A novel method for wide-area monitoring of power system broadband oscillations

Abstract

The application discloses a novel wide-area monitoring method for wide frequency oscillation of a power system, which comprises the following steps: collecting electric signal waveforms of key nodes of the power system at each substation; performing modal decomposition and feature extraction on the electric signal waveforms at each substation, identifying the type of wide frequency oscillation, obtaining a preliminary prediction result, and saving the corresponding wide frequency oscillation electric signal, namely, a to-be-compressed wide frequency oscillation electric signal; performing secondary subsampling on the to-be-compressed wide frequency oscillation electric signal at each substation to obtain a compressed signal, and uploading the compressed signal and the preliminary prediction result to a main station; performing feature extraction on the compressed signal by using a wide frequency oscillation classification and identification model of the main station, and obtaining the final type of wide frequency oscillation by comparing the preliminary prediction result; constructing a wide frequency oscillation signal reconstruction model of the main station, inputting the compressed signal uploaded by each substation, and outputting a reconstructed wide frequency oscillation signal; and comprehensively analyzing the reconstructed wide frequency oscillation signal by the main station, so that the novel wide-area global monitoring of the wide frequency oscillation of the power system is effectively realized.

Description

Technical Field

[0001] This invention relates to the technical field of wide-area monitoring of power systems, and in particular to a novel wide-area monitoring method for broadband oscillations in power systems. Background Technology

[0002] In recent years, the grid connection ratio of renewable energy and power electronic equipment has increased significantly, and the "dual-high" characteristics of the new power system dominated by new energy will become even more pronounced in the future. Compared with the traditional power system, the "source-grid-load" of the "dual-high" power system is developing towards power electronics. The interaction between a large number of renewable energy units and existing power generation equipment, transmission networks, power loads, and themselves has led to increasingly frequent broadband oscillations and the resulting engineering safety accidents. These new power system oscillations exhibit characteristics such as wide bandwidth, high noise, multi-mode operation, strong time-varying nature, strong nonlinearity, and spatiotemporal distribution. Moreover, their causes are complex and diverse, seriously threatening the stability of the power system and the safe and economical operation of electrical equipment, and severely restricting the effective absorption of renewable energy sources such as wind power and photovoltaics. Therefore, effective monitoring of broadband oscillations is of great significance for providing a data foundation for mechanism analysis, source tracing, online monitoring and early warning, and prevention and control of broadband oscillations.

[0003] Existing broadband oscillation monitoring methods still have many limitations in practical applications. First, synchronous phasor measurement units (PMUs), based on the Nyquist sampling theorem, can only monitor some sub- / super-synchronous oscillations in the 10-40Hz and 60-90Hz ranges. For sub- / super-synchronous oscillations with frequencies higher than low-frequency oscillations but within twice the power frequency range, accurate parameters of broadband oscillations cannot be directly obtained from the phasor data acquired by the PMU. Second, due to limitations in communication bandwidth and the Nyquist sampling theorem, the master station can only obtain oscillation signals within half the transmission frequency range from the slave station, and can only effectively monitor and analyze some sub- / super-synchronous oscillation signals within 50Hz of the power frequency. It cannot obtain complete oscillation information of broadband oscillations, making it difficult to effectively monitor the entire broadband oscillation. Furthermore, at the master station dispatch end, since different types of broadband oscillations may occur simultaneously, the lack of multi-source oscillation wide-area monitoring and identification methods makes it impossible to accurately trace the oscillation source, analyze the propagation path and its impact range, and provide support for the online monitoring, prevention, control and protection of the power grid, thus affecting the safe and stable operation of the power grid. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings and deficiencies of the prior art and provide a novel wide-area monitoring method for broadband oscillations in power systems, which can effectively realize global monitoring of broadband oscillations in new power systems.

