A method for identifying the behavior of power system terminals based on deep learning

Abstract

The present invention discloses a method for identifying terminal behavior in a power system based on deep learning, which relates to the technical field of power systems. It performs digital-to-graph conversion on the terminal data sequence, extracts feature points, and constructs a peak-point core sampling matrix; sets local boundary coefficients to perform local extraction on the peak-point core sampling matrix and conducts feature extraction with the terminal sequence coefficients to obtain a terminal feature coefficient group; performs information flow conversion on the terminal feature coefficient group and constructs a terminal vector storage library, conducts comparison on the terminal vector storage library to obtain the terminal code class distance, makes a determination and supplementation according to the terminal code class distance and restores it to the corresponding terminal feature stream segment to obtain a classified stream segment storage library, sets a threshold reference matrix to conduct behavior screening on the classified stream segment storage library to obtain abnormal data behaviors; accurately identifies abnormal behaviors of terminal devices, monitors the operating status of terminal devices in real time, and ensures the safe and stable operation of the power system.

Description

Technical Field

[0001] The present invention relates to the technical field of power systems, and specifically to a method for identifying terminal behaviors in a power system based on deep learning. Background Art

[0002] In modern society, as the core of the national energy infrastructure, with the rapid development of smart grids, the number of terminal devices in power systems has increased sharply, covering smart meters, distributed power access devices, electric vehicle chargers, etc. The behavioral characteristics of these terminal devices are complex, diverse, and dynamically changing, making the accurate identification of their behaviors a key issue for ensuring the safe, stable, and efficient operation of power systems.

[0003] Traditional methods for identifying terminal behaviors in power systems mainly rely on pre-set rules to judge terminal behaviors or analyze the statistical characteristics of historical data to identify behaviors. They have a certain effect when dealing with simple and clearly defined behaviors, but for complex and changing behavior patterns, it is difficult to formulate rules and they lack flexibility, making it difficult to meet the needs of the continuous development and change of power systems.

[0004] Deep learning technology can automatically learn deep feature representations from a large amount of raw data without the need for complex manual feature engineering. Therefore, applying deep learning technology to the identification of terminal behaviors in power systems, collecting terminal data, accurately and efficiently processing terminal data, and analyzing abnormal data in terminal data can provide early warnings of abnormal behaviors and provide strong guarantees for the safe operation of power systems. For this reason, a method for identifying terminal behaviors in a power system based on deep learning is provided. Summary of the Invention

[0005] The object of the present invention can be achieved by the following technical solutions:

[0006] A method for identifying terminal behaviors in a power system based on deep learning, comprising the following steps:

[0007] Step S1: Collect terminal comprehensive data and capture time, and construct a terminal data sequence;

[0008] Step S2: Perform digital graph extraction on the terminal data sequence to obtain terminal coefficient peak points, and perform morphological transformation according to the terminal coefficient peak points to obtain a peak point core sampling matrix;

[0009] Step S3: Set local boundary coefficients to perform local extraction on the peak point core sampling matrix to obtain a boundary peak point sampling matrix, and perform joint transformation on the boundary peak point sampling matrix and perform feature extraction with the terminal sequence coefficients to obtain a terminal feature coefficient group;

[0010] Step S4: Perform information flow conversion on the terminal feature coefficient group to obtain a terminal feature flow segment. Extract and construct the terminal feature flow segment to obtain a terminal vector repository. Conduct comparison and selection on the terminal vector repository to obtain a reference terminal vector. Traverse and compare the terminal vector repository with the reference terminal vector to obtain the terminal code class distance. Make a judgment and supplement according to the terminal code class distance to obtain a classification vector repository;

[0011] Step S5: Docking and restoring the classification vector repository according to the terminal feature flow segment to obtain a classification flow segment repository. Set a terminal data threshold for form replacement to obtain a threshold reference matrix. Screen the behavior of the classification flow segment repository according to the threshold reference matrix to obtain abnormal data behavior.

[0012] Preferably, the process of collecting terminal comprehensive data includes:

[0013] Comprehensively collect the power system terminals to obtain terminal comprehensive data, and mark the time of the comprehensively collected terminal comprehensive data to obtain the capture time;

[0014] Sort the terminal comprehensive data according to the obtained capture time to obtain a terminal data sequence.

[0015] Preferably, the process of performing digital graph extraction on the terminal data sequence includes:

[0016] Convert the coefficients of the terminal data sequence to obtain terminal sequence coefficients, and convert the obtained terminal sequence coefficients into a frequency spectrum diagram to obtain a terminal coefficient frequency spectrum diagram;

[0017] Traverse the peak points according to the obtained terminal coefficient frequency spectrum diagram to obtain terminal coefficient peak points.

[0018] Preferably, the process of performing morphological transformation according to the terminal coefficient peak points includes:

[0019] Sort the obtained terminal coefficient peak points to obtain a coefficient peak point sorting;

[0020] Set extraction conditions according to the obtained coefficient peak point sorting, and extract peak points from the coefficient peak point sorting based on the extraction conditions to obtain sample coefficient peak points;

[0021] Construct an original peak point sampling matrix based on the order of peak point extraction, and perform core filling on the original peak point sampling matrix according to the obtained sample coefficient peak points to obtain a peak point core sampling matrix.

