Power grid operation scene extraction method and device based on GMM improved clustering, terminal equipment and storage medium

Through the improved clustering method based on GMM, the clustering center is dynamically optimized, and the problem of boundary sample point classification fuzzy caused by the blurred initial clustering center in the traditional clustering method is solved, and a more accurate extraction of power grid operation scenarios is achieved.

CN119046708AActive Publication Date: 2024-11-29GUANGDONG POWER GRID CO LTD +1
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202411238923.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-09-05
Publication Date
2024-11-29
Estimated Expiration
2044-09-05

AI Technical Summary

Technical Problem

When processing time series data, traditional clustering methods easily lead to blurred initial clustering centers, which leads to blurred boundary sample points classification, making it difficult to accurately reflect the power grid operation scenario.

Method used

The improved clustering method based on GMM is adopted to obtain timing operation data, perform timing data mining and dimensionality reduction, build a GMM model, dynamically optimize the clustering center until the loss function converges, to obtain a more accurate grid operation scenario.

Benefits of technology

The problem of fuzzy boundary sample point classification caused by the fuzzy center of the initial cluster is effectively solved, which improves the accuracy and effectiveness of the clustering results, and can better reflect the true characteristics of the power grid operation.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119046708A_ABST
    Figure CN119046708A_ABST
Patent Text Reader

Abstract

The invention discloses a GMM improved clustering-based power grid operation scene extraction method and apparatus, a terminal device and a storage medium. The method comprises the steps of obtaining time sequence operation data of a to-be-extracted power grid; obtaining a time sequence operation characteristic according to the time sequence operation data and a preset sub-sequence length; projecting the time sequence operation features to a low-dimensional space to obtain time sequence feature indexes after dimension reduction; calculating time sequence weight vectors corresponding to the time sequence characteristic indexes according to the time sequence operation data, the time sequence characteristic indexes and the number of the time sequence characteristic indexes; constructing a GMM (Gaussian Mixture Model) according to the time sequence characteristic indexes and the time sequence weight vectors, and clustering to obtain a plurality of initial first clustering centers; and according to the plurality of initial first clustering centers, the time sequence characteristic indexes and the time sequence weight vectors, clustering center optimization operation is repeatedly carried out until a preset loss function converges, and a final time sequence characteristic clustering center is used as a power grid operation scene of the to-be-extracted power grid. According to the invention, a more accurate power grid operation scene is extracted.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of power systems, and in particular to a method, device, terminal equipment and storage medium for extracting power grid operation scenarios based on GMM improved clustering. Background Art

[0002] With the growing global demand for low-carbon and clean energy, the development of renewable energy has become an important trend in the new generation of power systems. However, the uncertainty of a high proportion of renewable energy and the large-scale access of power electronic equipment have brought multiple challenges to the power system, such as diversified operation modes, bidirectional power grid flow, and complex stability mechanisms. In addition, the pressure of renewable energy consumption and the growing demand for peak and frequency regulation have gradually highlighted the role of flexible resources such as energy storage in the power system. These new factors have begun to dominate the operation mode of the power grid. Traditional model-based and experience-based power system analysis methods are increasingly unable to cope with the massive and changing operation modes in power system operation, planning, protection and stability analysis. At present, the core issues of extracting power system operation scenarios include the similarity measurement of clustering data. Most existing studies use Euclidean distance and silhouette coefficient based on Euclidean distance, Ca linski-Harabaz index, Davis-Boul ld in index, etc. as evaluation indicators of clustering effectiveness, which are not enough to fully reflect the effectiveness of clustering results on time series data.

[0003] Traditional clustering methods usually use the mean of the curve within the cluster as the cluster center, which easily causes the loss of curve features in time series data. Due to the complexity and variability of time series, simple mean operations often smooth out the key features of the curve, resulting in a distorted shape of the clustering results, which cannot accurately reflect the true characteristics of the original data. Therefore, there is a problem of fuzzy initial cluster centers, which leads to fuzzy classification of boundary sample points, making it difficult to obtain effective power grid operation scenarios. Summary of the invention

[0004] The embodiments of the present invention provide a method, device, terminal device and storage medium for extracting power grid operation scenarios based on GMM improved clustering, which can effectively solve the problem in the prior art that the initial cluster center is vague, which leads to vague classification of boundary sample points, making it difficult to obtain effective power grid operation scenarios.

[0005] An embodiment of the present invention provides a method for extracting power grid operation scenarios based on GMM improved clustering, comprising:

[0006] Obtaining the time series operation data of the power grid to be extracted;

[0007] Performing time series data mining according to the time series operation data and a preset subsequence length to obtain time series operation features of the time series operation data;

[0008] Projecting the time series running features into a low-dimensional space for dimensionality reduction to obtain a time series feature index after dimensionality reduction;

[0009] Calculate a timing weight vector corresponding to each timing characteristic indicator according to the timing operation data, the timing characteristic indicator and the number of the timing characteristic indicators;

[0010] Constructing a GMM model according to the time series feature index and the time series weight vector, and clustering the time series feature index according to the constructed GMM model to obtain a plurality of initial first cluster centers;

[0011] According to a number of initial first cluster centers, the time series feature index and the time series weight vector, the cluster center optimization operation is repeatedly performed until a preset loss function converges to obtain a final time series feature cluster center;

[0012] The final time series feature clustering center is used as the power grid operation scenario of the power grid to be extracted.

[0013] Furthermore, time series data mining is performed according to the time series operation data and the preset subsequence length to obtain time series operation features of the time series operation data, including:

[0014] Taking a preset subsequence length as a sliding window length, and obtaining a plurality of time series operation subsequences of the time series operation data according to the time series operation data, a preset sliding window initial position, and a sliding window length;

[0015] According to the time series operation data and the time series operation subsequences, calculating distance values ​​between the time series operation subsequences and the time series operation data;

[0016] Determine a target timing operation feature according to the distance value, a plurality of timing operation subsequences and a preset distance threshold;

[0017] The distance value corresponding to the target timing operation feature is used as a feature value, and according to the feature value and the target timing operation feature, the target timing operation feature is converted into a vector space to obtain the final timing operation feature of the timing operation data.

[0018] Furthermore, the time series running features are projected into a low-dimensional space for dimensionality reduction to obtain time series feature indicators after dimensionality reduction, including:

[0019] Calculate, according to the time series operation characteristics, a first distance between adjacent time series operation characteristics and a second distance between any two time series operation characteristics;

[0020] According to the first distance, the second distance and a preset smoothing normalization factor, a weight between any two time series operation features corresponding to the second distance is calculated to obtain a weighted adjacency matrix;

[0021] Constructing a high-dimensional topology according to the weighted adjacency matrix and the timing operation characteristics to obtain a weighted neighbor graph;

[0022] Projecting the time series operation features into a low-dimensional space, and calculating a low-dimensional weight function of the time series operation features in the low-dimensional space;

[0023] According to the weighted neighbor graph and the low-dimensional weight function, a cross entropy function is constructed, and the cross entropy function is iteratively optimized by the stochastic gradient descent method until the preset loss function converges, and the time series feature index after dimensionality reduction is obtained according to the final cross entropy function.

