An on-line monitoring method for voltage transformer based on independent component analysis

Abstract

The application discloses an online monitoring method of a voltage transformer based on independent component analysis, which samples information of historical data, steady-state data and real-time data of signals output by the voltage transformer, constructs a data set from the historical data, and imports the data set into an initial SDAE network model based on a sparse denoising autoencoder to perform dynamic training, uses the steady-state data to fine-tune parameters based on the SDAE network model obtained through offline training, and thus obtains encoding data of the SDAE network model; uses an independent component analysis method to perform independent component decomposition by taking the encoding data as input; calculates sample statistics and overall statistical threshold, compares real-time statistics with the overall statistical threshold, and if the real-time statistics are lower than the overall statistical threshold, it is determined that the state of the voltage transformer is normal in this round of judgment; otherwise, it is considered that there is an abnormal voltage transformer in the voltage transformer group.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of online monitoring of power equipment, and particularly to an online monitoring method for voltage transformers based on independent component analysis. BACKGROUND

[0002] In modern power systems, voltage transformers, as a kind of vital electrical equipment, play an indispensable role. They are mainly used to convert high voltage into low voltage in proportion, and provide accurate and reliable voltage signals for the measurement, protection, control and other links of power systems. The accuracy of measurement and the reliability of operation of voltage transformers are directly related to the safe and stable operation of power systems, the evaluation of power quality and the accuracy of power metering.

[0003] In actual power system operation, voltage transformers face complex problems such as three-phase imbalance and harmonic dynamic disturbance. On the one hand, due to factors such as three-phase load asymmetry and single-phase large-capacity load access, three-phase voltage imbalance is common. Three-phase imbalance can cause three-phase excitation current imbalance of voltage transformers, thereby increasing the error of the transformers and affecting the measurement accuracy. If the transformers are in a three-phase imbalance operating environment for a long time, it may also accelerate the insulation aging and shorten the service life. On the other hand, with the widespread application of power electronic technology, a large number of nonlinear loads (such as frequency converters, rectifiers, electric arc furnaces, etc.) are connected to power systems, generating rich harmonics. Harmonics can saturate the core of voltage transformers, leading to increased excitation current and increased loss, while affecting the frequency response characteristics of the transformers, resulting in large errors in measurement results. Moreover, harmonics have dynamic variation characteristics, which makes the operating state of the transformers more complex and greatly increases the difficulty of monitoring and diagnosis.

[0004] Therefore, based on the above technical problems, the present application provides an online monitoring method for voltage transformers based on independent component analysis, which can effectively process data characteristics, improve monitoring accuracy and enhance model robustness. SUMMARY

[0005] The present application aims to overcome the shortcomings of the prior art and provide an online monitoring method for voltage transformers based on independent component analysis, which can effectively process data characteristics, improve monitoring accuracy and enhance model robustness.

[0006] To achieve the above-mentioned purpose, the present application provides an online monitoring method for voltage transformers based on independent component analysis, characterized by comprising the following steps:

[0007] S1 collects historical data, steady-state data and real-time data based on the output signals of the voltage transformer group of a substation, and constructs a data set through the historical data;

[0008] S2 creates an initial SDAE network model through sparse denoising auto-encoder;

[0009] S3 imports the data set into the initial SDAE network model created in step S2 for offline training, thereby obtaining an optimal SDAE network model;

[0010] S4 imports the steady-state data into the SDAE network model for correction training, and obtains a corrected SDAE network model when the number of training times of the SDAE network model reaches a preset value or the iteration effect reaches a preset purpose.

[0011] S5 imports the steady-state data into the SDAE network model for encoding, and then performs data standardization processing;

[0012] S6 takes the standardized encoded data as input, performs independent component decomposition through an independent component analysis method, and calculates sample statistics .

[0013] S7 calculates the overall statistical threshold of the sample statistics through kernel density estimation .

[0014] S8 collects real-time data, and calculates real-time statistics through steps S5-S6 .

[0015] S9 compares the real-time statistics with the overall statistical threshold to determine whether the voltage transformer set has failed, wherein when the real-time statistics are greater than the overall statistical threshold , the voltage transformer set has failed; otherwise, when the real-time statistics are less than the overall statistical threshold , the voltage transformer set has not failed.

[0016] S10 repeats steps S1-S9 to perform real-time evaluation on the voltage transformer set.

[0017] Further, the data obtained after mixing a certain proportion of noise signals into the input sample, the weight coefficient for controlling the sparse penalty term, the number of hidden layer neurons , the sparse parameter , and the average activation of the jth neuron under the sample data are introduced, thereby constructing the loss function of the SDAE network model.

