Transformer abnormal state identification method and system

Abstract

The invention provides a transformer abnormal state identification method and system, and the method comprises the steps: constructing a digital twinborn model corresponding to a transformer entity, and storing the operation information and electrical parameter information of transformer equipment in a twinborn database; training a feature extraction algorithm model based on orthogonal global and local preserving embedding by using initial fault feature sample data in a twin database to obtain a trained feature extraction algorithm model; inputting the initial fault features of the current time period in the twin database into the trained feature extraction algorithm model to obtain extracted fault features; according to the method, a transformer abnormal state recognition model based on an LOF-GMM improved algorithm is constructed, the extracted fault feature data is imported into the transformer abnormal state recognition model based on the LOF-GMM, a transformer abnormal state recognition result is output, and the transformer abnormal state recognition reliability is effectively improved.

Description

Technical Field

[0001] The present invention relates to the field of transformers, and in particular to a method and system for identifying abnormal states of transformers. Background Art

[0002] In the power system, transformers play a key role, responsible for voltage regulation and energy transmission. However, transformers may encounter various potential problems and abnormal conditions in long-term operation, which may have a negative impact on the operation of the power grid and even endanger the stability and life of the equipment. Therefore, it is crucial to accurately and quickly identify and process abnormal data of transformers. With the rapid development of intelligent algorithms in the field of data analysis and processing, it is possible to use these algorithms to perform abnormal diagnosis of transformers. As representatives of intelligent algorithms, Apriori algorithm and LSTM model show great potential in data mining and sequence data processing. By applying these algorithms to abnormal diagnosis of transformers, the ability to detect the state of the power system can be improved, thereby enhancing the safety and stability of the equipment. Therefore, research on abnormal diagnosis of transformers has become crucial, which can provide effective means for the management and maintenance of power systems to ensure the reliable operation of the power grid.

[0003] The operating status of power transformers is related to the safety of power grid operation and the stability of social electricity consumption. Oil chromatography online monitoring technology is an important means to reflect the operating status of transformers. The data information of dissolved gas in oil is collected and transmitted by the monitoring device to form a time series arranged in chronological order. However, due to the influence of environmental factors, abnormalities are inevitable in the monitoring data.

[0004] At present, there are many methods for identifying abnormal conditions, such as density-based detection algorithms, density-based algorithms, neural network-based methods, and cluster analysis methods, which generally have problems such as slow convergence speed, high computational and storage overhead, and easy early stagnation, resulting in large deviations in the optimization results and low reliability in identifying transformer abnormal conditions.

[0005] In view of this problem, the present invention provides a method and system for identifying abnormal status of a transformer to solve the above problem. Summary of the invention

[0006] In order to solve the problems existing in the prior art, the present invention innovatively proposes a transformer abnormal state identification method and system, which effectively solves the problem of low reliability of transformer abnormal state identification caused by the prior art and effectively improves the reliability of transformer abnormal state identification.

[0007] A first aspect of the present invention provides a method for identifying an abnormal state of a transformer, comprising:

[0008] Build a digital twin model corresponding to the transformer entity, obtain the transformer equipment operation information and electrical parameter information, and save the transformer equipment operation information and electrical parameter information into the twin database;

[0009] Using the initial fault feature sample data in the twin database, the feature extraction algorithm model based on orthogonal global and local preservation embedding is trained to obtain the trained feature extraction algorithm model;

[0010] Input the initial fault features of the current time period in the twin database into the trained feature extraction algorithm model to obtain the extracted fault features;

[0011] A transformer abnormal state recognition model based on the improved LOF-GMM algorithm is constructed, the extracted fault feature data is imported into the LOF-GMM transformer abnormal state recognition model, and the transformer abnormal state recognition results are output.

[0012] A second aspect of the present invention provides a transformer abnormal state identification system, comprising:

[0013] A construction module is used to construct a digital twin model corresponding to the transformer entity, obtain the transformer equipment operation information and electrical parameter information, and save the transformer equipment operation information and electrical parameter information into the twin database;

[0014] The training module uses the initial fault feature sample data in the twin database to train the feature extraction algorithm model based on orthogonal global and local preservation embedding to obtain the trained feature extraction algorithm model;

[0015] The input module inputs the initial fault sample feature data of the current time period in the twin database into the trained feature extraction algorithm model to obtain the extracted fault features;

[0016] Import and output modules, build a transformer abnormal state recognition model based on the improved LOF-GMM algorithm, import the extracted fault feature data into the LOF-GMM transformer abnormal state recognition model, and output the transformer abnormal state recognition results.

[0017] The technical solution adopted by the present invention includes the following technical effects:

[0018] 1. The present invention constructs a digital twin model corresponding to a transformer entity, obtains transformer equipment operation information and electrical parameter information, and saves the transformer equipment operation information and electrical parameter information into a twin database; uses the initial fault feature sample data in the twin database to train a feature extraction algorithm model based on orthogonal global and local preservation embedding to obtain a trained feature extraction algorithm model; inputs the initial fault features of the current time period in the twin database into the trained feature extraction algorithm model to obtain extracted fault features; constructs a transformer abnormal state recognition model based on the LOF-GMM improved algorithm, imports the extracted fault feature data into the transformer abnormal state recognition model of the LOF-GMM, and outputs the transformer abnormal state recognition result, effectively solving the problem of low reliability of transformer abnormal state recognition caused by the prior art, and effectively improving the reliability of transformer abnormal state recognition.

