of the intrinsic connections and semantic information of the data. However, the original historical execution records are mixed with data from multiple execution steps, and the semantic level is single, which makes it difficult to directly use them for feature extraction. Therefore, by splitting the execution process of the historical execution records and using multi-level metadata splitting, the data granularity and semantic representation are refined layer by layer, laying the foundation for subsequent association analysis and control network construction.

[0027] In the splitting process, first, the historical execution records are segmented according to the execution process to obtain the first-level metadata with a coarser granularity. On this basis, the first-level metadata is further split and refined to obtain the second-level metadata with a higher semantic level. And so on, the iterative splitting steps are repeated until the N-th level metadata with the finest granularity and richest semantics is obtained. Among them, the generation of each layer of metadata depends on the splitting result of the previous layer of metadata, showing the characteristics of layer-by-layer progression and gradual enhancement of semantics. The generation of the k+1th level metadata depends on the kth level metadata, showing a bottom-up progressive relationship. For example, for the historical execution records of power equipment fault diagnosis, the first-level metadata is the equipment type and the time of fault occurrence; the second-level metadata is refined on this basis into the specific model of the faulty equipment and the operating parameters before the fault occurs; the third-level metadata is further expanded into faulty components, fault modes, historical fault cases of equipment of the same type, etc. Among them, N and k are integer variables used to describe the metadata hierarchy; N represents the total number of metadata splitting levels, that is, starting from the original historical execution record, after layer-by-layer splitting and semantic enhancement, the final level of the finest-grained metadata belongs to; k is a loop iteration variable, which is used to refer to the current level in the metadata splitting process.

[0028] Through multi-level splitting, the originally mixed and redundant historical execution records are converted into semantically clear and structurally reasonable metadata sequences, fully mining the rich semantic information contained in the historical execution records, providing strong support for subsequent correlation analysis and relevance analysis, and creating conditions for accurately extracting power data quality characteristics.

[0029] S3: performing a correlation analysis on the metadata at the kth level according to the metadata at the k+1th level to obtain a correlation analysis result;

[0030] Further, the correlation analysis results are obtained, including:

[0031] Extracting a first value assignment sequence of a first data attribute of the metadata at the k+1th level through the historical execution record;

[0032] By using the historical execution records, performing execution frequency analysis on the k-th level metadata based on the first data attribute, obtaining a first high-frequency data attribute, a second high-frequency data attribute, and up to an M-th high-frequency data attribute of the k-th level metadata, where M is an integer;

[0033] Extracting the first high-frequency data attribute, the second high-frequency data attribute, and the Mth high-frequency data attribute through the historical execution record and the first assignment sequence, and obtaining a second assignment sequence, a third assignment sequence, and the M+1th assignment sequence;

[0034] Taking the first value assignment sequence as the reference sequence, and taking the second value assignment sequence, the third value assignment sequence to the M+1th value assignment sequence as the comparison sequence, a grey correlation analysis is performed to generate a first data attribute correlation analysis result, which is added to the correlation analysis result.

[0035] Furthermore, frequent analysis is performed, including:

[0036] According to the first data attribute, based on the historical execution record, a plurality of data attributes of the metadata of the kth level are marked with call frequencies to obtain a plurality of call frequency coefficients;

[0037] According to the calling frequency coefficient threshold, based on the calling frequency coefficients, the data attributes are sorted to obtain the first high-frequency data attribute, the second high-frequency data attribute, and so on to the Mth high-frequency data attribute.

[0038] In a feasible implementation, after completing the multi-level metadata splitting, each layer of metadata is composed of multiple data attributes and their assignments. For a certain metadata of the k+1th layer, select the first data attribute, that is, any one of the data attributes, and extract the assignment of the first data attribute at different historical points in time according to the historical execution records to form a complete first assignment sequence. For example, in the power equipment fault diagnosis task, the k+1th layer of metadata contains the data attribute of fault severity. Then, extract the severity assignment corresponding to each fault from the historical execution records, and arrange them in chronological order to form an ordered severity assignment sequence, which is the first assignment sequence.

[0039] Secondly, according to the obtained first data attribute and the first assignment sequence, the historical execution records are scanned and analyzed one by one. For each historical record, the value of the first data attribute of the k+1th layer metadata is checked, and the frequency of calling each attribute of the kth layer metadata under the value is counted, and the calling frequency of all historical execution records is accumulated to obtain the total calling frequency of each attribute of the kth layer metadata. In order to facilitate the comparison and sorting between different attributes, the calling frequency is normalized, for example, the calling frequency coefficient is calculated, that is, the calling frequency of each attribute is divided by the total number of historical execution records. The larger the calling frequency coefficient, the higher the correlation between the attribute and the first attribute, and the greater the possibility of being called when the value of the first attribute changes. For example, in the power equipment fault diagnosis task, the first data attribute of the k+1th layer metadata is the severity of the fault, corresponding to 1000 historical execution records, and the kth layer metadata contains multiple attributes such as current, voltage, and temperature. In the 1000 historical execution records, the current attribute is called 180 times, the voltage attribute is called 150 times, and the temperature attribute is called 120 times. Then the calling frequency coefficient of the current attribute is 180 / 1000=0.18, the calling frequency coefficient of the voltage attribute is 150 / 1000=0.15, and the calling frequency coefficient of the temperature attribute is 120 / 1000=0.12, thereby obtaining the numerical correlation measure between each data attribute of the k-th layer metadata and the first data attribute of the k+1-th layer metadata, and obtaining several calling frequency coefficients. Then, after obtaining several calling frequency coefficients of each attribute of the k-th layer metadata, the M high-frequency attributes with the highest correlation with the first data attribute are screened out. Specifically, first, according to the application scenario and empirical knowledge, a calling frequency coefficient threshold is set to indicate the minimum correlation between the attribute and the first data attribute; after determining the calling frequency coefficient threshold, all data attributes of the k-th layer metadata are scanned and judged one by one, and for each data attribute, the size relationship between its calling frequency coefficient and the calling frequency coefficient threshold is compared; if the calling frequency coefficient is greater than or equal to the calling frequency coefficient threshold, the attribute is marked as a candidate for the high-frequency attribute; if the calling frequency coefficient is less than the calling frequency coefficient threshold, the attribute is eliminated from the candidate range. Based on the candidate set, sort them according to the size of the call frequency coefficient, and obtain M data attributes whose call frequency coefficients are greater than the call frequency coefficient threshold, which are respectively used as the first high-frequency data attribute, the second high-frequency data attribute, and the Mth high-frequency data attribute. Among them, M is a positive integer, which represents the total number of high-frequency data attributes.