[0005] To achieve the above objectives, the technical solution provided by this invention is: a novel wide-area monitoring method for broadband oscillations in power systems, comprising the following steps:

[0006] S1. At each substation, the synchronous measurement unit (PMU) of the wide area measurement system is used to collect electrical signal waveforms of key nodes in the power system, and the collected electrical signal waveforms are stored as time series.

[0007] S2. Each substation performs mode decomposition and feature extraction on the electrical signal waveform, extracts the oscillation characteristics of the electrical signal waveform, identifies the broadband oscillation type, obtains preliminary prediction results, and saves the corresponding broadband oscillation electrical signal, i.e. the broadband oscillation electrical signal to be compressed.

[0008] S3. Each substation uses compressed sensing theory to perform secondary subsampling on the broadband oscillating electrical signal to be compressed, obtains the compressed signal, and uploads the compressed signal and preliminary prediction results to the main station through the communication network.

[0009] S4. At the main station, a broadband oscillation classification and recognition model based on deep dictionary learning is used to extract features from the uploaded compressed signal. By comparing the results with the preliminary prediction results uploaded by the sub-station, the final broadband oscillation type is obtained.

[0010] S5. At the main station, a broadband oscillation signal reconstruction model is constructed using compressed sensing. The input is the compressed signal uploaded by each substation, and the output is the reconstructed broadband oscillation signal. The main station performs comprehensive analysis on the reconstructed broadband oscillation signal to achieve global monitoring of broadband oscillation in the new power system.

[0011] Furthermore, in step S1, the electrical signal waveform refers to the electrical signal waveforms of each time period obtained by real-time sampling through the synchronous measurement unit (PMU) of the wide-area measurement system. The electrical signal includes line current, node voltage, active power, reactive power, and rotor angle.

[0012] Furthermore, step S2 includes the following steps:

[0013] S21. The electrical signal waveform is decomposed using an adaptive mode decomposition method to obtain the main mode components of the electrical signal waveform. Then, the main mode components are denoised using correlation analysis, kurtosis criterion and soft thresholding to obtain the denoised main mode components.

[0014] S22. Use Hilbert transform to extract features from the main modal components after noise reduction to obtain oscillation feature quantities, including oscillation amplitude and oscillation frequency.

[0015] S23. Based on the preset wideband oscillation frequency threshold and amplitude threshold, determine whether the obtained oscillation amplitude and oscillation frequency are greater than the preset wideband oscillation frequency threshold and amplitude threshold. If both the oscillation amplitude and oscillation frequency are greater than or equal to the preset wideband oscillation frequency threshold and amplitude threshold, then it is identified as wideband oscillation, a preliminary judgment result is obtained, and the corresponding wideband oscillation electrical signal, i.e., the wideband oscillation electrical signal to be compressed, is saved.

[0016] Furthermore, step S3 includes the following steps:

[0017] S31. At the substation, based on the broadband oscillating electrical signal to be compressed, a substation classification sparse basis, i.e., the substation classification sparse transform domain, is designed using the sparsity-adaptive K-SVD dictionary learning method, denoted as ψ=[ψ1,ψ2,…,ψ]. m ,…ψ K Let ψ be the set of sparse bases corresponding to various broadband oscillations. m Let be the classification sparse basis corresponding to the m-th type of wideband oscillation, where m = 1, 2, ..., K; where the sparsity-adaptive K-SVD dictionary learning method enables the sparsity, i.e. the number of non-zero coefficients in the dictionary representation of each signal, to be adaptively adjusted according to the characteristics of the signal.

[0018] S32. Construct a random measurement matrix Φ that is unrelated to the sparse basis of substation classification using the Logistic chaotic mapping method shown in Equation (1): First, determine the matrix dimension as M×N, and select appropriate initial values ​​and bifurcation parameters. Then, generate a chaotic sequence of length M×N using the Logistic chaotic mapping method shown in Equation (1). Fill the M×N matrix with the chaotic sequence in sequence to obtain the original random measurement matrix Φ'. Then, normalize the original random measurement matrix Φ' to obtain the random measurement matrix Φ that is unrelated to the sparse basis of substation classification. The random measurement matrix Φ has good numerical stability and can effectively project high-dimensional signals into low-dimensional space in compressed sensing.