[0019] 2. In the technical solution of the present invention, a virtual model is constructed by introducing digital twin technology, and its behavior and performance are consistent with the entity, which can reflect the operating conditions within the physical entity; the digital twin technology realizes functions such as data analysis and fault prediction based on the received data or historical data; the digital twin technology is used in combination with the improved LOF-GMM algorithm to map the operating status of the power transformer, and the digital twin model is optimized and updated through iterative training to realize status monitoring, realize abnormal status identification of the transformer, and save storage space.

[0020] 3. The technical solution of the present invention can determine the input of the transformer fault diagnosis model according to the feature selection results, and then import the data into the transformer abnormal state recognition model of LOF-GMM. The transformer abnormal state recognition model can realize automatic recognition of abnormal states through continuous iteration, accurately evaluate the state of the power transformer, judge whether there is a fault in advance, and arrange staff to carry out maintenance to prevent further threats to the safe operation of the power grid.

[0021] 4. In the technical solution of the present invention, based on the vector inner product operation, in the LOF algorithm calculation, the data object x new With y new The distance between them is calculated by the traditional algorithm using the Euclidean distance d(x new ,y new ) is improved to d'(x new ,y new ), further improving the reliability of transformer abnormal status identification.

[0022] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0024] Figure 1 This is a schematic diagram of a process of a method in Example 1 of the present invention;

[0025] Figure 2 A schematic diagram of the transformer fault diagnosis principle based on digital twins in the method of Example 1 in the solution of the present invention;

[0026] Figure 3 Schematic diagram of the process of step S2 in the method of embodiment 1 of the present invention;

[0027] Figure 4 Schematic diagram of the solution process of the mapping matrix in the method of Example 1 of the solution of the present invention;

[0028] Figure 5 Schematic diagram of the process of step S4 in the method of embodiment 1 of the present invention;

[0029] Figure 6 It is a schematic diagram of the structure of the device of Example 2 in the scheme of the present invention. DETAILED DESCRIPTION

[0030] In order to clearly illustrate the technical features of the present solution, the present invention is described in detail below through specific implementation methods and in conjunction with the accompanying drawings. The disclosure below provides many different embodiments or examples for realizing different structures of the present invention. In order to simplify the disclosure of the present invention, the components and settings of specific examples are described below. In addition, the present invention may repeat reference numbers and / or letters in different examples. This repetition is for the purpose of simplification and clarity, and does not itself indicate the relationship between the various embodiments and / or settings discussed. It should be noted that the components illustrated in the accompanying drawings are not necessarily drawn to scale. The present invention omits the description of known components and processing techniques and processes to avoid unnecessary limitations on the present invention.

[0031] Embodiment 1

[0032] like Figure 1 As shown, the present invention provides a method for identifying abnormal state of a transformer, comprising:

[0033] S1, construct a digital twin model corresponding to the transformer entity, obtain the transformer equipment operation information and electrical parameter information, and save the transformer equipment operation information and electrical parameter information into the twin database;

[0034] S2, using the initial fault feature sample data in the twin database, train the feature extraction algorithm model based on orthogonal global and local preservation embedding to obtain the trained feature extraction algorithm model;

[0035] S3, input the initial fault features of the current time period in the twin database into the trained feature extraction algorithm model to obtain the extracted fault features;

[0036] S4, construct a transformer abnormal state recognition model based on the LOF-GMM improved algorithm, import the extracted fault feature data into the LOF-GMM transformer abnormal state recognition model, and output the transformer abnormal state recognition result.

[0037] Among them, in step S1, if Figure 2 As shown in the figure, a digital twin model (virtual model) corresponding to the transformer entity is constructed. The digital twin integrates and fuses four layers of models: geometry, physics, behavior, and rules. In the digital twin, the physical part consists of a transformer and sensors, which collect equipment operation information through sensors and save it in the twin database. The database can store all information related to the transformer, and the diagnostic results can also be saved in real time. The behavior model and the rule model are the key to the twin.

[0038] The digital twin transformer is the digital embodiment of the physical transformer. Its biggest advantage is that it can obtain real-time change data of the physical transformer through sensors, and on this basis update, simulate, analyze and predict these data, and simulate the transformer operating status in real time to ensure the normal operation of the physical transformer.

[0039] Transformer fault diagnosis principle based on digital Li Sheng Figure 2 The digital twin model is a virtual model of the transformer. The collected transformer sample data and operating status are mapped to the digital twin model to simulate the transformer operating status. The LOF-GMM intelligent algorithm is used to iteratively optimize the digital twin model to update the abnormal status of the transformer.

[0040] The electrical parameters such as oil and gas data collected from the transformer entity are transmitted to the constructed digital twin model in real time through wireless transmission. After receiving the data, the digital twin model first pre-processes it, extracts and screens the fault features using a feature extraction algorithm based on orthogonal global and local preserving embedding (OGLPE), and determines the input of the transformer fault diagnosis model based on the feature selection results. Then the data is imported into the transformer abnormal state recognition model of LOF-GMM. The model realizes automatic recognition of abnormal states through continuous iteration. Accurately evaluate the status of the power transformer, determine whether there is a fault in advance, and arrange staff to repair it to prevent further threats to the safe operation of the power grid.

[0041] In step S2, in order to integrate the local and global structure information of the comprehensive sample, while ensuring data simplicity and maximizing the extraction of the fault characteristics of the transformer, a feature extraction algorithm based on orthogonal global and local preserving embedding (OGLPE) is adopted. This algorithm can complement the manifold learning method and the global variance maximization method, maintain the local non-linear structure of the data, and not destroy the global structure characteristics of the data. At the same time, information redundancy is removed through regularization requirements.