[0024] Furthermore, according to the time series operation data, the time series characteristic index and the number of the time series characteristic indexes, a time series weight vector corresponding to each time series characteristic index is calculated, including:

[0025] Calculate the entropy value according to the time series operation data and the number of time series characteristic indicators;

[0026] According to the entropy value and the number of the time series feature indicators, the entropy value is corrected to calculate a first weight vector for representing the size of the information entropy of the time series feature indicator;

[0027] According to the time series characteristic index, calculate the standard deviation and correlation coefficient of the time series characteristic index;

[0028] Calculating a second weight vector for representing the correlation between the time series feature indicators according to the standard deviation, the correlation coefficient and the number of the time series feature indicators;

[0029] A timing weight vector corresponding to each timing feature indicator is calculated based on the first weight vector and the second weight vector.

[0030] Furthermore, the time series feature indicators are clustered according to the constructed GMM model to obtain several initial cluster centers, including:

[0031] Determine the initial Gaussian distribution parameters of the GMM model according to the constructed GMM model;

[0032] Calculate the posterior probability according to the current Gaussian distribution parameters; and continuously update the current Gaussian distribution parameters according to the posterior probability until the preset first loss function converges to obtain the final Gaussian distribution parameters; cluster the time series feature indicators according to the final Gaussian distribution parameters to obtain several Gaussian distributions; randomly select a Gaussian distribution as the first first clustering center;

[0033] According to the current first cluster center and the position of the time series characteristic index in the GMM model, the clustering probability of the time series characteristic index becoming the next first cluster center is calculated, and a new first cluster center is generated according to the clustering probability; the remaining time series characteristic indexes are divided into the new first cluster center;

[0034] When all time series feature indicators are divided into new first cluster centers, several initial first cluster centers are obtained; among them, the current Gaussian distribution parameters of the initial calculation of the posterior probability are the initial Gaussian distribution parameters; the current cluster center of the initial calculation of the cluster probability is the first first cluster center.

[0035] Furthermore, the cluster center optimization operation includes:

[0036] According to the current second cluster center, the time series feature index and the time series weight vector, the cluster distance between the time series feature index and the current second cluster center is calculated; wherein the current second cluster center for the first calculation of the cluster distance is the initial first cluster center;

[0037] The current second cluster center corresponding to the minimum cluster distance of each time series feature indicator is used as the new current second cluster center corresponding to each time series feature indicator, until the current second cluster centers corresponding to all time series feature indicators remain unchanged, and the cluster center optimization operation is completed; and the final time series feature cluster center is obtained based on the current second cluster center.

[0038] Furthermore, it also includes:

[0039] According to the final time series feature cluster center and the time series feature index, the average distance between all the time series feature indexes and the time series feature cluster center is calculated;

[0040] According to the final time series feature clustering center, calculate the distance between every two time series feature clustering centers;

[0041] The clustering evaluation coefficient is calculated based on the average distance and the distance between the centers of each two time series feature clusters;

[0042] The final time series feature clustering center is evaluated according to the clustering evaluation coefficient.

[0043] As an improvement of the above solution, another embodiment of the present invention provides a power grid operation scenario extraction device based on GMM improved clustering, including:

[0044] A time series data acquisition module is used to acquire the time series operation data of the power grid to be extracted;

[0045] A time series feature mining module, used to perform time series data mining according to the time series operation data and a preset subsequence length, to obtain the time series operation features of the time series operation data;

[0046] A time series feature dimensionality reduction module is used to project the time series operation features into a low-dimensional space for dimensionality reduction, and obtain time series feature indicators after dimensionality reduction;

[0047] A timing weight calculation module, used to calculate a timing weight vector corresponding to each timing characteristic indicator according to the timing operation data, the timing characteristic indicator and the number of the timing characteristic indicators;

[0048] A cluster center initialization module, used to construct a GMM model according to the time series feature index and the time series weight vector, and cluster the time series feature index according to the constructed GMM model to obtain a plurality of initial first cluster centers;

[0049] A cluster center optimization module, used to repeatedly perform cluster center optimization operations according to a number of initial first cluster centers, the time series feature index and the time series weight vector, until a preset loss function converges, thereby obtaining a final time series feature cluster center;

[0050] The power grid operation scenario extraction module is used to use the final time series feature clustering center as the power grid operation scenario of the power grid to be extracted.

[0051] Another embodiment of the present invention provides a terminal device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, and when the processor executes the computer program, it implements a method for extracting power grid operation scenarios based on GMM improved clustering as described in the above embodiment.

[0052] Another embodiment of the present invention provides a computer-readable storage medium, which includes a stored computer program, wherein when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute a power grid operation scenario extraction method based on GMM improved clustering as described in the above embodiment.

[0053] By implementing the present invention, at least the following beneficial effects are achieved:

[0054] The present invention provides a method, device, terminal equipment and storage medium for extracting power grid operation scenarios based on GMM improved clustering. The method can obtain the time series operation data of the power grid to be extracted; perform time series data mining according to the time series operation data and a preset subsequence length to obtain the time series operation characteristics of the time series operation data; project the time series operation characteristics to a low-dimensional space for dimensionality reduction to obtain the time series characteristic indicators after dimensionality reduction; calculate the time series weight vector corresponding to each time series characteristic indicator according to the time series operation data, the time series characteristic indicators and the number of the time series characteristic indicators; construct a GMM model according to the time series characteristic indicators and the time series weight vector, and cluster the time series characteristic indicators according to the constructed GMM model to obtain a number of initial first clustering centers; repeatedly perform clustering center optimization operations according to the several initial first clustering centers, the time series characteristic indicators and the time series weight vector until the preset loss function converges to obtain the final time series characteristic clustering center; and use the final time series characteristic clustering center as the power grid operation scenario of the power grid to be extracted. By reducing the dimension of the time series operation characteristics of the time series operation data to obtain the time series characteristic index, different time series operation data can be better separated in the low-dimensional space, and the distinction between the time series characteristic indicators can be enhanced, thereby solving the problem of fuzzy classification of sample points during clustering; then, a GMM model is constructed according to the calculated time series weight vector and the time series characteristic index. The GMM model can be used to model the distribution of each cluster center. The GMM model is used to initialize the cluster center to obtain several initial first cluster centers. The initial first cluster center is located in an area with a high density of time series characteristic indicators to avoid clustering being affected by outlier time series characteristic indicators and causing cluster center offset. While retaining the true characteristics of the time series operation data, the problem of fuzzy classification of boundary sample points caused by the fuzzy initial cluster center is solved, and the time series characteristic cluster center is dynamically optimized to obtain a more accurate and effective power grid operation scenario. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is a flow chart of a method for extracting power grid operation scenarios based on GMM improved clustering provided by one embodiment of the present invention;