[0022] Preferably, the process of performing joint transformation on the boundary peak point sampling matrix and extracting features with the terminal sequence coefficients includes:

[0023] Set the local boundary coefficient according to the peak point core sampling array, perform local extraction on the peak point core sampling array according to the local boundary coefficient, and obtain the boundary peak point sampling array;

[0024] Perform adaptive adjustment on the boundary peak point sampling array to obtain a linear sampling segment, and perform sequential combination on the linear sampling segments based on the order of adaptive adjustment to obtain a linear boundary sampling group;

[0025] Upload the obtained linear boundary sampling group to the terminal sequence coefficient, and perform cyclic screening on the terminal sequence coefficient through the linear boundary sampling group to obtain a terminal characteristic coefficient group.

[0026] Preferably, the process of performing cyclic screening on the terminal sequence coefficient through the linear boundary sampling group includes:

[0027] Set the starting endpoint according to the terminal sequence coefficient, perform initial docking on the linear boundary sampling group based on the starting endpoint, and upload the linear boundary sampling group to the docking position;

[0028] Perform docking extraction on the terminal sequence coefficient through the linear boundary sampling group to obtain a segmented boundary coefficient;

[0029] Set the sliding distance according to the starting endpoint, perform sequential sliding on the linear boundary sampling group based on the terminal sequence coefficient according to the sliding distance, reach the next docking position, and perform docking extraction on the terminal sequence coefficient through the linear boundary sampling group to obtain a segmented boundary coefficient;

[0030] Perform sequential sliding on the linear boundary sampling group based on the terminal sequence coefficient according to the sliding distance until the terminal sequence coefficient is covered;

[0031] Combine the segmented boundary coefficients according to the order of sequential sliding to obtain a terminal characteristic coefficient group.

[0032] Preferably, the process of extracting and constructing the terminal characteristic stream segment includes:

[0033] Perform code class separation on the terminal characteristic stream segment to obtain the number of sub - coding times;

[0034] Perform code number induction on the obtained terminal characteristic stream segment to obtain the comprehensive characteristic code quantity;

[0035] Based on the terminal characteristic stream segment, perform separation and extraction of the number of sub - coding times according to the obtained comprehensive characteristic code quantity to obtain the characteristic code frequency, and construct a characteristic code vector according to the characteristic code frequency;

[0036] Set the capture period according to the obtained capture time, perform comprehensive statistics on the obtained characteristic code vector based on the capture period, and construct a terminal vector storage library.

[0037] Preferably, the process of determining and supplementing according to the terminal code class distance includes:

[0038] Set a fit threshold, determine the deviation of the terminal code class distance according to the fit threshold, obtain a matching feature vector, and construct a conforming vector set for the obtained matching feature vector;

[0039] Perform a comparison and screening on the terminal vector repository according to the conforming vector set to obtain a screened vector repository;

[0040] Based on the reference terminal vector, perform feature supplementation on the conforming vector set according to the obtained screened vector repository to obtain a classified vector repository.

[0041] Preferably, the process of setting a terminal data threshold for form replacement includes:

[0042] Based on the classified segment repository, set a terminal data threshold according to the terminal comprehensive data, perform a standard inspection on the terminal data threshold to obtain a standard threshold segment;

[0043] Based on the classified segment repository, perform encoding and segmentation on the terminal feature segment to obtain a terminal feature code element segment;

[0044] Perform format replacement on the obtained standard threshold segment to obtain a threshold reference matrix.

[0045] Preferably, the process of performing behavior screening on the classified segment repository according to the threshold reference matrix includes:

[0046] Based on the classified segment repository, perform an analogical judgment on the terminal feature code element segment according to the threshold reference matrix to obtain a reference judgment result;

[0047] Perform a forward match on the terminal feature segment according to the obtained reference judgment result to obtain abnormal terminal data, and perform behavior analogy according to the obtained abnormal terminal data to obtain abnormal data behavior.

[0048] Compared with the prior art, the beneficial effects of the present invention are:

[0049] 1. Combine the collected terminal data of the power system into a terminal data sequence in chronological order. By converting it into the form of a signal spectrogram, it is possible to capture the subtle and key feature differences in the spectrogram, extract feature segments in the spectrogram to construct a peak point core sampling matrix to extract features from the terminal data sequence, obtain a terminal feature coefficient group, and reduce the computational complexity by reducing the data dimension, greatly improving the efficiency of data processing;

[0050] 2. Capture the similarity of the terminal characteristic coefficient groups, classify them according to the similarity between different terminal characteristic coefficient groups to obtain a classification vector repository, classify the status of the device, facilitate the early maintenance of the device or the replacement of key components, avoid sudden device failures from affecting power supply, screen the classification vector repository by setting terminal data thresholds to obtain abnormal data; the abnormal data obtained through screening can early warn of possible device failures, gain valuable time for the maintenance and management of the power system, and ensure the safe and stable operation of the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0052] Figure 1 is the schematic diagram of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0053] The following will clearly and completely describe the technical solutions of the present invention in conjunction with the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0054] As Figure 1 shown, a method for identifying terminal behavior in a power system based on deep learning includes the following steps:

[0055] Step S1: Collect terminal comprehensive data and capture time to construct a terminal data sequence;

[0056] Step S2: Perform digital graph extraction on the terminal data sequence to obtain terminal coefficient peak points, and perform morphological transformation according to the terminal coefficient peak points to obtain a peak point core sampling matrix;

[0057] Step S3: Set local boundary coefficients to perform local extraction on the peak point core sampling matrix to obtain a boundary peak point sampling matrix. Through joint transformation of the boundary peak point sampling matrix and feature extraction with the terminal sequence coefficients, obtain a terminal characteristic coefficient group;

[0058] Step S4: Perform information flow conversion on the terminal feature coefficient group to obtain a terminal feature flow segment. Extract and construct the terminal feature flow segment to obtain a terminal vector repository. Conduct a comparison and selection on the terminal vector repository to obtain a reference terminal vector. Traverse and compare the terminal vector repository with the reference terminal vector to obtain a terminal code class distance. Make a determination and supplement based on the terminal code class distance to obtain a classification vector repository;

[0059] Step S5: Perform docking and restoration on the classification vector repository according to the terminal feature flow segment to obtain a classification flow segment repository. Set a terminal data threshold for form replacement to obtain a threshold reference matrix. Conduct behavior screening on the classification flow segment repository according to the threshold reference matrix to obtain abnormal data behaviors.