[0042] Specifically, as Figure 3 shown, using the initial fault feature sample data in the twin database to train the feature extraction algorithm model based on orthogonal global and local preserving embedding, the trained feature extraction algorithm model specifically includes:

[0043] S21, using the mapping function to transform the initial fault sample data set X in the twin database into the high-dimensional fault sample data set F in the high-dimensional kernel space;

[0044] Algorithm process: The initial fault sample data set X = [x 0 , x 1 , x 2 , x 3 , …, x m in the initial R (non-linear function) is transformed into the high-dimensional fault sample data set F in the high-dimensional kernel space R d1 , that is

[0045] S22, using the projection matrix A to map the initial fault sample data set X in the high-dimensional kernel space to the low-dimensional space, obtaining the low-dimensional embedded fault feature data Y corresponding to the initial fault sample data set X, and taking the low-dimensional embedded fault feature data Y as the extracted fault feature; among them, the projection matrix A is the product of the mapping function and the mapping matrix a;

[0046] Algorithm process: Using the projection matrix A to realize the mapping of the initial fault sample data set X in the high-dimensional kernel space to the low-dimensional space R d (d < d1), obtaining the formula of the low-dimensional embedded fault feature data Y of X as Y = [y 1 , y 2 , …, y m , and The core of this feature extraction algorithm is to construct an objective function representing its features and calculate the projection matrix A.

[0047] Using the projection matrix A to complete the conversion of the data set from the initial coordinates to the new coordinates: This ensures that the overall information is maximized, because the projection matrix A is composed of So there exists a mapping matrix a so that the projection matrix The complex problem of inner product operation of high-dimensional vectors is solved based on the kernel algorithm.

[0048] S23, according to the initial fault sample data set X in the twin database, the high-dimensional fault sample data set F in the high-dimensional kernel space, and the low-dimensional embedded fault feature data Y corresponding to the initial fault sample data set X, the feature extraction algorithm model based on orthogonal global and local preserving embedding is trained to obtain the mapping matrix a and the projection matrix A in the trained feature extraction algorithm model.

[0049] Among them, Figure 4 As shown, the determination process of the mapping matrix a is specifically as follows:

[0050] S201, define the objective function of the projection matrix A, where the initial objective function J of the projection matrix A is 1 for:

[0051]

[0052] in, is the expression of the high-dimensional fault sample data set F in the high-dimensional kernel space, tr refers to the trace of the matrix, usually expressed as tr(), and the calculation method is to add the elements on the diagonal of the matrix;

[0053] S202, define the kernel matrix K, convert the initial objective function of the projection matrix A into an intermediate objective function expressed by the mapping matrix a and the kernel matrix K, where the intermediate objective function J of the projection matrix A is 2 for:

[0054]

[0055] Among them, K ij is the i-th initial fault sample data x i With the jth initial fault sample data x j The corresponding Gaussian kernel function, δ is the width of the Gaussian kernel, is the i-th initial fault sample data x i Convert to high-dimensional fault sample data in high-dimensional kernel space, is the jth initial fault sample data x j Convert high-dimensional fault sample data into high-dimensional kernel space;

[0056] S203, adding orthogonal constraints, removing redundant feature information, and obtaining the mapping component a in the mapping matrix a kOrthogonality constraint equations;

[0057]

[0058] Among them, the mapping component a k is the kth column vector in the mapping matrix a, a j is the mapping component a in the mapping matrix a k The jth column vector that constitutes the orthogonal constraint; that is, the jth initial fault sample data x j The number of columns of the corresponding mapping matrix a when converted into low-dimensional embedded fault feature data;

[0059] S204, using the k-nearest neighbor method to establish a neighbor undirected graph, selecting neighbor initial fault samples in the converted high-dimensional kernel space, and calculating the neighbor similarity between the neighbor initial fault samples to establish a neighbor similarity matrix between the neighbor initial fault samples;

[0060] S205, using the Lagrange multiplier method, transforms the mapping component a in the mapping matrix a k The calculation problem is transformed into the problem of solving the constrained maximum value calculation; after multiple iterations, several orthogonal mapping vectors are obtained, and the kth mapping vector a of the mapping matrix a is solved. k , and get the mapping matrix a=[a 1 ,a 2 ,…,a k ].

[0061] Among them, in step S204, the k-nearest neighbor method is used to establish a neighbor undirected graph, the neighbor initial fault samples are selected in the converted high-dimensional kernel space, and the neighbor similarity between the neighbor initial fault samples is calculated, and the neighbor similarity matrix between the neighbor initial fault samples is established as follows:

[0062] Construct an adjacent undirected graph to satisfy the similarity between non-neighbor samples to be 0, and select high-dimensional fault sample data in the transformed high-dimensional kernel space through the Euclidean distance formula k nearest neighbors; among them, high-dimensional fault sample data The selection formula for k nearest neighbors is:

[0063]

[0064] in, for With each neighboring point The similarity between ij The acquisition function, w ij for With each neighboring point The similarity between is high-dimensional fault sample data With high-dimensional fault sample data The Euclidean distance between

[0065] Among them, the calculation formula of the Euclidean distance between high-dimensional fault sample data is specifically as follows:

[0066]

[0067] Using the kernel technique, the calculation formula of the Euclidean distance between high-dimensional fault sample data is converted into the expression of the kernel function:

[0068]

[0069] Among them, K ii is the i-th initial fault sample data x i With the i-th initial fault sample data x i The corresponding Gaussian kernel function, K ij is the i-th initial fault sample data x i With the jth initial fault sample data x j The corresponding Gaussian kernel function, K jj is the jth initial fault sample data x j With the jth initial fault sample data x j The corresponding Gaussian kernel function;

[0070] By solving the selection formula, we get With each neighboring point The similarity between ij (j=1,2,…,k-1), forming the similarity matrix W, it can be seen that Neighboring points with closer Euclidean distances are assigned greater similarity values.