[0018] Further, the independent component analysis method ICA performs independent component decomposition, including the following steps:

[0019] A1 steady-state data matrix construction, decomposing the steady-state data into a mixing matrix A and an independent component matrix S;

[0020] A2 data standardization, standardizing the steady-state data matrix to obtain a matrix , so that the mean of each column element is 0 and the standard deviation is 1;

[0021] A3 data whitening, eliminating data covariance through data whitening to achieve mutual independence and same variance between vectors;

[0022] A4 problem conversion, so as to convert the solution of the mixing matrix W to the solution of the conversion matrix B;

[0023] A5 solve the transformation matrix B, so as to realize the extraction of independent components;

[0024] A6 independent component decomposition, decomposing the demixing matrix W, the mixing matrix A and the independent component S;

[0025] A7 component separation, dividing the demixing matrix W and the conversion matrix B into main components and residual components;

[0026] A8 calculate the sample statistics ;

[0027] A9 calculate the overall statistical threshold of the corresponding confidence level by using kernel density estimation method .

[0028] The application adopts the above scheme, which has the beneficial effects of:

[0029] 1) Effectively process data characteristics and improve monitoring accuracy: by fusing the SDAE network model and the independent component analysis method, the nonlinear and interference characteristics of the mutual inductor monitoring data are effectively dealt with, and the accuracy of the voltage transformer operation state monitoring is greatly improved;

[0030] 2) Enhance model robustness and adapt to complex environment: in the SDAE network model training process, random noise is superimposed on the input data, and a sparse penalty term is introduced to make the model learn relatively sparse and simple digital features, which reduces the risk of overfitting, so that the model can better adapt to the complex and changeable operating environment, thereby enhancing the model's resistance to noise and interference;

[0031] 3) Realize model dynamic optimization, guarantee monitoring reliability: through adopting the way of incremental learning, based on the model obtained by offline training, using real-time data collected on line to fine-tune parameters, can adapt to the dynamic change of data in time, and continuously optimize model performance. BRIEF DESCRIPTION OF DRAWINGS

[0032] Figure 1 The method flowchart provided by the embodiment of the present application.

[0033] Figure 2 The flowchart of the independent component analysis method in the embodiment of the present application. DETAILED DESCRIPTION

[0034] In order to facilitate the understanding of the present application, the present application will be described more fully below with reference to the accompanying drawings. The preferred embodiments of the present application are shown in the drawings. However, the present application can be realized in many different forms and is not limited to the embodiments described herein. The purpose of providing these embodiments is to make the disclosure of the present application more thorough and comprehensive.

[0035] Referring to the accompanying Figure 1 As shown in the drawings, in the present embodiment, an independent component analysis-based voltage transformer online monitoring method is characterized in that it comprises the following steps:

[0036] S1, based on the voltage transformer group of the substation, the historical data, steady-state data and real-time data of the output signal are collected, and the data set is constructed through the historical data, so as to provide data support for the subsequent training of the SDAE network model and improve the accuracy of model prediction; Specifically, the historical data is the voltage transformer group output signal collected in the past and the related voltage transformer group output signal collected by the rest of the substation; The steady-state data is the data of running for one week (this kind of data is default steady-state), which is used to generate initial output weight and initial matrix; Real-time acquisition data is the real-time output signal of the voltage transformer group collected in time.

[0037] S2, an initial SDAE network model is created by using a sparse denoising autoencoder SDAE;

[0038] Specifically, the above-mentioned autoencoder (Autoencoder, AE) is a kind of neural network, including input layer, hidden layer and output layer, wherein the input layer is used for data input, the hidden layer represents learned feature expression, and the output layer represents the reconstruction value after decoding process;The purpose of the self-encoder as an unsupervised learning algorithm is to let the output value try to reconstruct the input value rather than the target value of the input x after training, and secondly, the self-encoder network is composed of two parts, one is an encoder represented by a function And a decoder for generating reconstruction In order to make the output reconstruct the input as much as possible, the autoencoder must extract the most important features representing the input, which usually requires the number of nodes in the hidden layer of the autoencoder to be less than the number of nodes in the input layer;

[0039] In the encoding process, the nonlinear activation function The input data x is mapped to the hidden layer h, and the mapping formula is:

[0040] ,

[0041] ,

[0042] Wherein is the weight matrix, is the bias vector;

[0043] In the decoding process, the activation function The hidden layer expression h is mapped to the input layer , and the mapping formula is:

[0044] ,

[0045] Wherein, Generally, and The function is consistent with the sigmoid activation function; is the weight matrix, which is preferably set to , which is also commonly called a binding weight; is the bias vector,

[0046] The sparse denoising autoencoder SDAE used in the embodiment combines the advantages of sparse autoencoder SAE and denoising autoencoder DAE. Firstly, a sparse penalty term is introduced on the basis of the autoencoder loss function, and relatively sparse and simple digital features are learned under the constraint of sparsity, so that the feature distribution of the input data can be effectively obtained. At the same time, random noise is superimposed into the input data to enhance the robustness of the process monitoring model and obtain strong generalization ability. The above method uses comprehensive statistical analysis information to improve the accuracy and robustness of the process monitoring model in view of the nonlinear and interference characteristics of the transformer monitoring data.

[0047] Further, relevant parameters are introduced to construct the loss function of the SDAE network model:

[0048] ,

[0049] The loss function of the autoencoder is the reconstruction error sum between the input data x and the reconstructed data , that is, to find optimal parameters of the SDAE network model;

[0050] wherein, L is the reconstruction error loss function, generally the sum of squares of errors (reflecting the performance of the model under the current model parameters through the loss value L); n is the number of samples; represents the i-th sample in the input data, represents the data obtained after mixing a certain proportion of noise signals to the input sample, represents the weight coefficient for controlling the sparsity penalty term, represents the number of hidden layer neurons, represents the sparsity parameter, represents the average activation of the j-th neuron under the sample data, wherein, The divergence describes the difference between two probability distributions, and its value monotonically increases with the distance between the two probability distributions and The specific expressions of and reach the minimum value;

[0051] wherein, and The specific expressions of

[0052] ,

[0053] ,

[0054] wherein, represents the activation value of the j-th neuron in the autoencoder.

[0055] S3 imports the data set into the initial SDAE network model created in step S2 for offline training, thereby obtaining the optimal SDAE network model. Through the initial SDAE network model created on the basis of historical data, relatively more data samples can be provided for related model establishment, thereby improving the accuracy of the model operation, reducing the required time for subsequent model correction, and relatively reducing the input cost.

[0056] S4 imports the steady-state data into the SDAE network model for correction training, and obtains the corrected SDAE network model when the training times of the SDAE network model reach the preset value or the iteration effect reaches the preset purpose. Specifically, the preset value of the first batch of training times of the SDAE network model or the preset purpose of the iteration effect can be set according to the actual situation, which is not limited here. Further correction of the SDAE network model by the steady-state data can reduce the operation error of the SDAE network model and improve its operation accuracy. Secondly, the steady-state data, i.e. the relatively constant output signal of the voltage transformer over a long period of time, is more suitable for the overall statistical threshold of the subsequent SDAE network model. The calculation is based on speculation.

[0057] S5 imports steady-state data into the SDAE network model for encoding, and then performs data standardization.

[0058] S6 takes the standardized coded data as input, performs independent component decomposition using the Independent Component Analysis (ICA) method, and then... (See appendix) Figure 2 As shown, calculate the sample statistic. ;

[0059] Specifically, the independent component analysis method described above includes the following steps for independent component decomposition:

[0060] A1 steady-state data matrix construction: A steady-state data matrix is ​​constructed by collecting steady-state data. Where n is the number of samples and m is the number of elements in a single sample collection; Based on this, the following data model is established, decomposing the original data into a mixture matrix A and an independent component matrix S (n×r, where r is the number of independent components), namely:

[0061] ;

[0062] Since the column vectors (i.e., the independent components) of the independent component matrix S satisfy statistical independence and possess zero mean and unit variance, the problem can be transformed into: transforming the sample matrix X through the unmixing matrix W, the resulting matrix S has statistical independence properties, i.e.:

[0063] ;

[0064] A2 data standardization involves standardizing the steady-state data matrix to obtain the matrix. This results in the mean of each column being 0 and the standard deviation being 1. Furthermore, it can replace X in the mathematical model in subsequent solutions, meaning the problem is transformed into:

[0065] ;

[0066] A3 data whitening eliminates data covariance, achieving mutual independence (uncorrelated vectors) and identical variance between vectors. The specific process of data whitening is as follows:

[0067] First, solve for the covariance matrix of the observation matrix. :

[0068] ;

[0069] For matrix Perform eigenvalue decomposition to obtain its eigenvector matrix. and the diagonal matrix formed by eigenvalues :

[0070] ;

[0071] Solve the whitening matrix :