[0040] Next, with the extracted first attribute assignment sequence as a reference, the historical execution records are scanned and extracted again. For each historical record, find the corresponding position in the first assignment sequence based on the first data attribute of its k+1th layer of metadata; at the same time, extract the values ​​of the M high-frequency data attributes of the kth layer of metadata in the historical record, and store them in M ​​independent assignment sequences respectively. By means of synchronous extraction, the consistency of different attribute assignment sequences on the time axis is ensured, laying the foundation for subsequent association analysis. After traversing and extracting all historical records, M high-frequency attribute assignment sequences corresponding to the first assignment sequence are obtained, which are successively called the second assignment sequence, the third assignment sequence, and the M+1th assignment sequence. Each assignment sequence records the specific values ​​of a high-frequency data attribute at different historical moments, reflecting the changes of the attribute over time.

[0041] After obtaining the first attribute value sequence and M high-frequency attribute value sequence, these sequences are subjected to grey correlation analysis to quantitatively characterize the correlation strength between the first data attribute of the metadata of the k+1th layer and the high-frequency data attribute of the metadata of the kth layer, and generate corresponding correlation results. Among them, the grey correlation analysis judges the correlation degree between different attributes by the similarity of the geometric shapes of the sequence curves. The higher the correlation degree, the greater the mutual influence between the attributes, and the closer the geometric shapes of the curves are; the first value sequence reflects the change trend of the first data attribute of the metadata of the k+1th layer, while the second value sequence, the third value sequence and the M+1 value sequence correspond to the change trend of the M high-frequency data attributes of the metadata of the kth layer. The first value sequence is regarded as the benchmark sequence, and the other M value sequences are regarded as the comparison sequences. By analyzing their similarity on the time scale, the correlation law between metadata attributes of different levels is obtained. Specifically, first, necessary data preprocessing is performed on each assignment sequence, including averaging and dimensionless conversion, to eliminate the influence of dimension and numerical range differences; then, the grey correlation coefficient between each comparison sequence and the benchmark sequence is calculated respectively to measure their geometric similarity at each moment; then, the grey correlation coefficients at all moments are comprehensively averaged to obtain the grey correlation degree between the comparison sequence and the benchmark sequence, which is used as a quantitative indicator of the correlation strength between the high-frequency data attribute and the first data attribute; the grey correlation degrees of each high-frequency data attribute are arranged in order from large to small to form the correlation analysis result of the first data attribute, and it is added to the correlation analysis result.

[0042] By processing each data attribute in each level of metadata in the above manner, a complete correlation analysis result can be obtained, which fully utilizes the changing trend information of cross-level metadata attributes on a time scale, obtains the intrinsic connection and influence mechanism between them, and lays the foundation for the subsequent extraction of power data quality features.

[0043] S4: performing a correlation analysis on the k-th level metadata according to the k+1-th level metadata to obtain a correlation analysis result;

[0044] Furthermore, the correlation analysis results are obtained, including:

[0045] Taking the first high-frequency data attribute as an independent variable, taking the second high-frequency data attribute to the Mth high-frequency data attribute as quantitative, taking the first data attribute as a dependent variable, assigning values ​​through the historical execution record, and obtaining an independent variable assignment sequence and a dependent variable assignment sequence;

[0046] Performing a Pearson correlation coefficient analysis on the independent variable value assignment sequence and the dependent variable value assignment sequence to obtain a first Pearson correlation coefficient pointing from the first high-frequency data attribute to the first data attribute;

[0047] According to the first Pearson correlation coefficient, an enhanced identification relationship, a suppressed identification relationship or no relevant identification is obtained and added into the correlation analysis result.

[0048] Furthermore, the independent variable assignment sequence and the dependent variable assignment sequence are obtained, including:

[0049] Retrieve the first high-frequency data attribute, the second high-frequency data attribute, up to the first distribution interval, the second distribution interval, up to the Mth distribution interval of the Mth high-frequency data attribute;

[0050] Traversing the second distribution interval to the Mth distribution interval, extracting the interval upper limit value, performing quantitative assignment on the second high-frequency data attribute to the Mth high-frequency data attribute, and generating a first quantitative assignment result;

[0051] Traversing the second distribution interval to the Mth distribution interval, extracting the interval lower limit value, performing quantitative assignment on the second high-frequency data attribute to the Mth high-frequency data attribute, and generating a second quantitative assignment result;

[0052] Traversing the second distribution interval to the Mth distribution interval, extracting the median of the interval, performing quantitative assignment on the second high-frequency data attribute to the Mth high-frequency data attribute, and generating a third quantitative assignment result;

[0053] The first quantitative assignment result, the second quantitative assignment result and the third quantitative assignment result are respectively used as constraints, the first high-frequency data attribute is used as an independent variable, and the first data attribute is used as a dependent variable. The historical execution records are used for assignment to obtain an independent variable assignment sequence and a dependent variable assignment sequence.

[0054] Further, obtaining an enhanced identification relationship, a suppressed identification relationship or no related identification according to the first Pearson correlation coefficient includes:

[0055] The first Pearson correlation coefficient includes a first sub-Pearson correlation coefficient, a second sub-Pearson correlation coefficient, and finally a third sub-Pearson correlation coefficient;

[0056] When at least two of the first sub-Pearson correlation coefficient, the second sub-Pearson correlation coefficient, and the third sub-Pearson correlation coefficient are enhanced relations, generating the enhanced identification relation;

[0057] When at least two of the first sub-Pearson correlation coefficient, the second sub-Pearson correlation coefficient, and the third sub-Pearson correlation coefficient are in an inhibition relationship, generating the inhibition identification relationship;

[0058] Otherwise, the no-correlation flag is generated.