[0056] Figure 2 It is another flow chart of a method for extracting power grid operation scenarios based on GMM improved clustering provided by one embodiment of the present invention;

[0057] Figure 3 It is a structural schematic diagram of a power grid operation scenario extraction device based on GMM improved clustering provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0058] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0059] See also Figure 1 , is a flow chart of a method for extracting power grid operation scenarios based on GMM improved clustering provided by an embodiment of the present invention, comprising:

[0060] S1. Obtaining the time series operation data of the power grid to be extracted;

[0061] S2. performing time series data mining according to the time series operation data and a preset subsequence length to obtain time series operation features of the time series operation data;

[0062] S3, projecting the time series running characteristics into a low-dimensional space for dimensionality reduction, and obtaining a time series characteristic index after dimensionality reduction;

[0063] S4. Calculate a timing weight vector corresponding to each timing characteristic indicator according to the timing operation data, the timing characteristic indicator and the number of the timing characteristic indicators;

[0064] S5. Constructing a GMM model according to the time series feature index and the time series weight vector, and clustering the time series feature index according to the constructed GMM model to obtain a plurality of initial first cluster centers;

[0065] S6, repeating the cluster center optimization operation according to the initial first cluster centers, the time series feature index and the time series weight vector until the preset loss function converges to obtain the final time series feature cluster center;

[0066] S7. The final time series feature clustering center is used as the power grid operation scenario of the power grid to be extracted.

[0067] Specifically, the time series operation data of the power grid to be extracted include generator output data, power grid operation load data, and cross-section flow data. Because the power system is a highly complex and intertwined engineering system, the power grid data covers a wide and detailed set of information, including but not limited to the performance parameters of various power generation facilities (such as generator output power and power angle), power grid component characteristics (transformer ratio, line resistance and reactance), network topology, load demand forecast, reactive voltage control parameters, and intermittent and uncertain characteristics of renewable energy power generation. These data are multi-dimensional. If they are not selected, using all data variables to characterize the characteristics of the power grid operation scenario will inevitably lead to a large amount of data redundancy and greatly reduce the calculation efficiency. Therefore, it is meaningful and necessary to select data in advance.

[0068] As the core component of the power system, the operation mode and active power output of the generator set play a vital role in the operation mode and stability of the power grid. Different types of power generation play different roles in power grid planning due to differences in response speed, output adjustability and environmental dependence. For example, hydropower generation is often used to provide fast frequency support due to its fast start-up and flexible regulation; new energy power generation, mainly wind power and photovoltaic power generation, is greatly affected by factors such as climate and time, and its output has obvious intermittent and uncertain characteristics, requiring the power grid to have higher flexibility and dispatching capabilities. Large-scale access is prone to frequent changes in power grid trends, causing difficulties in voltage adjustment and protection configuration of the power system, and bringing challenges to the safe and stable operation of the power system. Among the many influencing factors, the active output of the generator set has the most obvious impact on the power grid operation scenario. Therefore, when selecting attribute variables to characterize the power grid operation scenario, the output data of the generator set on the power generation side should be taken into account.

[0069] As the energy consumption terminal in the power system, the load side has a great impact on the power grid operation scenario due to its level and type. The load level directly restricts the output level of various types of units on the power generation side. With the ups and downs of social and economic activities and changes in user demand, the load level shows periodic and seasonal fluctuations. During the peak period of electricity consumption in summer, the power system needs to respond quickly to increase the output of thermal power, hydropower or start the standby units to avoid the pressure on the power grid and potential power outage risks caused by supply and demand. This dynamic adjustment process has a profound impact on the power flow distribution of the system. Loads can be divided into industrial loads, commercial loads, and residential loads. And each type of load has different characteristics, such as industrial loads with high power and strict requirements on power quality; commercial loads show obvious daily and weekly changes; and residential loads show double peak characteristics in the morning and evening. The access of different types of loads also tests the stability and reliability of the entire system. Therefore, load is an important attribute variable in the extraction of power grid operation scenarios.

[0070] As a key indicator reflecting the source-load distribution and power transmission path of the power grid, cross-sectional flow not only reveals the actual flow of electricity from the production end to the consumption end, but also helps identify potential bottlenecks in the power grid. Accurately grasping cross-sectional flow is crucial to studying power grid operation scenarios, ensuring safe and stable operation of the power grid, optimizing resource allocation, and preventing and responding to power system failures. In the context of energy transformation, the power system faces a high degree of uncertainty. Ensuring the rationality, safety, reliability and economy of the power grid operation scenarios has become a major challenge. Analyzing the distribution of cross-sectional flow is an indispensable means to solve this challenge. Therefore, cross-sectional flow data is an indispensable part of the study of power grid operation scenarios.

[0071] Time series operation features can effectively capture the local characteristics of time series operation data. They have the advantages of good interpretability, fast classification speed, and high accuracy, and are widely used in the field of time series classification. When processing large-scale time series operation data, one of the main challenges is the high dimensionality of the data, which mainly comes from the large number of scenarios, generators, and load nodes. At the same time, there are many highly similar active power variables that vary with time. If the power grid operation scenario is directly extracted, the computational efficiency will be seriously affected. Therefore, before applying the clustering algorithm, it is particularly important to take dimensionality reduction measures to select the time series feature indicators at each time point and simplify the original high-dimensional attributes.

[0072] The time series weight vector corresponding to each time series feature index can objectively reflect the importance of each index feature, and its configuration difference will greatly affect the clustering quality. At the same time, for different clustering curve cluster objects, the importance of the same time series feature index is also quite different. To this end, it is necessary to reasonably configure the time series weights of the time series feature index. Gaussian Mixture Model (GMM) is often used in combination with Expectation Maximization (EM). GMM assumes that the data consists of multiple Gaussian distributions, each corresponding to a potential category or cluster. Different from hard clustering methods such as K-means, GMM adopts a soft clustering method, allowing data points to belong to multiple clusters with a certain probability, which makes GMM more flexible in dealing with overlapping or fuzzy cluster boundaries.