[0060] It should be further noted that in the specific implementation process, the process of collecting terminal comprehensive data includes:

[0061] Comprehensively collect data from power system terminals to obtain terminal comprehensive data;

[0062] The terminal comprehensive data includes electrical quantity data, equipment status data, and environmental data. Among them, the electrical quantity data includes, but is not limited to, voltage, current, power, and frequency. The equipment status data includes, but is not limited to, switch status and running time. The environmental data includes, but is not limited to, temperature and humidity;

[0063] Mark the time of the comprehensively collected terminal comprehensive data to obtain the capture time;

[0064] Associate the obtained terminal comprehensive data with the corresponding capture time;

[0065] Sort the terminal comprehensive data according to the obtained capture time to obtain a terminal data sequence;

[0066] The time sorting means sorting according to the time order of the capture time of each terminal comprehensive data to obtain a time series sequence, which is the terminal data sequence. Then, according to the electrical quantity data, equipment status data, and environmental data included in the terminal comprehensive data, the terminal data sequence includes an electrical quantity data sequence, an equipment status data sequence, and an environmental data sequence.

[0067] Extract digital graphs from the terminal data sequence to obtain terminal coefficient peak points. The specific process includes:

[0068] Obtain the terminal data sequence, perform coefficient conversion on the obtained terminal data sequence to obtain a terminal sequence coefficient;

[0069] The coefficient conversion means converting the obtained terminal data sequence into a signal form, that is, the terminal sequence coefficient. According to the electrical quantity data sequence, device status data sequence, and environmental data sequence included in the terminal data sequence, the terminal sequence coefficient includes an electrical quantity sequence coefficient, a device status sequence coefficient, and an environmental sequence coefficient;

[0070] Perform a spectrogram conversion on the obtained terminal sequence coefficient to obtain a terminal coefficient spectrogram;

[0071] The spectrogram conversion means generating a spectrogram form of the signal based on the obtained terminal sequence coefficient, that is, the terminal coefficient spectrogram;

[0072] Furthermore, according to the electrical quantity sequence coefficient, device status sequence coefficient, and environmental sequence coefficient included in the terminal sequence coefficient, the terminal coefficient spectrogram includes an electrical quantity coefficient spectrogram, a device status coefficient spectrogram, and an environmental coefficient spectrogram;

[0073] Perform a peak point traversal on the obtained terminal coefficient spectrogram to obtain terminal coefficient peak points;

[0074] The peak point traversal means recording the spectrum at the intersection of the horizontal axis and the vertical axis in the terminal coefficient spectrogram as the peak point, and marking the obtained peak point as the terminal coefficient peak point.

[0075] Perform a morphological transformation based on the terminal coefficient peak points to obtain a peak point core sampling array. The specific process includes:

[0076] Sort the obtained terminal coefficient peak points in descending order to obtain a coefficient peak point sorting. Here, "order" means the values at the intersection of the terminal coefficient peak points and the vertical coordinate are sorted in descending order;

[0077] Set extraction conditions based on the obtained coefficient peak point sorting, and perform peak point extraction on the coefficient peak point sorting based on the extraction conditions to obtain sample coefficient peak points;

[0078] The extraction condition means the rule for selecting terminal coefficient peak points set for the coefficient peak point sorting. The extraction condition is to select the 1st, a-th, (a + 5)-th, (a + 10)-th, (a + 20)-th, -th, -th, -th, and n-th in the coefficient peak point sorting, where n is the last one of the terminal coefficient peak points in the coefficient peak point sorting, that is, there are n terminal coefficient peak points in the coefficient peak point sorting. a is a positive integer greater than 1, and the selection of a is related to the number of terminal coefficient peak points in the coefficient peak point sorting and must satisfy ;

[0079] Construct an original peak point sampling matrix based on the extraction order of peak points. The manifestation form of the original peak point sampling matrix is a 3-row and 3-column matrix;

[0080] Perform core filling on the original peak point sampling matrix according to the obtained sample coefficient peak points to obtain a peak point core sampling matrix;

[0081] The core filling means uploading the sample coefficient peak points to the original peak point sampling matrix in the order obtained by peak point extraction in sequence, and replacing the elements at the original positions to obtain a peak point core sampling matrix. Among them, the first sample coefficient peak point is uploaded to the position of the first row and the first column of the original peak point sampling matrix, the a-th sample coefficient peak point is uploaded to the position of the first row and the second column of the original peak point sampling matrix, the (a + 5)-th sample coefficient peak point is uploaded to the position of the first row and the third column of the original peak point sampling matrix, the (a + 10)-th sample coefficient peak point is uploaded to the position of the second row and the first column of the original peak point sampling matrix, the (a + 20)-th sample coefficient peak point is uploaded to the position of the second row and the second column of the original peak point sampling matrix, the -th sample coefficient peak point is uploaded to the position of the second row and the third column of the original peak point sampling matrix, the -th sample coefficient peak point is uploaded to the position of the third row and the first column of the original peak point sampling matrix, the -th sample coefficient peak point is uploaded to the position of the third row and the second column of the original peak point sampling matrix, and the n-th sample coefficient peak point is uploaded to the position of the third row and the third column of the original peak point sampling matrix.