[0071] In step S205, the mapping component a in the mapping matrix a is transformed into k The calculation problem is transformed into the problem of solving the constrained maximum value calculation; after multiple iterations, several orthogonal mapping vectors are obtained, and the kth mapping vector a of the mapping matrix a is solved. k , and get the mapping matrix a=[a 1 ,a 2 ,…,a k ]Specifically:

[0072] Calculate the diagonal matrix D and the intermediate matrix M of the similarity matrix W;

[0073] Among them, the diagonal elements in the diagonal matrix Intermediate matrix M = (DW) T (DW);

[0074] The intermediate objective function J of the joint projection matrix A2 , including the mapping component a in the mapping matrix a k The orthogonal constraint equation is transformed by applying the Lagrange multiplier to transform the mapping component a in the mapping matrix a k The calculation problem is to solve the constraint maximum value calculation problem. The solution function of the kth component of the mapping matrix a is expressed as follows:

[0075]

[0076] Among them, st. represents the introduction of restriction conditions to eliminate the influence of random scale factors, I is the unit matrix; λ and μ are both Lagrange multiplier vectors;

[0077] Using L(a k ) for the mapping component a k Take the partial derivative and put it after 0 to get the relationship equation:

[0078]

[0079] Define the first matrix variable A respectively k-1 =[a 1 ,a 2 ,…,a k-1 ] and the second matrix variable U k-1 =[μ 1 ,μ 2 ,…,μ k-1 ] T , a i is the i-th column vector in the mapping matrix a; that is, the i-th initial fault sample data x i The number of columns of the corresponding mapping matrix a when converted into low-dimensional embedded fault feature data;

[0080] Among them, the element a in the first matrix variable k-1 is the k-1th mapping vector of the mapping matrix a, μ k-1 is the Lagrange multiplier vector;

[0081] The calculation equation matrix is ​​established as:

[0082] (A k-1 ) T (K) -1 A k-1 U k-1 =2(A k-1 ) T (D) -1 Ma k (6)

[0083] After multiple iterations, several orthogonal mapping vectors are obtained, a k The calculation formula is as follows:

[0084] [I-2(A k-1 ) T (K) -1 (A k-1 ) -1 (A k-1 ) T ]×D -1 K(DM)Ka k =λa k .

[0085] By solving the eigenvector corresponding to the maximum eigenvalue in the above formula, the kth mapping vector a of the mapping matrix a can be solved: k , forming a mapping matrix a=[a 1 ,a 2 ,…,a k ], the projection matrix can be obtained And according to The low-dimensional embedded fault-sensitive features of each training sample are obtained.

[0086] Among them, in step S3, the initial fault features of the current time period in the twin database are input into the trained feature extraction algorithm model, and the extracted fault features are specifically:

[0087] For the initial fault feature data x in the current time period in the twin database new , input into the trained feature extraction algorithm model to obtain the corresponding extracted fault feature y new for:

[0088]

[0089] like Figure 5 As shown, in step S4, a transformer abnormal state recognition model based on the LOF-GMM improved algorithm is constructed, the extracted fault feature data is imported into the transformer abnormal state recognition model of the LOF-GMM, and the output transformer abnormal state recognition result specifically includes:

[0090] S41, obtain the initial fault characteristic data x of the current time period new , and the corresponding extracted fault feature data y new ;

[0091] S42, calculate the initial fault characteristic data x in the current time period new , and the corresponding extracted fault feature data y new The distance between

[0092] Among them, calculate the initial fault feature data object x in the current time period new , and the corresponding extracted fault feature data y newThe specific distance between them is:

[0093]

[0094] Among them, sim(x new ,y new ) is the initial fault feature data object x in the current time period new , and the corresponding extracted fault feature data y new The Euclidean distance between new ,y new ) is the initial fault feature data object x in the current time period new , and the corresponding extracted fault feature data y new The distance between is the data object of the i-th initial fault feature data in the current time period, It is the fault feature data extracted from the i-th initial fault feature data object in the current time period.

[0095] In the traditional LOF algorithm, when the sample data is fixed, the final LOF value of the algorithm is only related to its own parameter g. When g takes different values, it may lead to deviations in the judgment of abnormal data. Therefore, an improved LOF algorithm is proposed based on vector inner product operation. new With y new The distance between them is calculated by the traditional algorithm using the Euclidean distance d(x new ,y new ) is improved to d'(x new ,y new ). In the formula, the data object and are two vectors, d'(x new ,y new ) represents the distance between two vectors.