[0059] In a preferred embodiment, in order to obtain the correlation analysis result, first, for the first high-frequency data attribute, extract its minimum and maximum values ​​in all historical execution records, and determine the lower limit and upper limit of the value of the first high-frequency attribute accordingly, so as to obtain the first distribution interval, which reflects the variation range of the first high-frequency data attribute in the actual business scenario; similarly, for the second high-frequency attribute, by analyzing the minimum and maximum values ​​of its historical values, the second distribution interval can be determined to characterize the value boundary of the second high-frequency attribute; and so on, until the Mth distribution interval of the Mth high-frequency data attribute is determined. Then, in the order from the second high-frequency data attribute to the Mth high-frequency data attribute, traverse their respective corresponding distribution intervals in turn, that is, the second distribution interval to the Mth distribution interval, and for each distribution interval, extract its upper limit value as the quantitative assignment of the current high-frequency data attribute, and generate a quantitative assignment vector containing M-1 elements, which is called the first quantitative assignment result. Each element in the first quantitative assignment result represents the data feature when the corresponding high-frequency data attribute takes its historical maximum value. At the same time, also in the order from the second high-frequency attribute to the Mth high-frequency attribute, traverse their respective distribution intervals one by one, extract the lower limit value of each distribution interval as the quantitative assignment of the current high-frequency data attribute, and generate another quantitative assignment vector containing M-1 elements, called the second quantitative assignment result. Each element in the second quantitative assignment result represents the data feature when the corresponding high-frequency data attribute takes its historical minimum value. At the same time, still in the order from the second high-frequency attribute to the Mth high-frequency attribute, traverse their respective distribution intervals one by one, extract the median value of each distribution interval as the quantitative assignment of the current high-frequency attribute, and generate a third quantitative assignment vector containing M-1 elements, called the third quantitative assignment result. Each element in the third quantitative assignment result represents the data feature when the corresponding high-frequency data attribute takes its historical median value. Then, based on the first quantitative assignment result, search in the historical execution records for data records whose values ​​from the second high-frequency data attribute to the Mth high-frequency data attribute are equal to or close to the corresponding elements in the first quantitative assignment result, representing the corresponding values ​​of the first high-frequency data attribute and the first data attribute when the other high-frequency data attributes take the upper limit value, arrange the values ​​of the first high-frequency attribute in these data records in chronological order to form an independent variable assignment sequence, and arrange the values ​​of the first data attribute in the corresponding records in the same order to form a dependent variable assignment sequence, and these two sequences constitute a correlation analysis sample pair under the high value constraint. Similarly, with the second quantitative assignment result as the constraint condition, repeat the above process, extract the records that meet the constraint in the historical execution records, and form another set of independent variable assignment sequences and dependent variable assignment sequences from them, representing the corresponding relationship between the first high-frequency data attribute and the first data attribute under the low value constraint.Similarly, with the third quantitative assignment result as the constraint condition, the process is repeated again to extract the historical execution records that meet the median constraint, and generate the corresponding independent variable assignment sequence and dependent variable assignment sequence, reflecting the correlation characteristics of the two data attributes under normal conditions. Through the above three rounds of extraction and matching, three sets of paired independent variable assignment sequences and dependent variable assignment sequences are obtained, which correspond to the corresponding relationship between the first high-frequency data attribute and the first data attribute in high-value, low-value, and medium-value scenarios, respectively, and comprehensively and meticulously characterize the correlation performance of the two data attributes under different conditions, providing rich and targeted sample data for subsequent Pearson correlation analysis.

[0060] Subsequently, the Pearson correlation coefficients between each pair of sequences are calculated with the three sets of independent variable assignment sequences and dependent variable assignment sequences extracted as input. For the first set of independent variable assignment sequences and dependent variable assignment sequences, the Pearson correlation coefficient is calculated to obtain the correlation measure between the first high-frequency data attribute and the first data attribute under the high value constraint, which is used as the first sub-Pearson correlation coefficient to quantify the association strength and change trend consistency between the first high-frequency data attribute and the first data attribute when other high-frequency data attributes take the upper limit value; similarly, the Pearson correlation coefficient is calculated for the second and third sets of independent variable assignment sequences and dependent variable assignment sequences, respectively, to obtain the correlation measure between the first high-frequency data attribute and the first data attribute under the low value constraint and the median constraint, which are the second sub-Pearson correlation coefficient and the third sub-Pearson correlation coefficient, respectively, which characterize the association characteristics between the two data attributes when other high-frequency data attributes take the lower limit value and the median value. The three sub-Pearson correlation coefficients are combined together to obtain the complete Pearson correlation coefficient of the first high-frequency data attribute pointing to the first data attribute, which is the first Pearson correlation coefficient.

[0061] Next, check the signs and sizes of the three calculated sub-Pearson correlation coefficients. If at least two of the sub-Pearson correlation coefficients are positive and their absolute values ​​exceed the preset enhancement relationship threshold, it means that an increase in the value of the first high-frequency data attribute will lead to an increase in the value of the first data attribute, and there is a significant positive correlation between the two. The first high-frequency data attribute has an enhancement effect on the first data attribute. At this time, generate an enhancement identification relationship from the first high-frequency data attribute to the first data attribute, which is used to characterize the positive influence mechanism between the two attributes. Similarly, check the signs and sizes of the three sub-Pearson correlation coefficients. If at least two of the sub-Pearson correlation coefficients are negative and their absolute values ​​exceed the preset inhibition relationship threshold, it means that an increase in the value of the first high-frequency data attribute will lead to a decrease in the value of the first data attribute, and there is a significant negative correlation between the two. The first high-frequency data attribute has an inhibition effect on the first data attribute. At this time, generate an inhibition identification relationship from the first high-frequency data attribute to the first data attribute, which is used to characterize the reverse influence mechanism between the two attributes. If the positive and negative signs of the three sub-Pearson correlation coefficients are inconsistent, and the absolute values ​​are small, and do not exceed the judgment threshold of the enhancement relationship or the inhibition relationship, it means that a consistent correlation conclusion cannot be drawn, and there is no significant linear correlation between the first high-frequency data attribute and the first data attribute, or the correlation is very weak. At this time, a no correlation mark from the first high-frequency data attribute to the first data attribute is generated, indicating that there is no clear influence between the two attributes. Afterwards, the obtained enhancement mark relationship, inhibition mark relationship or no correlation mark is added to the correlation analysis result to indicate the correlation between the first high-frequency data attribute and the first data.