[0073] In a preferred embodiment of the present invention, the timing operation data of the power grid to be extracted is first obtained; then, timing data mining is performed based on the timing operation data and a preset subsequence length to obtain the timing operation characteristics of the timing operation data; then, the timing operation characteristics are projected into a low-dimensional space for dimensionality reduction to obtain the timing characteristic indicators after dimensionality reduction; then, based on the timing operation data, the timing characteristic indicators and the number of the timing characteristic indicators, the timing weight vector corresponding to each timing characteristic indicator is calculated; based on the timing characteristic indicators and the timing weight vector, a GMM model is constructed, and the timing characteristic indicators are clustered according to the constructed GMM model to obtain several initial first clustering centers; then, based on several initial first clustering centers, the timing characteristic indicators and the timing weight vector, the clustering center optimization operation is repeated until the preset loss function converges to obtain the final timing characteristic clustering center; finally, the final timing characteristic clustering center is used as the power grid operation scenario of the power grid to be extracted. By considering the difference degree and correlation of the timing operation characteristics of the timing operation data; proposing a timing weight vector, it can effectively improve the disadvantage of the traditional Euclidean distance calculation of the similarity of timing operation characteristics, reduce the sensitivity to the distribution form of the data, and effectively capture information, thereby improving the clustering quality; improving clustering through the weighted GMM model, solving the problem of fuzzy determination of the initial cluster center and fuzzy classification of boundary sample points in traditional clustering.

[0074] Preferably, performing time series data mining according to the time series operation data and a preset subsequence length to obtain time series operation features of the time series operation data includes:

[0075] Taking a preset subsequence length as a sliding window length, and obtaining a plurality of time series operation subsequences of the time series operation data according to the time series operation data, a preset sliding window initial position, and a sliding window length;

[0076] According to the time series operation data and the time series operation subsequences, calculating distance values ​​between the time series operation subsequences and the time series operation data;

[0077] Determine a target timing operation feature according to the distance value, a plurality of timing operation subsequences and a preset distance threshold;

[0078] The distance value corresponding to the target timing operation feature is used as a feature value, and according to the feature value and the target timing operation feature, the target timing operation feature is converted into a vector space to obtain the final timing operation feature of the timing operation data.

[0079] Specifically, the timing operation data T=t 1 ,t 2 ,...,t Mis a set of M real-valued data points arranged in time order, where adjacent data points t 1 ,t 2 ,...,t M The preset subsequence length is: Assuming a given time series T with a length of M, then a continuous sequence segment S with a length of l (l ≤ M) on T is called a subsequence of T, that is, S = t p ,t p+1 ,...,t p+l-1 , where 1≤p≤M-l+1. Assuming that a time series running data T of length M and a user-defined subsequence length l are given, the subsequence can be found through a sliding window of length l, and the subsequence defined by this sliding window is expressed as The superscript l represents the length of the subsequence, and the subscript p represents the starting position of the sliding window. The set of all time series running subsequences of length l in T is defined as S l (T), that is Dist(T,R) is a function that calculates the distance between two time series. When two time series T and R of equal length are input, Dist(T,R) returns the distance d (d≥0) between T and R, and Dist(T,R)=Dist(R,T). SubDist(T,S) is a function that calculates the distance d (d≥0) between the time series T and the subsequence S. SubDist(T,S)=min(Dist(S,S')). Then, based on the time series running data and several time series running subsequences, the distance values ​​between several time series running subsequences and the time series running data can be calculated. The preset distance threshold has an effect on other distance values ​​d. t ' h All satisfy: Gain(S,d OSP(D,S) )≥Gain(S,d t ' h ), the shapelet in the time series dataset D and its corresponding optimal distance threshold satisfy the following for any other subsequence S: Gain(Shapelet(D),d OSP(D,Shapelet(D)) )≥Gain(S,d OSP(D,S) )

[0080] , Shapelet(D) represents the shapelet in the data set D, that is, the target time series operation feature. Finally, the distance value corresponding to the target time series operation feature is used as the feature value, and according to the feature value and the target time series operation feature, the target time series operation feature is converted into the vector space to obtain the final time series operation feature of the time series operation data.

[0081] Preferably, the time series running characteristics are projected into a low-dimensional space for dimensionality reduction to obtain time series characteristic indicators after dimensionality reduction, including:

[0082] Calculate, according to the time series operation characteristics, a first distance between adjacent time series operation characteristics and a second distance between any two time series operation characteristics;

[0083] According to the first distance, the second distance and a preset smoothing normalization factor, a weight between any two time series operation features corresponding to the second distance is calculated to obtain a weighted adjacency matrix;

[0084] Constructing a high-dimensional topology according to the weighted adjacency matrix and the timing operation characteristics to obtain a weighted neighbor graph;

[0085] Projecting the time series operation features into a low-dimensional space, and calculating a low-dimensional weight function of the time series operation features in the low-dimensional space;

[0086] According to the weighted neighbor graph and the low-dimensional weight function, a cross entropy function is constructed, and the cross entropy function is iteratively optimized by the stochastic gradient descent method until the preset loss function converges, and the time series feature index after dimensionality reduction is obtained according to the final cross entropy function.

[0087] Specifically, UMAP is an advanced manifold learning technique that is often used for dimensionality reduction of high-dimensional data. It combines the concepts of Riemannian geometry and algebraic topology, and aims to reveal the intrinsic structure in the data through a nonlinear dimensionality reduction algorithm while maintaining the topological relationship between data points. Compared with the currently used t-SNE technology, UMAP shows significant advantages in dimensionality reduction quality and operating efficiency. The specific implementation process of UMAP is divided into two main stages, namely high-dimensional topological construction and low-dimensional projection. In the high-dimensional topological construction stage, the core content is to create a weighted k-nearest neighbor graph Among them, V is the vertex set, that is, the time series running features X={x 1 ,x 2 ,…,x N}; E is the edge set, that is, the set of directed edges that can be formed according to the time series operation characteristics; W is the weight function, and its calculation formula is: In the formula, s ij For x i With x j The distance between them, that is, the second distance between any two time series operation features; ρ i For x i The distance between adjacent points, that is, the first distance between adjacent timing operation features; σ i is the preset smoothing normalization factor. After calculating the weights of all time series operation features, in order to maintain the consistency of the weights between any two time series operation features, the expression is introduced: Where A is all W ij The weighted adjacency matrix composed of is the Hadamard product of the matrix; B is the final weighted adjacency matrix, and finally a high-dimensional topology is constructed according to the weighted adjacency matrix and the timing operation characteristics to obtain a weighted neighbor graph G. After the high-dimensional topology construction stage is completed, in order to realize the low-dimensional projection, it is necessary to construct a low-dimensional weight function in low dimension, that is, Where, d ij To project the time series running features into a low-dimensional space, the time series running features Y = {y 1 ,y 2 ,…,y N} i With y j The distance between them; a and b are both hyperparameters. In order to make the time series running characteristics after dimensionality reduction as close as possible to the original time series running characteristics, it is necessary to optimize V ij and W ij According to the weighted neighbor graph and the low-dimensional weight function, a cross entropy function is constructed, and the cross entropy function is: In the formula, the first term is the attractive force component of constraining the similar points to form clusters after dimensionality reduction; the second term is the repulsive force component to ensure that there is a sufficient distance between clusters after dimensionality reduction. The cross entropy function is iteratively optimized using the stochastic gradient descent method until the preset loss function converges, and the time series feature index after dimensionality reduction is obtained according to the final cross entropy function.