[0082] Obtain the peak point core sampling matrix, and set local boundary coefficients according to the peak point core sampling matrix. The manifestation form of the local boundary coefficients is in the form of a function; in particular, in this embodiment, the local boundary coefficients refer to wavelet functions, where the wavelet functions include but are not limited to Haar wavelets, Symlets wavelets, and Morlet wavelets;

[0083] Perform local extraction on the peak point core sampling matrix according to the obtained local boundary coefficients to obtain a boundary peak point sampling matrix;

[0084] The local extraction means multiplying the obtained local boundary coefficients with each element in the peak point core sampling matrix respectively to obtain a boundary peak point sampling matrix, and denoting the corresponding position elements after multiplying with the local boundary coefficients as local coefficient peak points;

[0085] Perform adaptive adjustment on the obtained boundary peak point sampling matrix to obtain a linear sampling segment, and the linear sampling segment includes a first sampling segment, a second sampling segment, and a third sampling segment;

[0086] The adaptation adjustment means splitting the boundary peak point sampling array into groups with each row as a group. Each row after splitting is denoted as a linear sampling segment. Then the first row is denoted as the first sampling segment, the second row as the second sampling segment, and the third row as the third sampling segment;

[0087] Perform sequential combination on the obtained linear sampling segments based on the order of adaptation adjustment to obtain a linear boundary sampling group;

[0088] The sequential combination means combining the first sampling segment, the second sampling segment, and the third sampling segment in sequence to obtain a vector with one row and nine columns, denoted as the linear boundary sampling group;

[0089] Upload the obtained linear boundary sampling group to the terminal sequence coefficient, and perform cyclic screening on the terminal sequence coefficient through the linear boundary sampling group to obtain a terminal characteristic coefficient group;

[0090] It should be further noted that in the specific implementation process, the process of the cyclic screening includes:

[0091] Set a starting endpoint according to the obtained terminal sequence coefficient. Here, the starting endpoint represents the first element of the terminal sequence coefficient, that is, the element at the first sampling point of the signal;

[0092] Perform initial docking on the linear boundary sampling group based on the starting endpoint and upload the linear boundary sampling group to the docking position. The initial docking means aligning the element in the first row and first column of the boundary peak point sampling array with the starting endpoint, and the docking position represents the aligned element in the first row and first column;

[0093] Perform docking extraction on the terminal sequence coefficient through the linear boundary sampling group to obtain a segmented boundary coefficient;

[0094] The docking extraction means multiplying the elements in the linear boundary sampling group with the corresponding position elements of the terminal sequence coefficient and summing the multiplication results at each position to obtain the segmented boundary coefficient. Then the segmented boundary coefficient represents the characteristic coefficient corresponding to the starting endpoint position;

[0095] Set a sliding distance according to the obtained starting endpoint. The sliding distance represents the distance that the linear boundary sampling group moves to the next docking position on the terminal sequence coefficient. In this embodiment, the sliding distance is one sampling point apart. That is, the next docking position one sliding distance apart from the starting endpoint is the second element of the terminal sequence coefficient, that is, the element at the second sampling point of the signal;

[0096] Based on the terminal sequence coefficients, perform sequential sliding on the linear boundary sampling group according to the obtained sliding distance, reach the next docking position, and extract the docking of the terminal sequence coefficients through the linear boundary sampling group to obtain segmented boundary coefficients. Here, sequential sliding means moving the linear boundary sampling group by the position of one sliding distance, and reaching the next docking position is the second element of the terminal sequence coefficients;

[0097] Based on the terminal sequence coefficients, perform sequential sliding on the linear boundary sampling group according to the obtained sliding distance until the terminal sequence coefficients are covered, that is, move the linear boundary sampling group to the third element of the terminal sequence coefficients, the fourth element of the terminal sequence coefficients, the fifth element of the terminal sequence coefficients,..., the last element of the terminal sequence coefficients at one time. Here, "the last element of the terminal sequence coefficients" means that the last element of the linear boundary sampling group coincides with the last element of the terminal sequence coefficients, that is, reaching the end point;

[0098] According to the order of sequential sliding, combine the obtained segmented boundary coefficients to obtain a terminal feature coefficient group;

[0099] The combination means that according to the order of sequential sliding, for each movement of one sliding distance, the segmented boundary coefficients at the corresponding positions are obtained, and then the segmented boundary coefficients are merged in sequence to obtain a terminal feature coefficient group. For example, if the segmented boundary coefficients obtained according to the sequential sliding are h1, h2, h3, h4, h5, h6, h7 in sequence, then the terminal feature coefficient group is [h1, h2, h3, h4, h5, h6, h7];

[0100] In particular, the characteristic information of the terminal sequence coefficients is extracted in matrix form, which greatly speeds up the calculation speed and improves the feature extraction speed. At the same time, through local feature extraction, it is beneficial to capture more subtle feature information of the data, improve the accuracy of data analysis, and facilitate the identification of the behaviors corresponding to the data.