[0062] S5: constructing a local gene regulatory network of the k-th level metadata to the k+1-th level metadata according to the association analysis result and the correlation analysis result;

[0063] Furthermore, this step includes:

[0064] According to the correlation analysis result, constructing a thickness feature of a connection line from the k-th level metadata to the k+1-th level metadata, wherein the greater the correlation, the thicker the connection line;

[0065] According to the correlation analysis result, construct a line pointing mark from the k-th level metadata to the k+1-th level metadata, wherein the enhanced mark relationship is an arrow mark from the k-th level metadata to the k+1-th level metadata, the suppressed mark relationship is a horizontal line perpendicular to the line from the k-th level metadata to the k+1-th level metadata, and the unrelated mark is no mark;

[0066] The metadata from the kth level to the k+1th level are distributed to a plurality of gene nodes, and the local gene regulatory network of the metadata from the kth level to the metadata from the k+1th level is constructed according to the line thickness feature and the line pointing identifier.

[0067] In a preferred embodiment, when it is implemented, for each data attribute of the k-level metadata, the gray correlation between it and the data attribute in the k+1-level metadata is checked, indicating the geometric similarity of the two attributes in the time series. The larger the correlation, the more consistent the change trend of the two attributes, and the closer the relationship. The size of the correlation in the correlation analysis result specifies the thickness level of the line between the k-level attribute and the k+1-level attribute. The level division presets several threshold intervals according to the distribution of the correlation. The higher the interval the correlation falls into, the thicker the corresponding line. For example, the correlation above 0.8 is defined as the thickest line, 0.6 to 0.8 is the second thickest line, 0.4 to 0.6 is the medium line, 0.2 to 0.4 is the second thinnest line, and 0.2 or less is the thinnest line. Through this mapping mechanism, the continuous correlation is quantified into discrete thickness levels, intuitively depicting the difference in the closeness of the relationship between cross-level metadata. The thicker the line, the stronger the correlation between the two attributes and the closer their positions in the semantic network; the thinner the line, the weaker the correlation between the two attributes and the more distant they are in the semantic network.

[0068] At the same time, for each data attribute of the metadata at the kth level, check the type of related identification between it and each data attribute of the metadata at the k+1th level. If it is an enhanced identification relationship, it means that the increase in the value of the attribute at the kth level will lead to the increase in the value of the attribute at the k+1th level, and there is a positive promotion mechanism between the two. At this time, add an arrow from the attribute at the kth level to the attribute at the k+1th level on the line connecting the two attributes, which intuitively indicates the positive direction of the causal relationship. If it is an inhibited identification relationship, it means that the increase in the value of the attribute at the kth level will lead to the decrease in the value of the attribute at the k+1th level, and there is a reverse constraint mechanism between the two. At this time, add a horizontal line perpendicular to the line connecting the two attributes to indicate the reverse direction of the causal relationship. If there is no related identification, it means that there is a lack of significant linear correlation between the two attributes, and there is no clear causal direction. At this time, no directional identification is added between them. Through qualitative symbolic directional identification, the causal information hidden in the data is intuitively presented, revealing the influence transmission mechanism between cross-level metadata.

[0069] Subsequently, the attributes of the metadata of the kth level and the metadata of the k+1th level are abstracted and mapped into several gene nodes, each of which represents a semantic unit corresponding to an attribute. Then, according to the thickness characteristics of the lines between different attributes, different thicknesses of edges are drawn between the corresponding gene nodes; thick edges indicate that the two attributes are highly correlated and close to each other; thin edges indicate that the two attributes are less correlated and relatively distant, which directly reflects the difference in the degree of correlation between different gene nodes in the network. Then, according to the direction mark of the line between different attributes, arrows, horizontal lines or no marks are added to the corresponding edges. Through the composition means of node mapping, roughness rendering, and direction mark, an intuitive and vivid cross-level metadata local gene regulatory network is generated, and the correlation law and influence mechanism between the metadata of the kth level and the metadata of the k+1th level are systematically displayed in a graphical way.

[0070] By adopting the gene regulatory network, the dependency logic between metadata attributes is abstracted, the internal connections and action rules between metadata are displayed intuitively and vividly, and the efficiency and depth of understanding of complex power data relationships are improved.

[0071] S6: Repeat the cyclic analysis, fuse several local gene regulatory networks, generate a gene regulatory knowledge base, perform unsupervised configuration, and generate a pre-trained language model;

[0072] Furthermore, a pre-trained language model is generated, including:

[0073] Using the gene regulation knowledge base as the analysis benchmark rule of the pre-trained language model;

[0074] Collecting historical data sets of power data, marking them through the gene regulation knowledge base, and generating gene node identification results;

[0075] Taking the power data history data set as input data, performing unsupervised learning based on the analysis benchmark rule to obtain a gene node classification output result;

[0076] According to the gene node identification result and the gene node classification output result, the proportion of the number of classified abnormal nodes is obtained;

[0077] When the proportion of the number of classified abnormal nodes for a consecutive preset number of times is less than or equal to the threshold of the proportion of the number of classified abnormal nodes, the pre-trained language model is generated.

[0078] In a feasible implementation, on the basis of constructing a single cross-level local gene regulatory network, the operations such as association analysis, correlation analysis and network construction are repeatedly performed between the metadata of the first level, the metadata of the second level, and the Nth level, between two adjacent levels of metadata, to obtain N-1 local gene regulatory networks covering all levels of metadata, that is, several local gene regulatory networks, which respectively depict the regulatory relationship between metadata attributes under a specific span, but there may be semantic separation and logical faults between them. In order to form a globally consistent, coherent and smooth semantic representation, several local gene regulatory networks are linked and integrated, and the scattered local gene regulatory networks are connected into a global, multi-level gene regulatory knowledge base, which systematically and completely describes the intrinsic semantic logic of the historical execution record.