[0096] S43, based on the initial fault characteristic data x in the current time period new , and the corresponding extracted fault feature data y new The distance between them is calculated to obtain the LOF value of each load point;

[0097] Among them, according to the initial fault characteristic data x in the current time period new , and the corresponding extracted fault feature data y new The distance between them is calculated, and the LOF average value of each load point is specifically:

[0098]

[0099] Among them, LOF(x new, ynew ) is the load point (x new, y new )’s LOF average, LOF g (x newi ) represents the i-th initial fault feature data x in the current time period newi The local anomaly factor, g is the input parameter of the algorithm itself, Minpts is the minimum value of the solvable range of parameter g, and Minpts is the maximum value of the solvable range of parameter g;

[0100] The i-th initial fault feature data x in the current time period newi The calculation formula of the local anomaly factor is as follows:

[0101]

[0102] in, is the i-th initial fault feature data point x in the current time period newi and surrounding data points y newi The density within the distance g (with respect to the data point x newi All data points y are within the distance of g. newi To data point x newi The inverse of the average distance of Represents the initial fault feature data x of the current time period i newi All data points y within the distance g newi The number of is the fault feature data point after extraction With surrounding data points The density within the distance g. Representation: With the extracted fault feature data points All data points x within a distance of g newi The number of LOF g (x newi ) means: with data point x newi Data points within the distance g and The average value of the ratio. g (x newi )The larger the value, the more likely the data point x newi The more likely it is to be an abnormal data point.

[0103] S44, use the GMM model to perform cluster analysis on the LOF value, perform clustering operations according to normal data and abnormal data, and use the LOF value of the sample normal data and abnormal data boundary obtained by cluster analysis as the abnormal judgment threshold; cyclically compare the LOF values ​​of all data in the load point in the current time period with the abnormal judgment threshold, and judge the load point with a LOF value less than the threshold as a normal data state, and judge the load point with a LOF value not less than the threshold as an abnormal data state; wherein the load point is the data object x of the initial fault feature data of the current time period new , and the corresponding extracted fault feature data y new The sample points composed of.

[0104] Among them, GMM is an unsupervised clustering algorithm that linearly combines multiple single Gaussian models to make the model fit a more responsible sample distribution. Since the parameter π of a single Gaussian model in the GMM model g ,μ g ,∑ g are all unknown quantities, π g is the mixing weight of the gth (same as the algorithm input parameter g) Gaussian distribution, μ g Mean, ∑ g Covariance matrix. Usually, the unknown parameters need to be iterated through the EM algorithm. The EM algorithm is divided into expectation calculation (Expectation, E step) and maximization (Maximization, M step): the first step is to estimate the hidden variables in each Gaussian model through the E step, randomly initialize the parameters or based on the results of the previous iteration; the second step is the M step, based on the hidden variable results estimated in the previous step, the parameters of the Gaussian distribution are inferred; finally, the EM algorithm steps are executed repeatedly until the GMM log-likelihood function converges or reaches the iteration limit. The specific calculation method is as follows:

[0105] ① Randomly initialize the parameters of the EM algorithm (The three parameters are the parameters of a single Gaussian model in the GMM model, g is the g-th Gaussian distribution, which is unique. u is the mean vector, and Σ is the covariance matrix), execute the E step in the algorithm, the formula is as follows. Among them, π g is the weight of the g-th Gaussian model in GMM, N(x new |μ g ,∑ g ) represents the g-th Gaussian distribution with mean μ g , the covariance matrix is ​​∑ g The single Gaussian model probability distribution function (N represents the single Gaussian model probability distribution function, and the covariance matrix is ​​expressed as ∑ g ) Calculate the LOF value of the sample point of the i-th initial fault feature data object in the current time period The posterior probability γ(t,g) of the Gaussian mixture model in the tth iteration, that is, after this iteration, each sample The probability of belonging to the g-th Gaussian model.

[0106] (G is the total number of Gaussian distributions (also called components)).

[0107] ②Execute the M step of the algorithm, and the calculation formula is as follows. The parameter estimate in the tth iteration is obtained by calculating the maximum likelihood function of the parameter According to the results, the relevant parameters in each Gaussian model are updated to obtain the GMM model after the tth iteration.

[0108]

[0109] ③ Calculate the log-likelihood function L of the GMM model after the tth iteration t , calculated as follows. Alternately execute the E and M steps of the algorithm until the likelihood function converges or the number of iterations reaches the set maximum value, and then terminate the algorithm. The significance of this step is to estimate the parameters by maximizing the joint probability of the LOF value sample set and select the best distribution model.

[0110]

[0111] The GMM algorithm uses a weighted combination of multiple Gaussian distribution functions to characterize the sample distribution. It can combine the spatial distance of the samples and the distribution density of the sample points, and has a good clustering effect. Therefore, GMM is selected as the LOF threshold partition clustering algorithm.

[0112] The GMM model is used to cluster the LOF values ​​into two categories: normal data and abnormal data. The LOF value of the sample normal data and abnormal data dividing line obtained by adaptive clustering is used as the abnormal judgment threshold. The LOF value of all data in the sample is cyclically compared with the abnormal threshold. Points with LOF values ​​less than the threshold are judged as normal data states, otherwise they are judged as abnormal data states.

[0113] Digital twins use the established digital models to update and dynamically present entity behaviors in real time, which can directly reflect the operating status of the system, as well as conduct real analysis and comprehensive detection of unpredictable situations. The present invention comprehensively utilizes methods such as data fusion analysis, digital twin technology, and LOF-GMM improved algorithm to provide a high-precision, high-efficiency, and intelligent operating mechanism and operation service for the transformer system. The collected transformer data is transferred to the virtual model to map the real state of the transformer. The digital twin model is updated by iterative optimization using the LOF-GMM improved algorithm to monitor the operating status of the transformer.