[0101] Obtain a terminal feature coefficient group, perform information flow conversion on the obtained terminal feature coefficient group to obtain a terminal feature flow segment;

[0102] The information flow conversion means converting the terminal feature coefficient group into a data stream composed of binary code elements, and the encoding rules for converting the terminal feature coefficient groups corresponding to each power system are the same;

[0103] Perform extraction and construction on the terminal feature flow segment to obtain a terminal vector storage library. The specific process includes:

[0104] Perform code class separation on the obtained terminal feature flow segment to obtain the number of code separations, and the number of code separations includes the zero code frequency and the non-zero code frequency;

[0105] The above-mentioned code class separation means to count the number of occurrences of binary code elements with a value of 1 in the terminal feature stream segment to obtain the non-zero code frequency, and count the number of occurrences of binary code elements with a value of 0 to obtain the zero code frequency;

[0106] Perform code number induction on the obtained terminal feature stream segment to obtain the comprehensive feature code quantity;

[0107] The above-mentioned code number induction means to count the number of binary code elements in the terminal feature stream segment to obtain the comprehensive feature code quantity;

[0108] Based on the terminal feature stream segment, separate and extract the number of sub-coding times according to the obtained comprehensive feature code quantity to obtain the feature code frequency, and the feature code frequency includes the zero code frequency and the non-zero code frequency;

[0109] Mark the obtained zero code frequency as F0, where, , represents the zero code frequency, and Z represents the comprehensive feature code quantity;

[0110] Mark the obtained non-zero code frequency as F1, where, , represents the non-zero code frequency;

[0111] Construct a feature code vector according to the obtained feature code frequency, and associate the obtained feature code vector with the corresponding terminal feature stream segment, indicating that each terminal feature stream segment has a corresponding feature code vector;

[0112] The above-mentioned vector construction means to form a two-dimensional vector according to the obtained zero code frequency and non-zero code frequency, which is the feature code vector, and mark the obtained feature code vector as , where, ;

[0113] Set a capture period according to the obtained capture time. The capture period is a time period during which the corresponding terminal data sequence is obtained. In order to make the data analysis more accurate, a sufficient number of terminal data sequences are required. Then the set capture period is long enough to obtain enough terminal data sequences for analysis reference;

[0114] Based on the capture period, perform comprehensive statistics on the obtained feature code vectors, and construct a terminal vector repository. Upload the obtained feature code vectors to the terminal vector repository, and record the feature code vectors in the terminal vector repository as , where i represents the number of the feature code vector in the terminal vector repository, i = 1, 2, 3,..., v1, and v1 is a positive integer.

[0115] Perform a comparison selection on the obtained terminal vector repository to obtain a reference terminal vector, and denote the obtained reference terminal vector as ;

[0116] The said comparison selection means randomly selecting a feature code vector in the terminal vector repository as the reference terminal vector, which is used as a reference vector to compare with the remaining vectors in the terminal vector repository, and dividing the vector types according to the comparison results;

[0117] Perform a traversal comparison on the terminal vector repository according to the obtained reference terminal vector to obtain the terminal code class distance;

[0118] Furthermore, the traversal comparison compares the reference terminal vector with each feature code vector in the terminal vector repository respectively, obtains the terminal code class distance corresponding to each feature code vector between the reference terminal vector and the terminal vector repository, and marks the obtained terminal code class distance as , where , represents the terminal code class distance between the reference terminal vector and the i-th feature code vector in the terminal vector repository;

[0119] In particular, when the reference terminal vector is compared with each feature code vector in the terminal vector repository respectively, it does not include the comparison with the reference terminal vector itself in the terminal vector repository;

[0120] Set a matching threshold, which is a threshold set according to the terminal code class distance, and is used to determine whether the distance between the reference terminal vector and the feature code vector in the terminal vector repository is within the specified threshold, that is, to determine whether which feature code vectors between the reference terminal vector and the terminal vector repository have a high similarity through the matching threshold;

[0121] Perform a deviation determination on the terminal code class distance according to the obtained matching threshold to obtain the matching feature vectors, and construct a conforming vector set for the obtained matching feature vectors, and upload the obtained matching feature vectors and the reference terminal vector to the conforming vector set;

[0122] The said deviation determination means denoting the feature code vectors corresponding to the terminal code class distances less than or equal to the matching threshold as the matching feature vectors;

[0123] Perform a comparison and screening on the terminal vector repository according to the obtained conforming vector set to obtain a screened vector repository;

[0124] The said comparison and screening means deleting the feature code vectors included in the conforming vector set in the terminal vector repository, and the remaining feature code vectors constitute the screened vector repository;

[0125] Based on the benchmark terminal vector, perform feature supplementation on the compliant vector set according to the obtained screening vector repository to obtain a classification vector repository;

[0126] It should be further noted that in the specific implementation process, the process of the feature supplementation includes:

[0127] Perform a comparison and selection on the compliant vector set to obtain a benchmark terminal vector, and repeat the process of obtaining the terminal code class distance for the screening vector repository according to the obtained benchmark terminal vector;

[0128] Perform a deviation determination on the obtained terminal code class distance according to the fit threshold, upload the feature code vector determined to be a matching feature vector to the compliant vector set, and perform a comparison and screening on the feature code vector determined to be a matching feature vector in the screening vector repository to obtain a secondarily supplemented compliant vector set, which represents a feature value vector with a qualified similarity to the benchmark terminal vector selected for the first time, that is, the terminal data sequences corresponding to the feature value vectors have a very high similarity and are determined to be of the same class. The secondarily supplemented compliant vector set represents continuing to match similar feature code vectors in the screening vector repository until all similar feature code vectors of the same class are supplemented into the compliant vector set. Then, a supplemented complete compliant vector set is recorded as a classification vector repository;