[0088] Specifically, according to the time series operation data, the time series characteristic index and the number of the time series characteristic indexes, a time series weight vector corresponding to each time series characteristic index is calculated, including:

[0089] Calculate the entropy value according to the time series operation data and the number of time series characteristic indicators;

[0090] According to the entropy value and the number of the time series feature indicators, the entropy value is corrected to calculate a first weight vector for representing the size of the information entropy of the time series feature indicator;

[0091] According to the time series characteristic index, calculate the standard deviation and correlation coefficient of the time series characteristic index;

[0092] Calculating a second weight vector for representing the correlation between the time series feature indicators according to the standard deviation, the correlation coefficient and the number of the time series feature indicators;

[0093] A timing weight vector corresponding to each timing feature indicator is calculated based on the first weight vector and the second weight vector.

[0094] Preferably, the entropy weight method is an objective method for determining the weight of each indicator in multi-indicator decision analysis, and is widely used in the fields of power quality assessment, competitiveness evaluation, etc. Its core principle is based on the concept of information entropy, and the importance of indicators is judged by quantifying the uncertainty of indicator data. This means that the higher the degree of data dispersion, the smaller the information entropy, the more effective information it covers, the more important its position in the comprehensive evaluation, and the greater its weight. Therefore, the entropy value h is first calculated based on the time series operation data and the number of time series characteristic indicators. j ,

[0095]

[0096] In the formula, m is the number of time series feature indicators; n is the number of data categories in the time series operation data; r ij is the time series running data; f ij is the contribution of the jth time series feature indicator of the ith data category.

[0097] Then, according to the entropy value and the number of the time series feature indicators, the entropy value is corrected to calculate a first weight vector for representing the information entropy size of the time series feature indicator: When the difference between the characteristic indicators of the same time series is small, h j The calculation result is close to 1. If the traditional calculation method is used, there will be a large error. Therefore, this embodiment corrects the entropy value to obtain the first weight vector used to represent the information entropy size of the time series feature indicator.

[0098]

[0099] In order to solve the problem that the entropy weight method cannot handle the correlation between the first weight vectors, the CRITIC weight method is introduced to solve this problem. The CRITIC weight method is an objective weighting technology, which is mainly evaluated through two core elements: contrast intensity and conflict. Contrast intensity is measured based on standard deviation. The greater the standard deviation fluctuation, the higher the importance of the indicator, and the higher the weight is assigned accordingly; conflict is based on the correlation coefficient between time series feature indicators. The larger the correlation coefficient, the smaller the independence or conflict between the time series feature indicators, and the lower the weight. According to the time series feature indicators, the standard deviation (indicating variability) S of the time series feature indicators is calculated. j , In the formula, s j is the standard deviation of the jth time series characteristic index, is the average value of the jth time series characteristic index. According to the time series characteristic index, the correlation coefficient r of the time series characteristic index is calculated. ij , calculate the conflict between time series characteristic indicators based on the correlation coefficient Then calculate the information content of each indicator Finally, a second weight vector for representing the correlation between the time series feature indicators is calculated. The combined use of the entropy weight method and the CRITIC weight method can better balance the weights of the time series feature indicators and the relationship between the time series feature indicators. The entropy weight method can calculate the weights of each time series feature indicator, and the CRITIC weight method can confirm the relationship between each time series feature indicator through comprehensive analysis, thereby providing a more accurate time series weight vector in the feature analysis process.

[0100] Where W j ' is the initial combined weight of the comprehensive entropy weight method and the CRITIC weight method; k is the coefficient of variation; It is the time series weight vector of the comprehensive entropy weight method and CRI TIC weight method.

[0101] Preferably, the time series feature indicators are clustered according to the constructed GMM model to obtain several initial cluster centers, including:

[0102] Determine the initial Gaussian distribution parameters of the GMM model according to the constructed GMM model;

[0103] Calculate the posterior probability according to the current Gaussian distribution parameters; and continuously update the current Gaussian distribution parameters according to the posterior probability until the preset first loss function converges to obtain the final Gaussian distribution parameters; cluster the time series feature indicators according to the final Gaussian distribution parameters to obtain several Gaussian distributions; randomly select a Gaussian distribution as the first first clustering center;

[0104] According to the current first cluster center and the position of the time series characteristic index in the GMM model, the clustering probability of the time series characteristic index becoming the next first cluster center is calculated, and a new first cluster center is generated according to the clustering probability; the remaining time series characteristic indexes are divided into the new first cluster center;

[0105] When all time series feature indicators are divided into new first cluster centers, several initial first cluster centers are obtained; among them, the current Gaussian distribution parameters of the initial calculation of the posterior probability are the initial Gaussian distribution parameters; the current cluster center of the initial calculation of the cluster probability is the first first cluster center.

[0106] In a preferred embodiment of the present invention, Kmeans++, as an improved method of Kmeans, mainly acts on the initialization of cluster centers. In traditional K-means, cluster centers are selected in a completely random manner, while K-means++ introduces the distance maximization principle, initially randomly selecting a data point as the first cluster center, and then the selection probability of the new cluster center is proportional to the distance of the current cluster center, that is, the farther the point is from the existing cluster center, the more likely it is to be selected as the new cluster center, ensuring that the distribution of the initial cluster center in the data space is more dispersed, thereby improving the algorithm performance and clustering quality. Gaussian Mixture Model (GMM) is often used in combination with Expectation Maximization (EM). GMM assumes that the data consists of multiple Gaussian distributions, each distribution corresponds to a potential category or cluster. Different from hard clustering methods such as K-means, GMM adopts a soft clustering method, allowing data points to belong to multiple clusters with a certain probability, which makes GMM more flexible in dealing with overlapping or fuzzy cluster boundaries.