[0079] Subsequently, the gene regulation knowledge base is established as the analysis benchmark rule for the pre-trained language model. At the same time, a historical data set of power data of a certain scale and time span is collected, and the historical data set of power data is matched and aligned with the gene regulation knowledge base, and the semantic attributes corresponding to each piece of data are identified and mapped to the corresponding gene nodes in the gene regulation knowledge base. For example, the initial association of data to nodes is achieved by using string matching and synonym expansion of attribute words in the gene regulation knowledge base, and then the confidence of the association is further confirmed by the numerical range of the attribute value, the regulation direction and other information, forming a data-node mapping relationship, and obtaining the gene node identification result. Then, the annotated historical data set of power data is used as a training sample, and the input data of the model is organized according to a certain batch size and order. Each sample consists of several data with gene node labels, reflecting the specific power regulation scenario. At the same time, the training objectives and evaluation indicators of the model are standardized and constrained by the established analysis benchmark rules, and the model is enabled to autonomously summarize the general rules of semantic mapping of power data through unsupervised learning. During the training process, the model identifies, maps, and encodes the gene nodes in the input power history data, and predicts, regulates, and optimizes the node relationship in combination with the analysis benchmark rules, outputs the category of each node, and reflects the semantic representation ability of the model in the current state, which is the gene node classification output result. After completing several rounds of unsupervised learning, the generated gene node identification results are used as the reference system for evaluation, and the output gene node classification results are used as the object system for evaluation. By comparing the differences and deviations between the two result sets, the current semantic representation ability and optimization space of the language model are examined. For each training sample, the number of inconsistent gene node identification and node classification is counted, and divided by the total number of sample nodes to obtain the proportion of classified abnormal nodes in the sample. The higher the proportion of classified abnormal nodes, the less accurate the semantic understanding and mapping of the pre-trained language model for the sample, and further learning optimization is needed; conversely, the lower the proportion of classified abnormal nodes, the better the model has mastered the semantic laws of the sample and is close to the optimal state. Then, according to application requirements and experience estimation, a threshold of the proportion of classified abnormal nodes is preset as a standard reference for judging model convergence, and an observation window is preset, that is, a preset number of consecutive times, such as 3 times, 5 times or 10 times. Only when the proportion of the number of classified abnormal nodes in the model within the preset number of consecutive times is less than or equal to the threshold of the proportion of the number of classified abnormal nodes, the model is considered to have converged and achieved the expected semantic encoding ability, thereby obtaining a pre-trained language model.

[0080] S7: Collecting power monitoring data, classifying it through the pre-trained language model, and obtaining gene node matching results;

[0081] Specifically, after completing the construction of the pre-trained language model, it is applied to actual power monitoring data.

[0082] First, through various IoT terminals, smart sensors and other collection devices, various monitoring data reflecting the operation status of the power grid are obtained in real time. These data come from multiple links such as generators, transmission lines, and distribution equipment, and are of various types as power monitoring data. Then, the collected power monitoring data are summarized and segmented according to certain time windows, spatial regions and other dimensions to form several independent data fragments, each of which corresponds to a specific power operation scenario. Next, each data fragment is input into the generated pre-trained language model. The model automatically identifies the semantic attributes of each power monitoring data in the fragment, maps it to the most relevant and matching gene node in the gene regulation knowledge base, classifies and identifies the power monitoring data, and reveals its inherent regulatory attributes and association rules. After the above processing, each monitoring data fragment is mapped to a gene node combination, which is the gene node matching result of the fragment. The gene node matching result contains the gene node corresponding to each power monitoring data, as well as the interaction mechanism and transmission path of the data in actual operation, forming a multi-level and multi-dimensional interpretation of the power monitoring data.

[0083] By using the pre-trained language model to automatically encode and semantically map the power monitoring data, and matching and associating it with the existing node system in the gene regulation knowledge base, we can achieve rapid annotation and classification of data semantics, and explore the association rules and regulation mechanisms therein to achieve accurate and efficient extraction of power data quality features.

[0084] S8: When the number of gene nodes in the gene node matching result is equal to 0, a cold start mark is performed on the power monitoring data.

[0085] Specifically, for each gene node matching result generated, check the number of successfully matched gene nodes. If this number is equal to 0, it means that the power monitoring data has not been successfully matched to any gene node, indicating that the data quality reflected by the power monitoring data segment is low, and key semantic elements are missing, resulting in the model being unable to capture its regulatory properties. At this time, these unmatched power monitoring data are marked as cold start for retrospective review and correction of the monitoring process. Among them, the specific implementation method of the cold start mark can be flexible and diverse, such as adding specific tags to the attribute table of the power monitoring data segment, and establishing a separate power monitoring data table in the database. With the introduction of the cold start mark, the ability to identify and respond to unknown working conditions is further enhanced.

[0086] In summary, the power data quality feature extraction method based on the pre-trained language model provided by the embodiment of the present invention has the following technical effects:

[0087] According to the target task type, historical execution records are collected to provide raw materials for subsequent data analysis and model training, laying the necessary data foundation. The execution process of historical execution records is split to obtain the first-level metadata, the second-level metadata, and the N-level metadata. The semantic levels of the first-level metadata, the second-level metadata, and the N-level metadata are enhanced in sequence, and the k+1-level metadata depends on the k-level metadata. N is an integer and k is an integer. By performing multi-level and shallow-to-deep metadata splitting on historical data, the inherent hierarchical structure and semantic relationship of the data are revealed, preparing for metadata association analysis. According to the k+1-level metadata, the k-level metadata is analyzed for association, and the association analysis results are obtained. By analyzing the association of adjacent metadata levels from bottom to top, the semantic connection between metadata of different granularities is established, providing a basis for building a metadata regulation network. According to the k+1-level metadata, the k-level metadata is analyzed for correlation, and the correlation analysis results are obtained. By analyzing the correlation of adjacent metadata levels from bottom to top, the similarity between metadata of different granularities is measured, providing quantitative indicators for metadata fusion. According to the results of association analysis and correlation analysis, a local gene regulatory network of the metadata of the kth level for the metadata of the k+1th level is constructed. Based on the results of association and correlation analysis, a local gene regulatory network is constructed to vividly depict the regulatory effect and influence mechanism between metadata. Repeat the cyclic analysis, fuse several local gene regulatory networks, generate a gene regulatory knowledge base, perform unsupervised configuration, generate a pre-trained language model, fuse the local regulatory network through cyclic iteration, build a complete gene regulatory knowledge base, and then train the language model in an unsupervised manner to enable it to learn the representation and extraction methods of power data quality characteristics. Collect power monitoring data, classify it through the pre-trained language model, obtain the gene node matching results, apply the pre-trained language model to the newly collected power monitoring data, automatically extract its quality characteristics, and realize fast and accurate processing of large-scale data. When the number of gene nodes in the gene node matching result is equal to 0, the power monitoring data is cold-started, and the data that cannot be classified is identified through the cold-start identification. By constructing a multi-level metadata gene regulatory network and using it to generate a pre-trained language model, the quality characteristics of power data can be accurately and efficiently extracted.