[0129] Continue to perform a comparison and selection in the secondarily supplemented compliant vector set to obtain a benchmark terminal vector, and repeat the process of obtaining the secondarily supplemented compliant vector set to obtain a thirdly supplemented compliant vector set;

[0130] Until there are no feature code vectors in the screening vector repository that meet the fit threshold, the feature supplementation of the compliant vector set is completed. Select any one in the remaining screening vector repository as the benchmark terminal vector, classify another type of feature code vector, and repeat the process of obtaining the classification vector repository to obtain another type of feature code vector until all feature code vectors in the terminal vector repository are supplemented into the corresponding compliant vector sets, that is, the classification of the feature code vectors is completed, and the classification of the terminal data sequences corresponding to each feature code vector is completed;

[0131] In particular, the same feature code vector is not selected for each comparison and selection, the feature code vector selected as the benchmark terminal vector cannot be repeatedly selected, and each time the compliant vector set is supplemented, the range of the fit threshold needs to be narrowed to ensure a high degree of similarity between the feature code vectors in the supplemented screening vector repository and the feature code vectors in the compliant vector set.

[0132] Perform docking and restoration on the classification vector repository according to the terminal feature stream segment to obtain a classification stream segment repository;

[0133] The docking reduction means obtaining the terminal feature stream segments associated with each feature code vector in the classification vector repository, and replacing the corresponding feature code vectors with the associated terminal feature stream segments to obtain the classification stream segment repository. Each classification stream segment repository represents terminal data sequences with high similarity, but it does not exclude that all are qualified terminal data sequences. It is necessary to perform anomaly recognition on the terminal data sequences to obtain abnormal terminal data, so that abnormal behaviors of the power system terminals can be determined based on the abnormal terminal data, that is, abnormal behaviors are identified through abnormal data;

[0134] Based on the classification stream segment repository, set the terminal data threshold according to the obtained terminal comprehensive data, and conduct a standardized investigation on the obtained terminal data threshold to obtain the standardized threshold stream segment;

[0135] It should be further noted that in the specific implementation process, the terminal data threshold represents setting the corresponding normal threshold according to the terminal comprehensive data corresponding to the terminal feature stream segments in the classification stream segment repository. That is, the set threshold is within the safe range and will not cause the comprehensive data generated by abnormal behaviors of the terminal, and it is denoted as the terminal data threshold. Then, for each classification stream segment repository, there is a corresponding standardized threshold stream segment, which is used to determine whether there are abnormal situations in the terminal feature stream segments in the classification stream segment repository; the standardized investigation means performing the same feature extraction on the terminal data threshold as on the terminal feature stream segments to obtain the standardized threshold stream segment with the same form, so as to conduct anomaly investigation on the classification stream segment repository through the standardized threshold stream segment. The specific process includes:

[0136] Perform coefficient conversion on the terminal data threshold to obtain the terminal threshold coefficient;

[0137] Upload the obtained linear boundary sampling group to the terminal threshold coefficient, and perform cyclic screening on the terminal threshold coefficient through the linear boundary sampling group to obtain the terminal threshold coefficient group;

[0138] Perform information flow conversion on the obtained terminal threshold coefficient group to obtain the standardized threshold stream segment;

[0139] Based on the classification stream segment repository, perform encoding segmentation on the obtained terminal feature stream segments to obtain the terminal feature code element segments;

[0140] The encoding segmentation means grouping in the terminal feature stream segments in groups of 7 binary code elements, and denoting them as terminal feature code element segments; in particular, if the last group has less than 7 binary code elements, zero code elements are supplemented at the front of the last group until 7 binary code elements are supplemented;

[0141] Perform format replacement on the obtained standardized threshold stream segment to obtain the threshold reference matrix;

[0142] Furthermore, the process of the format replacement includes:

[0143] Encode and segment the canonical threshold stream segment to obtain a threshold symbol segment, which means grouping the canonical threshold stream segment into groups of seven binary symbols, and supplementing the groups with less than seven binary symbols until each group has seven binary symbols, to obtain a threshold reference matrix;

[0144] Separate the symbols within each group of the obtained threshold symbol segment to obtain fixed symbols and variable symbols;

[0145] The within-group separation means denoting the first four binary symbols of the threshold symbol segment as fixed symbols and the last three binary symbols as variable symbols;

[0146] Based on the canonical threshold stream segment, perform joint matching on the obtained threshold symbol segment to obtain a joint threshold matrix;

[0147] The joint matching means constructing a matrix in the canonical threshold stream segment with three threshold symbol segments as a group, generating a matrix with three rows and seven columns, denoted as the joint threshold matrix. If the last group has less than three threshold symbol segments, then combine the last three threshold symbol segments in the canonical threshold stream segment into a matrix, denoted as the joint threshold matrix;

[0148] Perform array transformation on the joint threshold matrix according to the fixed symbols and variable symbols to obtain a threshold reference matrix;

[0149] The array transformation means denoting the three-row and four-column matrix corresponding to the fixed symbols in the joint threshold matrix as the fixed symbol group, and denoting the three-row and three-column matrix corresponding to the variable symbols as the variable symbol group. Then perform modulo 2 addition on the rows of the variable symbol group based on the joint threshold matrix until the variable symbol group becomes an identity matrix. At the same time, record the values of the fixed symbol group after the change, and form a threshold adjustment matrix with the identity matrix. Then perform matrix transpose operation on the threshold adjustment matrix to obtain the threshold reference matrix, where the identity matrix is a three-row and three-column matrix with elements of 1 only in the first row and first column, the second row and second column, and the third row and third column, and the remaining elements are 0;