[0114] Abnormal data identification of transformers based on digital twin technology can achieve virtual and real synchronization, and timely obtain the operating status of the transformer, providing a strong guarantee for the smooth operation of the transformer.

[0115] By continuously collecting the real-time operating status of the transformer, information fusion is realized in the digital platform, equipment fault information is extracted, and fault events are quickly perceived using specific algorithms based on the separated feature data.

[0116] The present invention constructs a digital twin model corresponding to a transformer entity, obtains transformer equipment operation information and electrical parameter information, and saves the transformer equipment operation information and electrical parameter information into a twin database; uses initial fault feature sample data in the twin database to train a feature extraction algorithm model based on orthogonal global and local preservation embedding to obtain a trained feature extraction algorithm model; inputs the initial fault features of the current time period in the twin database into the trained feature extraction algorithm model to obtain extracted fault features; constructs a transformer abnormal state recognition model based on an improved LOF-GMM algorithm, imports the extracted fault feature data into the transformer abnormal state recognition model of LOF-GMM, and outputs a transformer abnormal state recognition result, thereby effectively solving the problem of low reliability of transformer abnormal state recognition caused by the prior art and effectively improving the reliability of transformer abnormal state recognition.

[0117] In the technical solution of the present invention, a virtual model is constructed by introducing digital twin technology, and its behavior and performance are consistent with the entity, which can reflect the operating conditions within the physical entity; the digital twin technology realizes functions such as data analysis and fault prediction based on the received data or historical data; the digital twin technology is used in combination with the LOF-GMM improved algorithm to map the operating status of the power transformer, and the digital twin model is optimized and updated through iterative training to realize status monitoring, realize abnormal status identification of the transformer, and save storage space.

[0118] The technical solution of the present invention can determine the input quantity of the transformer fault diagnosis model according to the feature selection result, and then import the data into the transformer abnormal state recognition model of LOF-GMM. The transformer abnormal state recognition model realizes automatic recognition of abnormal state through continuous iteration, accurately evaluates the state of the power transformer, judges in advance whether there is a fault, and arranges staff to carry out maintenance to prevent further threats to the safe operation of the power grid.

[0119] In the technical solution of the present invention, based on the vector inner product operation, in the LOF algorithm calculation, the data object x new With y new The distance between them is calculated by the traditional algorithm using the Euclidean distance d(x new ,y new) is improved to d'(x new ,y new ), further improving the reliability of transformer abnormal status identification.

[0120] Embodiment 2

[0121] like Figure 6 As shown, the technical solution of the present invention also provides a transformer abnormal state identification system, comprising:

[0122] Building module 101, building a digital twin model corresponding to the transformer entity, obtaining transformer equipment operation information and electrical parameter information, and saving the transformer equipment operation information and electrical parameter information into a twin database;

[0123] The training module 102 uses the initial fault feature sample data in the twin database to train the feature extraction algorithm model based on orthogonal global and local preservation embedding to obtain a trained feature extraction algorithm model;

[0124] An input module 103 inputs the initial fault sample feature data of the current time period in the twin database into the trained feature extraction algorithm model to obtain the extracted fault features;

[0125] The import and output module 104 constructs a transformer abnormal state recognition model based on the LOF-GMM improved algorithm, imports the extracted fault feature data into the transformer abnormal state recognition model of the LOF-GMM, and outputs the transformer abnormal state recognition result.

[0126] The implementation process in the construction module 101, the training module 102, the input module 103, and the import and output module 104 in this embodiment corresponds to the specific process of the steps in the first embodiment, and will not be repeated in this embodiment.

[0127] The present invention constructs a digital twin model corresponding to a transformer entity, obtains transformer equipment operation information and electrical parameter information, and saves the transformer equipment operation information and electrical parameter information into a twin database; uses initial fault feature sample data in the twin database to train a feature extraction algorithm model based on orthogonal global and local preservation embedding to obtain a trained feature extraction algorithm model; inputs the initial fault features of the current time period in the twin database into the trained feature extraction algorithm model to obtain extracted fault features; constructs a transformer abnormal state recognition model based on an improved LOF-GMM algorithm, imports the extracted fault feature data into the transformer abnormal state recognition model of LOF-GMM, and outputs a transformer abnormal state recognition result, thereby effectively solving the problem of low reliability of transformer abnormal state recognition caused by the prior art and effectively improving the reliability of transformer abnormal state recognition.

[0128] In the technical solution of the present invention, a virtual model is constructed by introducing digital twin technology, and its behavior and performance are consistent with the entity, which can reflect the operating conditions within the physical entity; the digital twin technology realizes functions such as data analysis and fault prediction based on the received data or historical data; the digital twin technology is used in combination with the LOF-GMM improved algorithm to map the operating status of the power transformer, and the digital twin model is optimized and updated through iterative training to realize status monitoring, realize abnormal status identification of the transformer, and save storage space.

[0129] The technical solution of the present invention can determine the input quantity of the transformer fault diagnosis model according to the feature selection result, and then import the data into the transformer abnormal state recognition model of LOF-GMM. The transformer abnormal state recognition model realizes automatic recognition of abnormal state through continuous iteration, accurately evaluates the state of the power transformer, judges in advance whether there is a fault, and arranges staff to carry out maintenance to prevent further threats to the safe operation of the power grid.

[0130] In the technical solution of the present invention, based on the vector inner product operation, in the LOF algorithm calculation, the data object x new With y new The distance between them is calculated by the traditional algorithm using the Euclidean distance d(x new ,y new ) is improved to d'(x new ,y new ), further improving the reliability of transformer abnormal status identification.