[0088] Embodiment 2

[0089] Based on the same inventive concept as the power data quality feature extraction method based on the pre-trained language model in the aforementioned embodiment, Figure 2 As shown, an embodiment of the present invention provides a power data quality feature extraction system based on a pre-trained language model, the system comprising:

[0090] The record collection module 11 is used to collect historical execution records according to the target task type;

[0091] A process splitting module 12 is used to perform execution process splitting on the historical execution record to obtain first-level metadata, second-level metadata, and up to N-level metadata, wherein the semantic levels of the first-level metadata, the second-level metadata, and up to the N-level metadata are sequentially enhanced, and the k+1-th level metadata is generated depending on the k-th level metadata, where N is an integer and k is an integer;

[0092] The association analysis module 13 is used to perform an association analysis on the k-th level metadata according to the k+1-th level metadata to obtain an association analysis result;

[0093] A similarity analysis module 14 is used to perform a correlation analysis on the k-th level metadata according to the k+1-th level metadata to obtain a correlation analysis result;

[0094] A network construction module 15 is used to construct a local gene regulation network of the k-th level metadata to the k+1-th level metadata according to the association analysis result and the correlation analysis result;

[0095] A model generation module 16 is used for repeated cycle analysis, fusing several local gene regulatory networks, generating a gene regulatory knowledge base, performing unsupervised configuration, and generating a pre-trained language model;

[0096] A data classification module 17 is used to collect power monitoring data, classify it through the pre-trained language model, and obtain gene node matching results;

[0097] The cold start identification module 18 is used to perform a cold start identification on the power monitoring data when the number of gene nodes in the gene node matching result is equal to 0.

[0098] Furthermore, the association analysis module 13 includes the following execution steps:

[0099] Extracting a first value assignment sequence of a first data attribute of the metadata at the k+1th level through the historical execution record;

[0100] By using the historical execution records, performing execution frequency analysis on the k-th level metadata based on the first data attribute, obtaining a first high-frequency data attribute, a second high-frequency data attribute, and up to an M-th high-frequency data attribute of the k-th level metadata, where M is an integer;

[0101] Extracting the first high-frequency data attribute, the second high-frequency data attribute, and the Mth high-frequency data attribute through the historical execution record and the first assignment sequence, and obtaining a second assignment sequence, a third assignment sequence, and the M+1th assignment sequence;

[0102] Taking the first value assignment sequence as the reference sequence, and taking the second value assignment sequence, the third value assignment sequence to the M+1th value assignment sequence as the comparison sequence, a grey correlation analysis is performed to generate a first data attribute correlation analysis result, which is added to the correlation analysis result.

[0103] Furthermore, the association analysis module 13 further includes the following execution steps:

[0104] According to the first data attribute, based on the historical execution record, a plurality of data attributes of the metadata of the kth level are marked with call frequencies to obtain a plurality of call frequency coefficients;

[0105] According to the calling frequency coefficient threshold, based on the calling frequency coefficients, the data attributes are sorted to obtain the first high-frequency data attribute, the second high-frequency data attribute, and so on to the Mth high-frequency data attribute.

[0106] Furthermore, the similarity analysis module 14 includes the following execution steps:

[0107] Taking the first high-frequency data attribute as an independent variable, taking the second high-frequency data attribute to the Mth high-frequency data attribute as quantitative, taking the first data attribute as a dependent variable, assigning values ​​through the historical execution record, and obtaining an independent variable assignment sequence and a dependent variable assignment sequence;

[0108] Performing a Pearson correlation coefficient analysis on the independent variable value assignment sequence and the dependent variable value assignment sequence to obtain a first Pearson correlation coefficient pointing from the first high-frequency data attribute to the first data attribute;

[0109] According to the first Pearson correlation coefficient, an enhanced identification relationship, a suppressed identification relationship or no relevant identification is obtained and added into the correlation analysis result.

[0110] Furthermore, the similarity analysis module 14 further includes the following execution steps:

[0111] Retrieve the first high-frequency data attribute, the second high-frequency data attribute, up to the first distribution interval, the second distribution interval, up to the Mth distribution interval of the Mth high-frequency data attribute;

[0112] Traversing the second distribution interval to the Mth distribution interval, extracting the interval upper limit value, performing quantitative assignment on the second high-frequency data attribute to the Mth high-frequency data attribute, and generating a first quantitative assignment result;

[0113] Traversing the second distribution interval to the Mth distribution interval, extracting the interval lower limit value, performing quantitative assignment on the second high-frequency data attribute to the Mth high-frequency data attribute, and generating a second quantitative assignment result;

[0114] Traversing the second distribution interval to the Mth distribution interval, extracting the median of the interval, performing quantitative assignment on the second high-frequency data attribute to the Mth high-frequency data attribute, and generating a third quantitative assignment result;

[0115] The first quantitative assignment result, the second quantitative assignment result and the third quantitative assignment result are respectively used as constraints, the first high-frequency data attribute is used as an independent variable, and the first data attribute is used as a dependent variable. The historical execution records are used for assignment to obtain an independent variable assignment sequence and a dependent variable assignment sequence.