[0072] ;

[0073] Solve for the matrix Z after data whitening:

[0074] ;

[0075] The A4 problem transformation involves matrix B, which is a transformation of the mixture matrix A. Compared to the full-rank mixture matrix A, matrix B is orthogonal, thus requiring fewer parameters to be estimated. The specific derivation is as follows:

[0076] ,

[0077] Based on the above derivation, it can be deduced that the independent component matrix S and the whitened matrix Z satisfy the following:

[0078] ,

[0079] Further, solving for the unmixing matrix W is transformed into solving for the transformation matrix B:

[0080] ,

[0081] A5 solves for the transformation matrix B. The independent component decomposition method separates independent source signals through the maximum independence criterion, and independence can be achieved by minimizing mutual information or maximizing non-Gaussianity.

[0082] Since the Gaussian distribution has the maximum entropy, non-Gaussianity can be used as an indicator of independence. By maximizing non-Gaussianity, independent components can be extracted. The independent component analysis method utilizes the above idea to solve matrix B through the following iterative process:

[0083] ① Determine the number r of independent components that need to be estimated, and denote i = 1;

[0084] ② Randomly select the initial unit vector (Column vector of B);

[0085] ③ Use the following formula to Calculate the amplitude, where E{} represents the expectation;

[0086] ,

[0087] in, express The derivative, It is a non-linear function, and can be selected from the following functions:

[0088] ,

[0089] ,

[0090] ,

[0091] ④ Regarding the income Perform orthogonalization;

[0092] ⑤ If the result is If convergence has not occurred, return to step ③;

[0093] ⑥If the result is Convergence, output vector ;

[0094] ⑦ If i < r, then let i = i+1 and return to step ②; if i = r, end the above process.

[0095] ⑧ Output all column vectors This forms matrix B.

[0096] A6 Independent Component Decomposition decomposes the unmixing matrix W, the mixing matrix A, and the independent component components S. The specific calculation method is as follows:

[0097] ,

[0098] ,

[0099] .

[0100] A7 component separation divides the unmixing matrix W and the transformation matrix B into principal components and residual components. Specifically, it calculates the L2 norm of the column vectors of the mixing matrix A. (i = 1, 2, ..., r, where r is the number of independent components);

[0101] Secondly, arrange the L2 norms in descending order and calculate the contribution of the features. and cumulative contribution ;

[0102] ,

[0103] ;

[0104] For example, the preferred standard is to divide all components into principal components (column vectors corresponding to features that contribute more than 85% of the cumulative contribution) and residual components.

[0105] The unmixing matrix W is divided into principal components according to the row vectors (the row indices mapped by the column indices of A). and remnants ;

[0106] Divide the transformation matrix B into principal components according to the column vectors (the column indices corresponding to the column indices of A). and remnants .

[0107] A8 calculates sample statistics. In calculating sample statistics In (sample squared prediction error), the i-th sample vector is calculated. The formula for the statistic is:

[0108] ,

[0109] in, This represents the reconstructed variables of the principal components isolated by independent component analysis at this moment:

[0110] ,

[0111] According to the formula Will The calculation is further simplified to:

[0112] ,

[0113] S7 calculates the sample statistic using the kernel density estimation method. Overall statistical threshold ;

[0114] A9 uses kernel density estimation to calculate the overall statistical threshold for the corresponding confidence level. The confidence level mentioned above refers to the degree to which a specific individual believes in the truth of a specific proposition; that is, probability is a measure of the rationality of an individual's belief.

[0115] S8 collects raw data in real time and calculates real-time statistics through steps S5-S6. ;

[0116] S9 will provide real-time statistics. Compared with the overall statistical threshold In comparison, to determine whether a voltage transformer bank has malfunctioned, the real-time statistics... Greater than the overall statistical threshold When the voltage transformer group fails, the voltage transformer bank will malfunction; conversely, when the real-time statistics show... Less than the overall statistical threshold If so, then the voltage transformer group has not failed;

[0117] S10 repeats steps S1-S9 to perform real-time evaluation of the voltage transformer group.

[0118] Through continuous training and real-time evaluation and comparison using the above steps, the SDAE network model can be trained with imported data, improving the robustness of the process monitoring model's accuracy. In practical applications, it is also necessary to combine various strategies such as processing superimposed random noise in updated data, model architecture selection, and relevant parameter tuning to further improve the model's performance.