[0107] Assume X = {x 1 ,x 2 ,…,x n} is the time series characteristic index, and the mathematical expression of the GMM model can be expressed as:

[0108]

[0109] Where P(x) is the probability density function of the Gaussian mixture model; K is the number of Gaussian components; π k is the weight coefficient of the kth component; is the multivariate probability density function of the kth component; μ k is the mean of the kth component; ∑ k is the covariance of the kth component. The Expectation Maximization (EM) algorithm is used to estimate the initial Gaussian distribution parameter π k ,μ k and∑ k The GMM model can define the number of categories K and first set π k , μ k and∑ k Initial value; calculate the posterior probability P(Z) based on the current Gaussian distribution parameters nk ):

[0110]

[0111] Then maximize the maximum likelihood and update π k ,μk and∑ k :

[0112]

[0113] Until the preset first loss function converges, the final Gaussian distribution parameters are obtained. Cluster the time series feature indicators according to the final Gaussian distribution parameters to obtain several Gaussian distributions; randomly select a Gaussian distribution as the first first cluster center; calculate the clustering probability of the time series feature indicator becoming the next first cluster center according to the current first cluster center and the position of the time series feature indicator in the GMM model, and generate a new first cluster center according to the clustering probability; divide the remaining time series feature indicators into new first cluster centers; when all time series feature indicators are divided into new first cluster centers, several initial first cluster centers are obtained; among which, the current Gaussian distribution parameters for the initial calculation of the posterior probability are the initial Gaussian distribution parameters; the current cluster center for the initial calculation of the clustering probability is the first first cluster center.

[0114] Schematically, the cluster center optimization operation includes:

[0115] According to the current second cluster center, the time series feature index and the time series weight vector, the cluster distance between the time series feature index and the current second cluster center is calculated; wherein the current second cluster center for the first calculation of the cluster distance is the initial first cluster center;

[0116] The current second cluster center corresponding to the minimum cluster distance of each time series feature indicator is used as the new current second cluster center corresponding to each time series feature indicator, until the current second cluster centers corresponding to all time series feature indicators remain unchanged, and the cluster center optimization operation is completed; and the final time series feature cluster center is obtained based on the current second cluster center.

[0117] In a preferred embodiment of the present invention, the K-means++ algorithm is run according to the current second cluster center, the time series feature index and the time series weight vector to calculate the cluster distance between the time series feature index and the current second cluster center, and the minimum cluster distance R (C i )=min{dis(C i ,z n )} is used as the new current second cluster center, that is, the roulette method is used to select the next second cluster center, and the probability of selection is Until the number of selected second cluster centers reaches the required K, the current second cluster centers corresponding to all time series feature indicators remain unchanged, the cluster center optimization operation is completed, and the distance from each time series feature indicator to the second cluster center is calculated. In the formula, dis(Ci ,z n ) * is the weighted Euclidean distance, is the time series weight vector of time series feature index j; ∑ j dis 2 (C i,j ,z n,j ) is the distance between the time series feature index and the second cluster center. The time series feature index close to the second cluster center is divided into one category to obtain the final time series feature cluster center.

[0118] Indicatively, it also includes:

[0119] According to the final time series feature cluster center and the time series feature index, the average distance between all the time series feature indexes and the time series feature cluster center is calculated;

[0120] According to the final time series feature clustering center, calculate the distance between every two time series feature clustering centers;

[0121] The clustering evaluation coefficient is calculated based on the average distance and the distance between the centers of each two time series feature clusters;

[0122] The final time series feature clustering center is evaluated according to the clustering evaluation coefficient.

[0123] In a preferred embodiment of the present invention, the clustering evaluation coefficient

[0124] In the formula, σ i is the average distance between all time series feature indicators of the i-th category and the cluster center of the time series feature, d(μ i ,μ j ) is the distance between the cluster center of the i-th time series feature and the cluster center of the j-th time series feature. The smaller the DBI, the closer the intra-class distance and the longer the inter-class distance, and the better the clustering performance.

[0125] In another preferred embodiment of the present invention, the silhouette coefficient In the formula, a i is the average distance between the time series feature index and all other time series feature indexes in the same category, b i It is the average distance between the time series feature index and the closest time series feature index in different categories. The value of the silhouette coefficient is between [-1,1]. The closer it is to 1, the closer the distance is in the same category, and the farther the distance is in different categories, the better the clustering effect. CHI index Where B(k) is the sum of squared deviations between clusters, i.e., the weighted average of the sum of squared deviations between all cluster centroids and the total centroid. W(k) is the sum of squared deviations within clusters, i.e., the sum of the sum of squared deviations from all time series feature indicators in each cluster to the centroid of the cluster. N is the total number of time series feature indicators; k is the number of clusters. The larger the CHI, the better the compactness between the same clusters, the stronger the dispersion between different clusters, and the better the clustering effect.

[0126] In a preferred embodiment of the present invention, Figure 2 As shown in the figure, the attribute variables of the power grid are first selected, and then the data preprocessing and data standardization of the attribute variables are performed, and then the Shapelet time series feature extraction is performed, and then the time series features are subjected to UMP dimensionality reduction. According to the weight coefficient calculated by combining the entropy weight method with the CRITIC weight method, clustering is performed based on the GMM improved K-means++ algorithm to obtain the power grid operation scenario. The Euclidean distance calculation method that improves the entropy weight method and integrates the CRITIC weight method is proposed, which can effectively improve the disadvantages of the traditional Euclidean distance calculation feature similarity, reduce the sensitivity to the distribution form of the data, and can more effectively capture information. The weight obtained is more representative and global, and the feature similarity calculated is more accurate, thereby improving the clustering quality. The Shapelet transform can identify the key local patterns (i.e., shape l ets) in the time series data by using the Shapelet and UMAP algorithms to extract and reduce the dimensionality of the time series features. These patterns are highly representative for distinguishing different categories. UMAP is an efficient data dimensionality reduction technology. It can retain the local structure of high-dimensional data, so that data points of different categories can be better separated in low-dimensional space. It can solve the problems of high dimensionality and multiple types of power grid operation data. The combination of the two can extract the typical characteristics of multi-dimensional time series variables, realize the mapping of complex high-dimensional features to low-dimensional space, and retain the time series feature information to the greatest extent. Weighted GMM improves the K-means++ algorithm. GMM can estimate the probability density function of data. Outliers are usually located in low-density areas. These low-probability points can be marked as outliers to avoid their negative impact on clustering results. K-means is easily affected by outliers because outliers may become centroids, resulting in clustering results offset. GMM initialization can help K-means avoid this situation because GMM tends to place the centroid in areas with higher data density, so it can solve the problem of initial cluster center determination and classification ambiguity of boundary sample points in the traditional K-means clustering algorithm.