[0116] Furthermore, the similarity analysis module 14 further includes the following execution steps:

[0117] The first Pearson correlation coefficient includes a first sub-Pearson correlation coefficient, a second sub-Pearson correlation coefficient, and finally a third sub-Pearson correlation coefficient;

[0118] When at least two of the first sub-Pearson correlation coefficient, the second sub-Pearson correlation coefficient, and the third sub-Pearson correlation coefficient are enhanced relations, generating the enhanced identification relation;

[0119] When at least two of the first sub-Pearson correlation coefficient, the second sub-Pearson correlation coefficient, and the third sub-Pearson correlation coefficient are in an inhibition relationship, generating the inhibition identification relationship;

[0120] Otherwise, the no-correlation flag is generated.

[0121] Furthermore, the network construction module 15 includes the following execution steps:

[0122] According to the correlation analysis result, constructing a thickness feature of a connection line from the k-th level metadata to the k+1-th level metadata, wherein the greater the correlation, the thicker the connection line;

[0123] According to the correlation analysis result, construct a line pointing mark from the k-th level metadata to the k+1-th level metadata, wherein the enhanced mark relationship is an arrow mark from the k-th level metadata to the k+1-th level metadata, the suppressed mark relationship is a horizontal line perpendicular to the line from the k-th level metadata to the k+1-th level metadata, and the unrelated mark is no mark;

[0124] The metadata from the kth level to the k+1th level are distributed to a plurality of gene nodes, and the local gene regulatory network of the metadata from the kth level to the metadata from the k+1th level is constructed according to the line thickness feature and the line pointing identifier.

[0125] Furthermore, the data classification module 17 includes the following execution steps:

[0126] Using the gene regulation knowledge base as the analysis benchmark rule of the pre-trained language model;

[0127] Collecting historical data sets of power data, marking them through the gene regulation knowledge base, and generating gene node identification results;

[0128] Taking the power data history data set as input data, performing unsupervised learning based on the analysis benchmark rule to obtain a gene node classification output result;

[0129] According to the gene node identification result and the gene node classification output result, the proportion of the number of classified abnormal nodes is obtained;

[0130] When the proportion of the number of classified abnormal nodes for a consecutive preset number of times is less than or equal to the threshold of the proportion of the number of classified abnormal nodes, the pre-trained language model is generated.

[0131] Any steps of the method described above may be stored as computer instructions or programs in an unlimited computer memory, and may be called and recognized by an unlimited computer processor to implement any method in the embodiments of the present invention, without further limitation.

[0132] Furthermore, the first or second mentioned above may not only represent an order relationship, but may also represent a specific concept, and / or refer to multiple elements that can be selected individually or in full. Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the present invention and its equivalents, the present invention is intended to include these modifications and variations.

Claims

1. A method for extracting power data quality features based on a pre-trained language model, characterized in that: include: Collect historical execution records according to the target task type; the target task type includes power equipment fault diagnosis, and the historical execution records include: historical equipment operation parameter data, alarm data and maintenance records; Performing execution process splitting on the historical execution record to obtain first-level metadata, second-level metadata, and up to N-level metadata, wherein the semantic levels of the first-level metadata, the second-level metadata, and up to the N-level metadata are sequentially enhanced, and the k+1-th level metadata is generated depending on the k-th level metadata, where N is an integer and k is an integer; Performing a correlation analysis on the k-th level metadata according to the k+1-th level metadata to obtain a correlation analysis result; Performing a correlation analysis on the k-th level metadata according to the k+1-th level metadata to obtain a correlation analysis result; Constructing a local gene regulatory network of the k-th level metadata to the k+1-th level metadata according to the association analysis result and the correlation analysis result; Repeat the cycle analysis, fuse several local gene regulatory networks, generate a gene regulatory knowledge base, perform unsupervised configuration, and generate a pre-trained language model; Collecting power monitoring data, classifying it through the pre-trained language model, and obtaining gene node matching results; When the number of gene nodes in the gene node matching result is equal to 0, a cold start mark is performed on the power monitoring data; Performing a correlation analysis on the k-th level metadata according to the k+1-th level metadata to obtain a correlation analysis result includes: Extracting a first value assignment sequence of a first data attribute of the metadata at the k+1th level through the historical execution record; By using the historical execution records, performing execution frequency analysis on the k-th level metadata based on the first data attribute, obtaining a first high-frequency data attribute, a second high-frequency data attribute, and up to an M-th high-frequency data attribute of the k-th level metadata, where M is an integer; Extracting the first high-frequency data attribute, the second high-frequency data attribute, and the Mth high-frequency data attribute through the historical execution record and the first assignment sequence, and obtaining a second assignment sequence, a third assignment sequence, and the M+1th assignment sequence; Taking the first value assignment sequence as a reference sequence, and taking the second value assignment sequence, the third value assignment sequence, and the M+1th value assignment sequence as comparison sequences, performing grey correlation analysis, generating a first data attribute correlation analysis result, and adding the result to the correlation analysis result; Performing a correlation analysis on the k-th level metadata according to the k+1-th level metadata to obtain a correlation analysis result includes: Taking the first high-frequency data attribute as an independent variable, taking the second high-frequency data attribute to the Mth high-frequency data attribute as quantitative, taking the first data attribute as a dependent variable, assigning values ​​through the historical execution record, and obtaining an independent variable assignment sequence and a dependent variable assignment sequence; Performing a Pearson correlation coefficient analysis on the independent variable value assignment sequence and the dependent variable value assignment sequence to obtain a first Pearson correlation coefficient pointing from the first high-frequency data attribute to the first data attribute; According to the first Pearson correlation coefficient, an enhanced identification relationship, a suppressed identification relationship or no related identification is obtained, and added into the correlation analysis result; According to the association analysis result and the correlation analysis result, constructing a local gene regulatory network of the k-th level metadata to the k+1-th level metadata, including: According to the correlation analysis result, constructing a thickness feature of a connection line from the k-th level metadata to the k+1-th level metadata, wherein the greater the correlation, the thicker the connection line; According to the correlation analysis result, construct a line pointing mark from the k-th level metadata to the k+1-th level metadata, wherein the enhanced mark relationship is an arrow mark from the k-th level metadata to the k+1-th level metadata, the suppressed mark relationship is a horizontal line perpendicular to the line from the k-th level metadata to the k+1-th level metadata, and the unrelated mark is no mark; Distributing the metadata from the kth level to the k+1th level to a plurality of gene nodes, and constructing the local gene regulatory network of the metadata from the kth level to the metadata from the k+1th level according to the line thickness feature and the line pointing identifier; Generate a pre-trained language model, including: Using the gene regulation knowledge base as the analysis benchmark rule of the pre-trained language model; Collecting a historical data set of power data, marking it through the gene regulation knowledge base, and generating a gene node identification result; wherein the historical data set of power data includes current, voltage, active power, and reactive power; Taking the power data history data set as input data, performing unsupervised learning based on the analysis benchmark rule to obtain a gene node classification output result; According to the gene node identification result and the gene node classification output result, the proportion of the number of classified abnormal nodes is obtained; When the proportion of the number of classified abnormal nodes for a consecutive preset number of times is less than or equal to the threshold of the proportion of the number of classified abnormal nodes, the pre-trained language model is generated.