[0131] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A method for identifying abnormal state of a transformer, characterized in that: include: Build a digital twin model corresponding to the transformer entity, obtain the transformer equipment operation information and electrical parameter information, and save the transformer equipment operation information and electrical parameter information into the twin database; Using the initial fault feature sample data in the twin database, the feature extraction algorithm model based on orthogonal global and local preservation embedding is trained to obtain the trained feature extraction algorithm model; Input the initial fault features of the current time period in the twin database into the trained feature extraction algorithm model to obtain the extracted fault features; A transformer abnormal state recognition model based on the improved LOF-GMM algorithm is constructed, the extracted fault feature data is imported into the LOF-GMM transformer abnormal state recognition model, and the transformer abnormal state recognition results are output.

2. A transformer abnormal state identification method according to claim 1, characterized in that: Using the initial fault feature sample data in the twin database, the feature extraction algorithm model based on orthogonal global and local preservation embedding is trained. The trained feature extraction algorithm model specifically includes: Using the mapping function Convert the initial fault sample data set X in the twin database to a high-dimensional fault sample data set F in the high-dimensional kernel space; Use the projection matrix A to map the initial fault sample data set X in the high-dimensional kernel space to the low-dimensional space, obtain the low-dimensional embedded fault feature data Y corresponding to the initial fault sample data set X, and use the low-dimensional embedded fault feature data Y as the extracted fault feature; where the projection matrix A is the mapping function The product with the mapping matrix a; According to the initial fault sample data set X in the twin database, the high-dimensional fault sample data set F in the high-dimensional kernel space, and the low-dimensional embedded fault feature data Y corresponding to the initial fault sample data set X, the feature extraction algorithm model based on orthogonal global and local preserving embedding is trained to obtain the mapping matrix a and the projection matrix A in the trained feature extraction algorithm model.

3. A transformer abnormal state identification method according to claim 2, characterized in that: The specific process of determining the mapping matrix a is: Define the objective function of the projection matrix A, where the initial objective function J1 of the projection matrix A is: in, is the expression of the high-dimensional fault sample data set F in the high-dimensional kernel space; Define the kernel matrix K, and convert the initial objective function of the projection matrix A into an intermediate objective function expressed by the mapping matrix a and the kernel matrix K, where the intermediate objective function J2 of the projection matrix A is: Among them, K ij is the i-th initial fault sample data x i With the jth initial fault sample data x j The corresponding Gaussian kernel function, δ is the width of the Gaussian kernel, is the i-th initial fault sample data x i Convert to high-dimensional fault sample data in high-dimensional kernel space, is the jth initial fault sample data x j Convert high-dimensional fault sample data into high-dimensional kernel space; Add orthogonal constraints, remove redundant feature information, and obtain the mapping component a in the mapping matrix a k Orthogonality constraint equations; Among them, the mapping component a k is the kth column vector in the mapping matrix a, a j is the mapping component a in the mapping matrix a k The j-th column vector constituting the orthogonality constraint; The k-nearest neighbor method is used to establish a neighbor undirected graph, and the neighbor initial fault samples are selected in the converted high-dimensional kernel space, and the neighbor similarity between the neighbor initial fault samples is calculated to establish the neighbor similarity matrix between the neighbor initial fault samples; By using the Lagrange multiplier method, the mapping component a in the mapping matrix a is k The calculation problem is transformed into the problem of solving the constrained maximum value calculation; after multiple iterations, several orthogonal mapping vectors are obtained, and the kth mapping vector a of the mapping matrix a is solved. k , and get the mapping matrix a=[a1,a2,…,a k ].

4. A transformer abnormal state identification method according to claim 3, characterized in that: The k-nearest neighbor method is used to establish a neighbor undirected graph. The neighbor initial fault samples are selected in the converted high-dimensional kernel space, and the neighbor similarity between the neighbor initial fault samples is calculated. The neighbor similarity matrix between the neighbor initial fault samples is established as follows: Construct an adjacent undirected graph to satisfy the similarity between non-neighbor samples to be 0, and select high-dimensional fault sample data in the transformed high-dimensional kernel space through the Euclidean distance formula k nearest neighbors; among them, high-dimensional fault sample data The selection formula for k nearest neighbors is: in, for With each neighboring point The similarity between ij The acquisition function, w ij for With each neighboring point The similarity between is high-dimensional fault sample data With high-dimensional fault sample data The Euclidean distance between Among them, the calculation formula of the Euclidean distance between high-dimensional fault sample data is specifically as follows: Using the kernel technique, the calculation formula of the Euclidean distance between high-dimensional fault sample data is converted into the expression of the kernel function: Among them, K ii is the i-th initial fault sample data x i With the i-th initial fault sample data x i The corresponding Gaussian kernel function, K ij is the i-th initial fault sample data x i With the jth initial fault sample data x j The corresponding Gaussian kernel function, K jj is the jth initial fault sample data x j With the jth initial fault sample data x j The corresponding Gaussian kernel function; By solving the selection formula, we get With each neighboring point The similarity between ij , forming a similarity matrix W.