[0019] h n+1 =μh n (1-h n (1)

[0020] In the formula, h n+1 h n , respectively, are the values ​​of the chaotic variables in the (n+1)th and nth iterations, and μ is the bifurcation parameter;

[0021] S33, Based on a substation classification sparse base Ψ m The random measurement matrix Φ and the compressed sensing theory shown in equation (2) are used to compress the broadband oscillating electrical signal to be compressed, i.e., secondary subsampling, to obtain the compressed signal y, i.e., the one-dimensional observation signal;

[0022] y = Φx = ΦΨ m s=Θs (2)

[0023] In the formula, x is the original signal of length N, i.e., the broadband oscillating electrical signal to be compressed; M represents the number of rows in the observation matrix, i.e., the dimension of the compressed signal y, satisfying N >> M; S is the classification sparse basis Ψ of the original signal x at a certain substation. m The coefficient vector is sparse, with sparsity G, satisfying G << N; Θ = ΦΨ, which is an M×N dimensional perception matrix.

[0024] Further, in step S4, at the main station, a broadband oscillation classification and recognition model is constructed using deep dictionary learning, and this model is used to extract oscillation features from the uploaded compressed signal to identify the broadband oscillation type. First, the main station normalizes the received compressed one-dimensional observation signals of different broadband oscillation types from each substation, and sets classification labels based on the preliminary prediction results of the broadband oscillation type, forming a one-dimensional observation signal sample dataset. Then, the Gram corner field method is used to convert the one-dimensional observation signals in the one-dimensional observation signal sample dataset into two-dimensional Gram feature maps, and a corresponding classification label is set for each two-dimensional Gram feature map. The two-dimensional Gram feature maps and the corresponding classification labels are used to construct a two-dimensional image sample dataset, which is divided into a training set and a test set.

[0025] Based on deep learning theory and dictionary learning theory, a deep dictionary learning method is proposed to construct a broadband oscillation classification and recognition model. The model structure includes:

[0026] Input module: The input consists of two-dimensional Gram feature maps of each category, providing basic data for subsequent deep dictionary learning;

[0027] The classification deep learning dictionary module contains multiple categories of deep learning dictionary modules. The input is the two-dimensional gram feature map of each category in the input module, and the output is the updated deep learning dictionary for each category. Each category's deep learning dictionary module consists of multiple cascaded deep dictionary learning modules. Each deep dictionary learning module includes a sparse coding layer and a dictionary update layer. It progressively learns and extracts local and global features from the two-dimensional gram feature map through multiple dictionary layers, improving feature extraction capabilities. The input of each deep dictionary learning module is the two-dimensional gram feature map, and the output is the updated deep learning dictionary for each category. Specifically, the input of the sparse coding layer is the output data of the previous layer (i.e., the input module or the previous deep dictionary learning module) and the dictionary matrix of this layer. It optimizes and solves the output sparse representation coefficients to extract key features of the input signal. The input of the dictionary update layer is the sparse representation coefficients output by the sparse coding layer. It uses a dictionary learning algorithm based on nonlinear transformation to update the dictionary matrix of this layer, obtaining the updated dictionary for this layer, thereby improving nonlinear expressive power and suppressing noise and insignificant features in the signal.

[0028] The sparse representation coefficient calculation module takes as input the original data of each gram 2D ​​feature map and the updated deep learning dictionary for each category. It calculates the sparse representation coefficients of each gram 2D ​​feature map under each category deep learning dictionary using the sparse representation method, and obtains the reconstructed gram 2D ​​feature map under each category deep learning dictionary using the signal reconstruction algorithm. It then compares the reconstructed gram 2D ​​feature map with the original data and finally outputs the reconstruction residual of each gram 2D ​​feature map under each category deep learning dictionary. The smaller the error, the higher the matching degree with the category.