2. The method according to claim 1, characterized in that By using the historical execution record, performing frequent analysis on the k-th level metadata based on the first data attribute, obtaining a first high-frequency data attribute, a second high-frequency data attribute, and up to an M-th high-frequency data attribute of the k-th level metadata, where M is an integer, including: According to the first data attribute, based on the historical execution record, a plurality of data attributes of the metadata of the kth level are marked with call frequencies to obtain a plurality of call frequency coefficients; According to the calling frequency coefficient threshold, based on the calling frequency coefficients, the data attributes are sorted to obtain the first high-frequency data attribute, the second high-frequency data attribute, and so on to the Mth high-frequency data attribute.

3. The method according to claim 1, characterized in that Taking the first high-frequency data attribute as an independent variable, taking the second high-frequency data attribute to the Mth high-frequency data attribute as quantitative, taking the first data attribute as a dependent variable, assigning values ​​through the historical execution record, and obtaining an independent variable assignment sequence and a dependent variable assignment sequence, including: Retrieve the first high-frequency data attribute, the second high-frequency data attribute, up to the first distribution interval, the second distribution interval, up to the Mth distribution interval of the Mth high-frequency data attribute; Traversing the second distribution interval to the Mth distribution interval, extracting the interval upper limit value, performing quantitative assignment on the second high-frequency data attribute to the Mth high-frequency data attribute, and generating a first quantitative assignment result; Traversing the second distribution interval to the Mth distribution interval, extracting the interval lower limit value, performing quantitative assignment on the second high-frequency data attribute to the Mth high-frequency data attribute, and generating a second quantitative assignment result; Traversing the second distribution interval to the Mth distribution interval, extracting the median of the interval, performing quantitative assignment on the second high-frequency data attribute to the Mth high-frequency data attribute, and generating a third quantitative assignment result; The first quantitative assignment result, the second quantitative assignment result and the third quantitative assignment result are respectively used as constraints, the first high-frequency data attribute is used as an independent variable, and the first data attribute is used as a dependent variable. The historical execution records are used for assignment to obtain an independent variable assignment sequence and a dependent variable assignment sequence.

4. The method according to claim 1, characterized in that According to the first Pearson correlation coefficient, an enhanced identification relationship, a suppressed identification relationship or a non-correlated identification is obtained and added to the correlation analysis result, including: The first Pearson correlation coefficient includes a first sub-Pearson correlation coefficient, a second sub-Pearson correlation coefficient, and finally a third sub-Pearson correlation coefficient; When at least two of the first sub-Pearson correlation coefficient, the second sub-Pearson correlation coefficient, and the third sub-Pearson correlation coefficient are enhanced relations, generating the enhanced identification relation; When at least two of the first sub-Pearson correlation coefficient, the second sub-Pearson correlation coefficient, and the third sub-Pearson correlation coefficient are in an inhibition relationship, generating the inhibition identification relationship; Otherwise, the no-correlation flag is generated.

5. The power data quality feature extraction system based on the pre-trained language model is characterized by: The method for extracting power data quality features based on a pre-trained language model according to any one of claims 1 to 4 comprises: A record collection module, which is used to collect historical execution records according to the target task type; a process splitting module, the process splitting module is used to perform execution process splitting on the historical execution record, and obtain first-level metadata, second-level metadata, and finally N-level metadata, wherein the semantic levels of the first-level metadata, the second-level metadata, and finally the N-level metadata are sequentially enhanced, and the k+1-th level metadata is generated depending on the k-th level metadata, where N is an integer and k is an integer; A correlation analysis module, wherein the correlation analysis module is used to perform a correlation analysis on the k-th level metadata according to the k+1-th level metadata to obtain a correlation analysis result; A similarity analysis module, wherein the similarity analysis module is used to perform a correlation analysis on the k-th level metadata according to the k+1-th level metadata to obtain a correlation analysis result; A network construction module, wherein the network construction module is used to construct a local gene regulation network of the k-th level metadata to the k+1-th level metadata according to the association analysis result and the correlation analysis result; A model generation module, wherein the model generation module is used to repeatedly perform cyclic analysis, fuse several local gene regulatory networks, generate a gene regulatory knowledge base, perform unsupervised configuration, and generate a pre-trained language model; A data classification module, wherein the data classification module is used to collect power monitoring data, classify it through the pre-trained language model, and obtain gene node matching results; A cold start identification module, wherein the cold start identification module is used to perform a cold start identification on the power monitoring data when the number of gene nodes in the gene node matching result is equal to 0.

Citation Information

Patent Citations

  • Gene similarity intrusion detection method based on electric power network IEC protocol

    CN109600365A

  • Genetic intelligent optimization neural network wind power generation equipment temperature prediction method

    CN114595884A

  • Analysis method for predicting potential pathogenic mechanism of single-gene hereditary disease based on phenotype semantic association gene clustering regulatory network

    CN116343913A

  • Automatic compression method and platform for multilevel knowledge distillation-based pre-trained language model

    WO2022126797A1

  • Method and apparatus for drug retargeting model generation, storage medium, and computer device

    WO2023040150A1