[0127] By implementing this embodiment, the timing operation characteristics of the timing operation data are reduced in dimension to obtain timing characteristic indicators, so that different timing operation data can be better separated in a low-dimensional space, and the distinction between the timing characteristic indicators is enhanced, thereby solving the problem of fuzzy classification of sample points during clustering; then, a GMM model is constructed according to the calculated timing weight vector and the timing characteristic indicator, and the distribution of each cluster center can be modeled using the GMM model. The cluster center is initialized using the GMM model to obtain a number of initial first cluster centers, and the initial first cluster center is located in an area with a high density of timing characteristic indicators, so as to avoid the clustering being affected by outlier timing characteristic indicators and causing the cluster center to shift, thereby retaining the true characteristics of the timing operation data while solving the problem of fuzzy classification of boundary sample points caused by fuzzy initial cluster centers, dynamically optimizing the timing characteristic cluster centers, and obtaining a more accurate and effective power grid operation scenario.

[0128] See also Figure 3 , is a schematic diagram of a structure of a power grid operation scenario extraction device based on GMM improved clustering provided by an embodiment of the present invention, comprising:

[0129] A time series data acquisition module is used to acquire the time series operation data of the power grid to be extracted;

[0130] A time series feature mining module, used to perform time series data mining according to the time series operation data and a preset subsequence length, to obtain the time series operation features of the time series operation data;

[0131] A time series feature dimensionality reduction module is used to project the time series operation features into a low-dimensional space for dimensionality reduction, and obtain time series feature indicators after dimensionality reduction;

[0132] A timing weight calculation module, used to calculate a timing weight vector corresponding to each timing characteristic indicator according to the timing operation data, the timing characteristic indicator and the number of the timing characteristic indicators;

[0133] A cluster center initialization module, used to construct a GMM model according to the time series feature index and the time series weight vector, and cluster the time series feature index according to the constructed GMM model to obtain a plurality of initial first cluster centers;

[0134] A cluster center optimization module, used to repeatedly perform cluster center optimization operations according to a number of initial first cluster centers, the time series feature index and the time series weight vector, until a preset loss function converges, thereby obtaining a final time series feature cluster center;

[0135] The power grid operation scenario extraction module is used to use the final time series feature clustering center as the power grid operation scenario of the power grid to be extracted.

[0136] The present invention provides a power grid operation scene extraction device based on GMM improved clustering, which obtains the time series operation data of the power grid to be extracted according to the time series data acquisition module; in the time series feature mining module, the time series data mining is performed according to the time series operation data and the preset subsequence length to obtain the time series operation characteristics of the time series operation data; then, through the time series feature dimension reduction module, the time series operation characteristics are projected to a low-dimensional space for dimension reduction to obtain the time series feature index after dimension reduction; then, in the time series weight calculation module, each time series feature index is calculated according to the time series operation data, the time series feature index and the number of the time series feature indexes. The timing weight vector corresponding to the label; in the cluster center initialization module, a GMM model is constructed according to the timing feature index and the timing weight vector, and the timing feature index is clustered according to the constructed GMM model to obtain a number of initial first cluster centers; in the cluster center optimization module, the cluster center optimization operation is repeated according to the number of initial first cluster centers, the timing feature index and the timing weight vector until the preset loss function converges to obtain the final timing feature cluster center; finally, through the power grid operation scenario extraction module, the final timing feature cluster center is used as the power grid operation scenario of the power grid to be extracted. By reducing the dimension of the time series operation characteristics of the time series operation data to obtain the time series characteristic index, different time series operation data can be better separated in the low-dimensional space, and the distinction between the time series characteristic indicators can be enhanced, thereby solving the problem of fuzzy classification of sample points during clustering; then, a GMM model is constructed according to the calculated time series weight vector and the time series characteristic index. The GMM model can be used to model the distribution of each cluster center. The GMM model is used to initialize the cluster center to obtain several initial first cluster centers. The initial first cluster center is located in an area with a high density of time series characteristic indicators to avoid clustering being affected by outlier time series characteristic indicators and causing cluster center offset. While retaining the true characteristics of the time series operation data, the problem of fuzzy classification of boundary sample points caused by the fuzzy initial cluster center is solved, and the time series characteristic cluster center is dynamically optimized to obtain a more accurate and effective power grid operation scenario.

[0137] It should be noted that the device embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. In addition, in the accompanying drawings of the device embodiments provided by the present invention, the connection relationship between the modules indicates that there is a communication connection between them, which may be specifically implemented as one or more communication buses or signal lines. A person of ordinary skill in the art may understand and implement it without paying any creative effort.

[0138] Those skilled in the art can clearly understand that, for the sake of convenience and brevity, the specific working process of the device described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0139] Another embodiment of the present invention further provides a terminal device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein when the processor executes the computer program, a method for extracting power grid operation scenarios based on GMM improved clustering as described in the above embodiment is implemented. The terminal device may be a computing device such as a desktop computer, a notebook, a PDA, and a cloud server. The terminal device may include, but is not limited to, a processor and a memory.

[0140] The processor may be a central processing unit (CPU), or other general-purpose processors, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc. The processor is the control center of the terminal device, and various interfaces and lines are used to connect various parts of the entire terminal device.

[0141] The memory can be used to store the computer program, and the processor realizes various functions of the terminal device by running or executing the computer program stored in the memory and calling the data stored in the memory. The memory can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, at least one application required for a function, etc.; the data storage area can store data created according to the use of the mobile phone, etc. In addition, the memory can include a high-speed random access memory, and can also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart memory card (Smart Med i aCard, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device or other volatile solid-state storage device.

[0142] Another embodiment of the present invention provides a computer-readable storage medium, which includes a stored computer program, wherein when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute a power grid operation scenario extraction method based on GMM improved clustering as described in the above embodiment.

[0143] The storage medium is a computer-readable storage medium, and the computer program is stored in the computer-readable storage medium. When the computer program is executed by the processor, the steps of each of the above-mentioned method embodiments can be implemented. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium.

[0144] The above is a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements and modifications without departing from the principle of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A power grid operation scenario extraction method based on GMM improved clustering, It is characterized in that include: Obtaining the time series operation data of the power grid to be extracted; Performing time series data mining according to the time series operation data and a preset subsequence length to obtain time series operation features of the time series operation data; Projecting the time series running features into a low-dimensional space for dimensionality reduction to obtain a time series feature index after dimensionality reduction; Calculate a timing weight vector corresponding to each timing characteristic indicator according to the timing operation data, the timing characteristic indicator and the number of the timing characteristic indicators; Constructing a GMM model according to the time series feature index and the time series weight vector, and clustering the time series feature index according to the constructed GMM model to obtain a plurality of initial first cluster centers; According to a number of initial first cluster centers, the time series feature index and the time series weight vector, the cluster center optimization operation is repeatedly performed until a preset loss function converges to obtain a final time series feature cluster center; The final time series feature clustering center is used as the power grid operation scenario of the power grid to be extracted.