[0029] Classification Decision Module: The input is the reconstruction residual of each Gram 2D ​​feature map under each category deep learning dictionary. The class with the smallest reconstruction residual is selected as the sample class, and the output is the class label of each Gram 2D ​​feature map.

[0030] Finally, the broadband oscillation classification and recognition model was tested using a test set divided from the two-dimensional image sample dataset. The model parameters were further optimized and adjusted to obtain the broadband oscillation classification and recognition model with the best performance.

[0031] Furthermore, step S5 includes the following steps:

[0032] S51. Based on the one-dimensional observation signal sample dataset, the main station uses the sparsity-adaptive K-SVD dictionary learning method to design the main station classification sparse basis, that is, the main station classification sparse transform domain.

[0033] S52. The main station uses the random measurement matrix generation model constructed based on the Logistic chaotic mapping method shown in Equation (1) to reconstruct the random measurement matrix, and multiplies it by the main station classification sparse basis to obtain the perception matrix.

[0034] S53. Based on compressed sensing theory and sensing matrix, a broadband oscillation signal reconstruction model is constructed. The model structure is as follows:

[0035] Input layer: The input is a one-dimensional observation signal;

[0036] Signal reconstruction layer: The one-dimensional observed signal is reconstructed using a signal reconstruction algorithm based on sparse representation. The reconstructed sparse signal is obtained by using deep learning dictionaries of various categories and sparse representation coefficients.

[0037] Signal recovery layer: Recovers the reconstructed sparse signal and outputs the recovered wideband oscillation signal, i.e., the reconstructed wideband oscillation signal;

[0038] Finally, the broadband oscillation signal reconstruction model was trained and tested using the training and test sets, and the model parameters were optimized to obtain the broadband oscillation signal reconstruction model with the best performance.

[0039] S54. Utilize the best-performing broadband oscillation signal reconstruction model, with the input being the compressed signal uploaded by each substation and the output being the reconstructed broadband oscillation signal.

[0040] S55. The main station performs comprehensive analysis based on the reconstructed broadband oscillation signal, including time-frequency analysis, wide-area monitoring, mechanism analysis, disturbance source tracing and location, data mining, scheduling decision-making, and prevention, protection and control.

[0041] Furthermore, the signal reconstruction algorithms based on sparse representation include matching pursuit algorithm, orthogonal matching pursuit algorithm, regularized matching pursuit algorithm, basis pursuit algorithm, and alternating direction multiplier method.

[0042] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0043] This invention utilizes an adaptive mode decomposition method at the substation to perform mode decomposition on broadband oscillating electrical signal waveforms, obtaining the main mode components of the waveform. Then, it combines correlation analysis, kurtosis criteria, and soft thresholding to denoise these main mode components, yielding denoised main mode components. Hilbert transform is then used to extract features from the denoised main mode components, obtaining oscillation characteristic quantities. Combined with preset thresholds, broadband oscillations are identified, providing preliminary prediction results. By employing the adaptive mode decomposition method and Hilbert transform, the oscillation characteristics and types of various electrical signals can be quickly obtained, leading to rapid preliminary prediction results, improving the signal quality transmitted to the main station, and reducing the computational and storage requirements of the substation's industrial control computer.

[0044] This invention utilizes a sparse basis with atomic self-updating and self-optimizing functions and a random measurement matrix at the substation to perform dimensionality reduction processing on the detected broadband oscillation data, which greatly improves the signal compression ratio, reduces the storage and transmission burden caused by the fixed sparse basis, and improves data compression and transmission efficiency. It is beneficial to realize the transmission of high-frequency broadband oscillation data under existing communication bandwidth and PMU data transmission frequency, and ensures the real-time performance and validity of the data.