[0119] In summary, the advantages mentioned above include:

[0120] 1) Effectively handle data characteristics and improve monitoring accuracy: By integrating the SDAE network model and independent component analysis method, the nonlinear and interference characteristics of the transformer monitoring data are effectively addressed, which greatly improves the accuracy of monitoring the operating status of voltage transformers.

[0121] 2) Enhance model robustness and adapt to complex environments: During the training of the SDAE network model, random noise is superimposed on the input data, and a sparse penalty term is introduced to enable the model to learn relatively sparse and concise numerical features. This reduces the risk of overfitting and enables the model to better adapt to complex and ever-changing operating environments, thereby enhancing the model's resistance to noise and interference.

[0122] 3) Achieve dynamic model optimization and ensure monitoring reliability: By adopting incremental learning, the model is trained offline and then fine-tuned using real-time data collected online. This allows the model to adapt to dynamic changes in the data in a timely manner and continuously optimize its own performance.

[0123] This embodiment takes advantage of the relatively stable three-phase voltage imbalance of the instrument transformer. It analyzes the corresponding data based on the Sparse Denoising Autoencoder-Independent Component Analysis (SDAE-ICA) method to simultaneously address the nonlinear problems that may be caused by three-phase imbalance or harmonic dynamic disturbances. The SDAE method is an improvement on the traditional encoder method and can be effectively used for data feature extraction and dimensionality reduction of nonlinear processes, while the Independent Component Analysis method (ICA method) can extract non-Gaussian components in the data and enhance the separation of fault features.

[0124] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Any modifications or variations made by those skilled in the art, without departing from the scope of the present invention, using the disclosed technical content, are equivalent embodiments of the present invention. Therefore, all equivalent changes made based on the concept of the present invention without departing from the scope of the present invention should be covered within the protection scope of the present invention.

Claims

1. A method for online monitoring of voltage transformers based on independent component analysis, characterized in that: Includes the following steps: S1 collects historical data, steady-state data, and real-time data based on the output signals of the voltage transformer group in the substation, and constructs a dataset using the historical data. S2 uses a sparse denoising autoencoder to create the initial SDAE network model; S3 imports the dataset into the initial SDAE network model created in step S2 for offline training, thereby obtaining the optimal SDAE network model. S4 imports the steady-state data into the optimal SDAE network model for correction training. When the optimal SDAE network model reaches a preset number of training times or the iteration effect reaches a preset objective, the corrected SDAE network model is obtained. . S5 imports the steady-state data into the modified SDAE network model for encoding, and then performs data standardization processing. S6 takes the standardized coded data as input, performs independent component decomposition using the independent component analysis method, and then calculates the sample statistic. ; S7 calculates the sample statistic using the kernel density estimation method. Overall statistical threshold ; S8 collects raw data in real time and calculates real-time statistics through steps S5-S6. ; S9 will provide real-time statistics. Compared with the overall statistical threshold In comparison, to determine whether a voltage transformer bank has malfunctioned, the real-time statistics... Greater than the overall statistical threshold If this happens, the voltage transformer group will malfunction; Conversely, when real-time statistics Less than the overall statistical threshold If so, then the voltage transformer group has not failed; S10 repeats steps S1-S9 to perform real-time evaluation of the voltage transformer group.

2. The online monitoring method for voltage transformers based on independent component analysis according to claim 1, characterized in that: Introducing data to represent the input sample after it has been mixed with a certain proportion of noise signal. The weighting coefficients used to represent the control of the sparsity penalty term. , representing the number of neurons in the hidden layer , used to represent sparsity parameters and the average activation level of the j-th neuron in the sample data. This allows us to construct the loss function for the SDAE network model.

3. The online monitoring method for voltage transformers based on independent component analysis according to claim 1, characterized in that: The independent component analysis method (ICA) for independent component decomposition includes the following steps: A1 steady-state data matrix construction: decompose the steady-state data into a mixture matrix A and independent component matrices S; A2 data standardization involves standardizing the steady-state data matrix to obtain the matrix. This results in the mean of each column being 0 and the standard deviation being 1. A3 data whitening eliminates data covariance, achieving mutual independence and identical variance between vectors; A4 solves the problem transformation, thus transforming the solution to the unmixing matrix W into the solution to the transformation matrix B; A5 solves for the transformation matrix B, thereby enabling the extraction of independent components; A6 Independent Component Decomposition decomposes the unmixing matrix W, the mixing matrix A, and the independent component components S. A7 component separation divides the unmixing matrix W and the transformation matrix B into principal components and residual components; A8 calculates sample statistics. ; A9 uses kernel density estimation to calculate the overall statistical threshold for the corresponding confidence level. .

Images