[0150] Based on the classification stream segment repository, perform analogical judgment on the terminal feature symbol segment according to the obtained threshold reference matrix to obtain a reference judgment result, and the reference judgment result includes a compliant symbol segment and an abnormal symbol segment;

[0151] It should be further noted that in the specific implementation process, the process of the analogical judgment includes:

[0152] In the classification stream segment repository, in the upload order of the terminal feature stream segments, sequentially compare the first terminal feature stream segment with the threshold reference matrix to obtain a feature threshold segment;

[0153] The sequential comparison means that the first terminal feature code element segment of the terminal feature stream segment is subjected to a matrix multiplication operation with the threshold reference matrix to obtain a feature threshold segment, then the second terminal feature code element segment is subjected to a matrix multiplication operation with the threshold reference matrix to obtain a feature threshold segment, and then the third, fourth,... until all the terminal feature code element segments have completed the matrix multiplication operation with the threshold reference matrix, and the obtained feature threshold segments are statistically analyzed;

[0154] Segment code decision is performed on the obtained feature threshold segments to obtain a reference decision result;

[0155] The segment code decision means that when the feature threshold segment is a zero matrix, it is recorded as a compliant code element segment; when the feature threshold segment is a non-zero matrix, it is recorded as an abnormal code element segment;

[0156] Based on the classification stream segment repository, in the upload order of the terminal feature stream segments, the next terminal feature stream segment and the threshold reference matrix repeat the process of obtaining the reference decision result until all the terminal feature stream segments in the classification stream segment repository have been judged with the threshold reference matrix to obtain the reference decision result, then the analogical decision is completed;

[0157] Forward matching is performed on the terminal feature stream segment according to the obtained abnormal code element segment to obtain abnormal terminal data, and behavior analogy is performed according to the obtained abnormal terminal data to obtain abnormal data behavior;

[0158] The forward matching means that when there is an abnormal code element segment in the terminal feature stream segment, the terminal feature stream segment is recorded as abnormal terminal data, which represents the terminal comprehensive data corresponding to the terminal feature stream segment within an abnormal range, that is, the terminal comprehensive data outside the safe range, indicating that there is an abnormal situation in the comprehensive data generated by the terminal behavior and early warning is required to avoid the safety hazards of the power system caused by the abnormal terminal data. And according to the acquisition point position of the terminal comprehensive data, the problem can be quickly located, so as to take measures to solve the problem in time, ensure the normal power consumption of users, and improve the power supply reliability;

[0159] On the contrary, it means that the terminal feature stream segment is in a normal situation, that is, the terminal comprehensive data corresponding to the terminal feature stream segment is within the normal range, there is no abnormal data, indicating that there is no abnormal situation in the comprehensive data generated by the terminal behavior.

[0160] The preferred embodiments of the present invention disclosed above are only used to help illustrate the present invention. The preferred embodiments do not elaborate all the details, nor limit the invention to only the specific embodiments. Obviously, many modifications and variations can be made according to the content of this specification. These embodiments are selected and specifically described in this specification to better explain the principle and practical application of the present invention, so that those skilled in the art can well understand and utilize the present invention. The present invention is only limited by the claims and their full scope and equivalents.

Claims

1. A method for identifying the behavior of power system terminals based on deep learning, characterized in that, It includes the following steps: Step S1: Collect terminal comprehensive data and capture time. The terminal comprehensive data includes electrical quantity data, equipment status data, and environmental data, and construct a terminal data sequence; Step S2: Perform digital graph extraction on the terminal data sequence to obtain terminal coefficient peak points; The process of performing digital graph extraction on the terminal data sequence includes: Perform coefficient conversion on the terminal data sequence to obtain terminal sequence coefficients. The coefficient conversion means converting the terminal data sequence into a signal form; Perform spectrogram conversion on the obtained terminal sequence coefficients to obtain a terminal coefficient spectrogram; Perform peak point traversal based on the obtained terminal coefficient spectrogram to obtain terminal coefficient peak points; Perform morphological transformation based on the terminal coefficient peak points to obtain a peak point core sampling matrix; Step S3: Set local boundary coefficients to perform local extraction on the peak point core sampling matrix to obtain a boundary peak point sampling matrix. By performing joint transformation on the boundary peak point sampling matrix and performing feature extraction with the terminal sequence coefficients, obtain a terminal feature coefficient group; Step S4: Perform information flow conversion on the terminal feature coefficient group to obtain a terminal feature flow segment. The information flow conversion means converting the terminal feature coefficient group into a data stream composed of binary code elements; Perform extraction and construction on the terminal feature flow segment to obtain a terminal vector storage library; The process of performing extraction and construction on the terminal feature flow segment includes: Perform code class separation on the terminal feature flow segment to obtain the number of sub - codes. The code class separation means counting the number of occurrences of binary code elements 1 or 0 in the terminal feature flow segment; Perform code number induction on the obtained terminal feature flow segment to obtain the comprehensive feature code quantity. The code number induction means counting the number of binary code elements in the terminal feature flow segment; Based on the terminal feature flow segment, perform separation and extraction of the number of sub - codes according to the obtained comprehensive feature code quantity to obtain the feature code frequency. Construct a feature code vector according to the feature code frequency; Set a capture period according to the obtained capture time. Based on the capture period, perform comprehensive statistics on the obtained feature code vectors to construct a terminal vector storage library; Perform a control selection on the terminal vector repository to obtain a reference terminal vector. Traverse and compare the terminal vector repository with the reference terminal vector to obtain the terminal code class distance, and mark the obtained terminal code class distance as , where , represents the terminal code class distance between the reference terminal vector and the i-th feature code vector in the terminal vector repository. The feature code vectors in the terminal vector repository are denoted as , and the reference terminal vector is denoted as ; Perform determination and supplementation according to the terminal code class distance to obtain a classification vector storage library; The process of performing determination and supplementation according to the terminal code class distance includes: Set a fit threshold. Perform deviation determination on the terminal code class distance according to the fit threshold to obtain a matching feature vector, and construct a conforming vector set for the obtained matching feature vector; Perform comparison and screening on the terminal vector storage library according to the conforming vector set to obtain a screened vector storage library; Based on the reference terminal vector, perform feature supplementation on the conforming vector set according to the obtained screened vector storage library to obtain a classification vector storage library; Step S5: Perform docking and restoration on the classification vector storage library according to the terminal feature flow segment to obtain a classified flow segment storage library. The docking and restoration means obtaining the terminal feature flow segment associated with each feature code vector in the classification vector storage library, and replacing the corresponding feature code vector with the associated terminal feature flow segment to obtain a classified flow segment storage library; Set a terminal data threshold for form replacement to obtain a threshold reference matrix. Perform behavior screening on the classified flow segment storage library according to the threshold reference matrix to obtain abnormal data behaviors.