[0045] This invention proposes a broadband oscillation identification and reconstruction method based on Gram angle field, compressed sensing, and deep dictionary learning. The master station can directly utilize compressed data to quickly and accurately extract and identify the multi-mode oscillation features and types of broadband oscillations. While ensuring the ability to distinguish multi-mode features of multiple oscillation types, the extracted fault features have good hierarchy and physical meaning. Simultaneously, based on the compressed data uploaded by each substation, the original broadband oscillation signal can be accurately reconstructed using signal reconstruction algorithms, providing support for subsequent oscillation source tracing, suppression, and power grid stability control. Attached Figure Description

[0046] Figure 1 This is an overall flowchart of the method of the present invention. Detailed Implementation

[0047] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0048] like Figure 1 As shown in the figure, this embodiment discloses a novel wide-area monitoring method for broadband oscillations in power systems, the specific details of which are as follows:

[0049] S1. At each substation, the synchronous measurement unit (PMU) of the wide-area measurement system is used to collect electrical signal waveforms of key nodes in the power system, and the collected electrical signal waveforms are stored as time series; wherein, the electrical signal waveforms refer to the electrical signal waveforms of each time period obtained by real-time sampling through the synchronous measurement unit (PMU) of the wide-area measurement system, and the electrical signals include line current, node voltage, active power, reactive power and rotor angle.

[0050] S2. Each substation performs mode decomposition and feature extraction on the electrical signal waveform, extracts the oscillation characteristics of the electrical signal waveform, identifies the broadband oscillation type, obtains preliminary prediction results, and saves the corresponding broadband oscillation electrical signal, i.e., the broadband oscillation electrical signal to be compressed. This includes the following steps:

[0051] S21. The electrical signal waveform is decomposed using an adaptive mode decomposition method to obtain the main mode components of the electrical signal waveform. Then, the main mode components are denoised using correlation analysis, kurtosis criterion and soft thresholding to obtain the denoised main mode components.

[0052] S22. Use Hilbert transform to extract features from the main modal components after noise reduction to obtain oscillation feature quantities, including oscillation amplitude and oscillation frequency.

[0053] S23. Based on the preset wideband oscillation frequency threshold and amplitude threshold, determine whether the obtained oscillation amplitude and oscillation frequency are greater than the preset wideband oscillation frequency threshold and amplitude threshold. If both the oscillation amplitude and oscillation frequency are greater than or equal to the preset wideband oscillation frequency threshold and amplitude threshold, then it is identified as wideband oscillation, a preliminary judgment result is obtained, and the corresponding wideband oscillation electrical signal, i.e., the wideband oscillation electrical signal to be compressed, is saved.

[0054] S3. Each substation uses compressed sensing theory to perform secondary subsampling on the broadband oscillating electrical signal to be compressed, obtaining a compressed signal. The compressed signal and the preliminary prediction results are then uploaded to the main station via the communication network, including the following steps:

[0055] S31. At the substation, based on the broadband oscillating electrical signal to be compressed, a substation classification sparse basis, i.e., the substation classification sparse transform domain, is designed using the sparsity-adaptive K-SVD dictionary learning method, denoted as ψ=[ψ1,ψ2,…,ψ].m ,…ψ K Let ψ be the set of sparse bases corresponding to various broadband oscillations. m Let be the classification sparse basis corresponding to the m-th type of wideband oscillation, where m = 1, 2, ..., K; where the sparsity-adaptive K-SVD dictionary learning method enables the sparsity, i.e. the number of non-zero coefficients in the dictionary representation of each signal, to be adaptively adjusted according to the characteristics of the signal.