2. A method for extracting power grid operation scenarios based on GMM improved clustering as claimed in claim 1, It is characterized in that Performing time series data mining according to the time series operation data and the preset subsequence length to obtain time series operation features of the time series operation data includes: Taking a preset subsequence length as a sliding window length, and obtaining a plurality of time series operation subsequences of the time series operation data according to the time series operation data, a preset sliding window initial position, and a sliding window length; According to the time series operation data and the time series operation subsequences, calculating distance values ​​between the time series operation subsequences and the time series operation data; Determine a target timing operation feature according to the distance value, a plurality of timing operation subsequences and a preset distance threshold; The distance value corresponding to the target timing operation feature is used as a feature value, and according to the feature value and the target timing operation feature, the target timing operation feature is converted into a vector space to obtain the final timing operation feature of the timing operation data.

3. A method for extracting power grid operation scenarios based on GMM improved clustering as claimed in claim 1, It is characterized in that The time series running features are projected into a low-dimensional space for dimensionality reduction to obtain time series feature indicators after dimensionality reduction, including: Calculate, according to the time series operation characteristics, a first distance between adjacent time series operation characteristics and a second distance between any two time series operation characteristics; According to the first distance, the second distance and a preset smoothing normalization factor, a weight between any two time series operation features corresponding to the second distance is calculated to obtain a weighted adjacency matrix; Constructing a high-dimensional topology according to the weighted adjacency matrix and the timing operation characteristics to obtain a weighted neighbor graph; Projecting the time series operation features into a low-dimensional space, and calculating a low-dimensional weight function of the time series operation features in the low-dimensional space; According to the weighted neighbor graph and the low-dimensional weight function, a cross entropy function is constructed, and the cross entropy function is iteratively optimized by the stochastic gradient descent method until the preset loss function converges, and the time series feature index after dimensionality reduction is obtained according to the final cross entropy function.

4. A method for extracting power grid operation scenarios based on GMM improved clustering as claimed in claim 1, It is characterized in that According to the time series operation data, the time series characteristic index and the number of the time series characteristic indexes, a time series weight vector corresponding to each time series characteristic index is calculated, including: Calculate the entropy value according to the time series operation data and the number of time series characteristic indicators; According to the entropy value and the number of the time series feature indicators, the entropy value is corrected to calculate a first weight vector for representing the size of the information entropy of the time series feature indicator; According to the time series characteristic index, calculate the standard deviation and correlation coefficient of the time series characteristic index; Calculating a second weight vector for representing the correlation between the time series feature indicators according to the standard deviation, the correlation coefficient and the number of the time series feature indicators; A timing weight vector corresponding to each timing feature indicator is calculated based on the first weight vector and the second weight vector.

5. A method for extracting power grid operation scenarios based on GMM improved clustering as claimed in claim 1, It is characterized in that The time series feature indicators are clustered according to the constructed GMM model to obtain several initial cluster centers, including: Determine the initial Gaussian distribution parameters of the GMM model according to the constructed GMM model; Calculate the posterior probability according to the current Gaussian distribution parameters; and continuously update the current Gaussian distribution parameters according to the posterior probability until the preset first loss function converges to obtain the final Gaussian distribution parameters; cluster the time series feature indicators according to the final Gaussian distribution parameters to obtain several Gaussian distributions; randomly select a Gaussian distribution as the first first clustering center; According to the current first cluster center and the position of the time series characteristic index in the GMM model, the clustering probability of the time series characteristic index becoming the next first cluster center is calculated, and a new first cluster center is generated according to the clustering probability; the remaining time series characteristic indexes are divided into the new first cluster center; When all time series feature indicators are divided into new first cluster centers, several initial first cluster centers are obtained; among them, the current Gaussian distribution parameters of the initial calculation of the posterior probability are the initial Gaussian distribution parameters; the current cluster center of the initial calculation of the cluster probability is the first first cluster center.

6. A method for extracting power grid operation scenarios based on GMM improved clustering as claimed in claim 1, It is characterized in that The cluster center optimization operation includes: According to the current second cluster center, the time series feature index and the time series weight vector, the cluster distance between the time series feature index and the current second cluster center is calculated; wherein the current second cluster center for the first calculation of the cluster distance is the initial first cluster center; The current second cluster center corresponding to the minimum cluster distance of each time series feature indicator is used as the new current second cluster center corresponding to each time series feature indicator, until the current second cluster centers corresponding to all time series feature indicators remain unchanged, and the cluster center optimization operation is completed; and the final time series feature cluster center is obtained based on the current second cluster center.

7. A method for extracting power grid operation scenarios based on GMM improved clustering as claimed in claim 1, It is characterized in that Also includes: According to the final time series feature cluster center and the time series feature index, the average distance between all the time series feature indexes and the time series feature cluster center is calculated; According to the final time series feature clustering center, calculate the distance between every two time series feature clustering centers; The clustering evaluation coefficient is calculated based on the average distance and the distance between the centers of each two time series feature clusters; The final time series feature clustering center is evaluated according to the clustering evaluation coefficient.

8. A power grid operation scenario extraction device based on GMM improved clustering, It is characterized in that include: A time series data acquisition module is used to acquire the time series operation data of the power grid to be extracted; A time series feature mining module, used to perform time series data mining according to the time series operation data and a preset subsequence length, to obtain the time series operation features of the time series operation data; A time series feature dimensionality reduction module is used to project the time series operation features into a low-dimensional space for dimensionality reduction, and obtain time series feature indicators after dimensionality reduction; A timing weight calculation module, used to calculate a timing weight vector corresponding to each timing characteristic indicator according to the timing operation data, the timing characteristic indicator and the number of the timing characteristic indicators; A cluster center initialization module, used to construct a GMM model according to the time series feature index and the time series weight vector, and cluster the time series feature index according to the constructed GMM model to obtain a plurality of initial first cluster centers; A cluster center optimization module, used to repeatedly perform cluster center optimization operations according to a number of initial first cluster centers, the time series feature index and the time series weight vector, until a preset loss function converges, thereby obtaining a final time series feature cluster center; The power grid operation scenario extraction module is used to use the final time series feature clustering center as the power grid operation scenario of the power grid to be extracted.

9. A terminal device, It is characterized in that It comprises a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein when the processor executes the computer program, it implements a method for extracting power grid operation scenarios based on GMM improved clustering as described in any one of claims 1 to 7.

10. A computer-readable storage medium, It is characterized in that The computer-readable storage medium includes a stored computer program, wherein when the computer program is executed, the device where the computer-readable storage medium is located is controlled to execute a method for extracting power grid operation scenarios based on GMM improved clustering as described in any one of claims 1 to 7.

Citation Information

Cited By

  • Abnormal user behavior detection method based on space-time behavior characteristic data analysis

    CN119807974A

  • State detection method, device and equipment based on electronic nose and storage medium

    CN120429740A