5. A transformer abnormal state identification method according to claim 3, characterized in that: By using the Lagrange multiplier method, the mapping component a in the mapping matrix a is k The calculation problem is transformed into the problem of solving the constrained maximum value calculation; after multiple iterations, several orthogonal mapping vectors are obtained, and the kth mapping vector a of the mapping matrix a is solved. k , and get the mapping matrix a=[a1,a2,…,a k ]Specifically: Calculate the diagonal matrix D and the intermediate matrix M of the similarity matrix W; Among them, the diagonal elements in the diagonal matrix ), intermediate matrix M = (DW) T (DW); The intermediate objective function J2 of the joint projection matrix A includes the mapping component a in the mapping matrix a k The orthogonal constraint equation is transformed by applying the Lagrange multiplier to transform the mapping component a in the mapping matrix a k The computational problem is to solve the constraint maximum value calculation problem, and the kth component of the mapping matrix a solves the function L(a k ) is expressed as follows: Among them, st. represents the introduction of restriction conditions, I is the unit matrix; λ and μ are both Lagrange multiplier vectors; Using L(a k ) for the mapping component a k Take the partial derivative and put it after 0 to obtain the relationship equation: Define the first matrix variable A respectively k-1 =[a1,a2,…,a k-1 ] and the second matrix variable U k-1 =[μ1,μ2,…,μ k-1 ] T , Among them, the element a in the first matrix variable k-1 is the k-1th mapping vector of the mapping matrix a, μ k-1 is the Lagrange multiplier vector; The calculation equation matrix is ​​established as: (A k-1 ) T (K) -1 A k-1 U k-1 =2(A k-1 ) T (D) -1 Ma k (6), After multiple iterations, several orthogonal mapping vectors are obtained, a k The calculation formula is as follows: [I-2(A k-1 ) T (K) -1 (A k-1 ) -1 (A k-1 ) T ]×D -1 K(D-M)Ka k =λa k 。 6. A transformer abnormal state identification method according to claim 2, characterized in that: The initial fault features of the current time period in the twin database are input into the trained feature extraction algorithm model, and the extracted fault features are as follows: For the initial fault feature data x in the current time period in the twin database new , input into the trained feature extraction algorithm model to obtain the corresponding extracted fault feature y new for: y new =A T φ(x new )。 7. A transformer abnormal state identification method according to claim 1, characterized in that: A transformer abnormal state recognition model based on the improved LOF-GMM algorithm is constructed, and the extracted fault feature data is imported into the transformer abnormal state recognition model of LOF-GMM. The output transformer abnormal state recognition results include: Get the initial fault feature data x for the current time period new , and the corresponding extracted fault feature data y new ; Calculate the initial fault feature data x for the current time period new , and the corresponding extracted fault feature data y new The distance between According to the initial fault characteristic data x in the current time period new , and the corresponding extracted fault feature data y new After calculating the LOF value of each load point, the GMM model is used to perform cluster analysis on the LOF value, and clustering operations are performed according to normal data and abnormal data. The LOF value of the boundary between normal data and abnormal data obtained by cluster analysis is used as the abnormal judgment threshold; the LOF value of all data in the load point in the current time period is cyclically compared with the abnormal judgment threshold, and the load point with a LOF value less than the threshold is judged as a normal data state, and the load point with a LOF value not less than the threshold is judged as an abnormal data state; where the load point is the data object x of the initial fault feature data of the current time period new , and the corresponding extracted fault feature data y new The sample points composed of.

8. A transformer abnormal state identification method according to claim 7, characterized in that: Calculate the initial fault feature data object x for the current time period new , and the corresponding extracted fault feature data y new The specific distance between them is: Among them, sim(x new ,y new ) is the initial fault feature data object x in the current time period new , and the corresponding extracted fault feature data y new The Euclidean distance between new ,y new ) is the initial fault feature data object x in the current time period new , and the corresponding extracted fault feature data y new The distance between.

9. A transformer abnormal state identification method according to claim 8, characterized in that: According to the initial fault characteristic data x in the current time period new , and the corresponding extracted fault feature data y new The distance between them is calculated, and the LOF average value of each load point is specifically: Among them, LOF(x new, y new ) is the load point (x new, y new )’s LOF average, LOF g (x newi ) represents the i-th initial fault feature data x in the current time period newi The local anomaly factor, g is the input parameter of the algorithm itself, Minpts is the minimum value of the solvable range of parameter g, and Minpts is the maximum value of the solvable range of parameter g; The i-th initial fault feature data x in the current time period newi The calculation formula of the local anomaly factor is as follows: in, is the i-th initial fault feature data point x in the current time period newi and surrounding data points y newi The density within the distance g; N g (x newi ) represents the initial fault feature data x of the current time period i newi All data points y within the distance g newi The number of g (y newi ) is the extracted fault feature data point y newi With surrounding data points x newi The density within the distance g; N g (y newi ) represents the fault feature data point y after extraction newi All data points x within a distance of g newi The number of .

10. A transformer abnormal state identification system, characterized in that: include: A construction module is used to construct a digital twin model corresponding to the transformer entity, obtain the transformer equipment operation information and electrical parameter information, and save the transformer equipment operation information and electrical parameter information into the twin database; The training module uses the initial fault feature sample data in the twin database to train the feature extraction algorithm model based on orthogonal global and local preservation embedding to obtain the trained feature extraction algorithm model; The input module inputs the initial fault sample feature data of the current time period in the twin database into the trained feature extraction algorithm model to obtain the extracted fault features; Import and output modules, build a transformer abnormal state recognition model based on the improved LOF-GMM algorithm, import the extracted fault feature data into the LOF-GMM transformer abnormal state recognition model, and output the transformer abnormal state recognition results.

Images