2. The method for identifying the behavior of a power system terminal based on deep learning according to claim 1, wherein The process of collecting comprehensive data of the acquisition terminal includes: Comprehensively collect the power system terminal to obtain the comprehensive terminal data, and mark the time of the comprehensively collected comprehensive terminal data to obtain the capture time; Sort the terminal comprehensive data according to the obtained capture time to obtain the terminal data sequence.

3. A method for identifying the behavior of a power system terminal based on deep learning according to claim 1, characterized in that The process of morphological transformation according to the peak points of the terminal coefficient includes: Sort the obtained peak points of the terminal coefficient to obtain the sorted peak points of the coefficient; Set the extraction conditions according to the obtained sorted peak points of the coefficient, and extract the peak points based on the extraction conditions for the sorted peak points of the coefficient to obtain the sample peak points of the coefficient; Construct the original peak point sampling array based on the order of peak point extraction, and perform core filling on the original peak point sampling array according to the obtained sample peak points of the coefficient to obtain the peak point core sampling array; The core filling means uploading the sample peak points of the coefficient to the original peak point sampling array in sequence according to the order of peak point extraction, and replacing the elements at the original positions to obtain the peak point core sampling array.

4. A method for identifying the behavior of a power system terminal based on deep learning according to claim 1, characterized in that The process of joint transformation of the boundary peak point sampling array and feature extraction with the terminal sequence coefficient includes: Set the local boundary coefficient according to the peak point core sampling array, and perform local extraction on the peak point core sampling array according to the local boundary coefficient to obtain the boundary peak point sampling array; Perform adaptive adjustment on the boundary peak point sampling array to obtain a linear sampling segment, and perform sequential combination on the linear sampling segment based on the order of adaptive adjustment to obtain a linear boundary sampling group; Upload the obtained linear boundary sampling group to the terminal sequence coefficient, and perform cyclic screening on the terminal sequence coefficient through the linear boundary sampling group to obtain the terminal characteristic coefficient group.

5. The method for identifying the behavior of a power system terminal based on deep learning according to claim 4, wherein, The process of cyclic screening of the terminal sequence coefficient by the linear boundary sampling group includes: Set the starting endpoint according to the terminal sequence coefficient, and perform initial docking on the linear boundary sampling group based on the starting endpoint. The initial docking means aligning the element in the first row and first column of the boundary peak point sampling array with the starting endpoint, and uploading the linear boundary sampling group to the docking position; Perform docking extraction on the terminal sequence coefficient through the linear boundary sampling group to obtain the segmented boundary coefficient. The docking extraction means multiplying the elements in the linear boundary sampling group by the corresponding position elements of the terminal sequence coefficient, and summing the multiplication results at each position to obtain the segmented boundary coefficient; Set the sliding distance according to the starting endpoint, and perform sequential sliding on the linear boundary sampling group based on the terminal sequence coefficient according to the sliding distance to reach the next docking position, and perform docking extraction on the terminal sequence coefficient through the linear boundary sampling group to obtain the segmented boundary coefficient; Perform sequential sliding on the linear boundary sampling group based on the terminal sequence coefficient according to the sliding distance until the terminal sequence coefficient is covered; Combine the segmented boundary coefficients according to the order of sequential sliding to obtain the terminal characteristic coefficient group.

6. A method for identifying the behavior of a power system terminal based on deep learning according to claim 1, characterized in that, The process of setting the terminal data threshold for form replacement includes: Set the terminal data threshold based on the classification flow segment repository according to the terminal comprehensive data, and perform standard inspection on the terminal data threshold to obtain the standard threshold flow segment; Encode and segment the terminal characteristic flow segment based on the classification flow segment repository to obtain the terminal characteristic code element segment; Perform format replacement on the obtained specification threshold stream segment to obtain a threshold reference matrix.

7. A method for identifying the behavior of a power system terminal based on deep learning according to claim 6, characterized in that, The process of performing behavior screening on the classification stream segment repository according to the threshold reference matrix includes: Based on the classification stream segment repository, perform analogical judgment on the terminal feature code element segment according to the threshold reference matrix to obtain a reference judgment result; Perform forward matching on the terminal feature stream segment according to the obtained reference judgment result to obtain abnormal terminal data, and perform behavior analogy according to the obtained abnormal terminal data to obtain abnormal data behaviors.

Images