[0056] S32. Construct a random measurement matrix Φ that is unrelated to the sparse basis of substation classification using the Logistic chaotic mapping method shown in Equation (1): First, determine the matrix dimension as M×N, and select appropriate initial values ​​and bifurcation parameters. Then, generate a chaotic sequence of length M×N using the Logistic chaotic mapping method shown in Equation (1). Fill the M×N matrix with the chaotic sequence in sequence to obtain the original random measurement matrix Φ'. Then, normalize the original random measurement matrix Φ' to obtain the random measurement matrix Φ that is unrelated to the sparse basis of substation classification. The random measurement matrix Φ has good numerical stability and can effectively project high-dimensional signals into low-dimensional space in compressed sensing.

[0057] h n+1 =μh n (1-h n (1)

[0058] In the formula, h n+1 h n , respectively, are the values ​​of the chaotic variables in the (n+1)th and nth iterations, and μ is the bifurcation parameter;

[0059] S33, Based on a substation classification sparse base Ψ m The random measurement matrix Φ and the compressed sensing theory shown in equation (2) are used to compress the broadband oscillating electrical signal to be compressed, i.e., secondary subsampling, to obtain the compressed signal y, i.e., the one-dimensional observation signal;

[0060] y = Φx = ΦΨ m s=Θs(2)

[0061] In the formula, x is the original signal of length N, i.e., the broadband oscillating electrical signal to be compressed; Φ is an M×N dimensional random measurement matrix, where M represents the number of rows in the observation matrix, i.e., the dimension of the compressed signal y, satisfying N >> M; S is the classification sparse basis Ψ of the original signal x at a certain substation. m The coefficient vector is sparse, with sparsity G, satisfying G << N; Θ = ΦΨ, which is an M×N dimensional perception matrix.

[0062] S4. On the main station, a broadband oscillation classification and recognition model based on deep dictionary learning is used to extract oscillation features from the uploaded compressed signal, identify the broadband oscillation type, and compare it with the preliminary prediction results uploaded by the sub-station to obtain the final broadband oscillation type, as follows:

[0063] First, the main station normalizes the received compressed one-dimensional observation signals of different broadband oscillation types from each substation and sets classification labels based on the preliminary prediction results of the broadband oscillation types, forming a one-dimensional observation signal sample dataset. Then, the Gram corner field method is used to convert the one-dimensional observation signals in the one-dimensional observation signal sample dataset into two-dimensional Gram feature maps, and a corresponding classification label is set for each two-dimensional Gram feature map. The two-dimensional Gram feature maps and the corresponding classification labels are used to construct a two-dimensional image sample dataset, which is divided into a training set and a test set.

[0064] Based on deep learning theory and dictionary learning theory, a deep dictionary learning method is proposed to construct a broadband oscillation classification and recognition model. The model structure includes:

[0065] Input module: The input consists of two-dimensional Gram feature maps of each category, providing basic data for subsequent deep dictionary learning;

[0066] The classification deep learning dictionary module contains multiple categories of deep learning dictionary modules. The input is the two-dimensional gram feature map of each category in the input module, and the output is the updated deep learning dictionary for each category. Each category's deep learning dictionary module consists of multiple cascaded deep dictionary learning modules. Each deep dictionary learning module includes a sparse coding layer and a dictionary update layer. It progressively learns and extracts local and global features from the two-dimensional gram feature map through multiple dictionary layers, improving feature extraction capabilities. The input of each deep dictionary learning module is the two-dimensional gram feature map, and the output is the updated deep learning dictionary for each category. Specifically, the input of the sparse coding layer is the output data of the previous layer (i.e., the input module or the previous deep dictionary learning module) and the dictionary matrix of this layer. It optimizes and solves the output sparse representation coefficients to extract key features of the input signal. The input of the dictionary update layer is the sparse representation coefficients output by the sparse coding layer. It uses a dictionary learning algorithm based on nonlinear transformation to update the dictionary matrix of this layer, obtaining the updated dictionary for this layer, thereby improving nonlinear expressive power and suppressing noise and insignificant features in the signal.

[0067] The sparse representation coefficient calculation module takes as input the original data of each gram 2D ​​feature map and the updated deep learning dictionary for each category. It calculates the sparse representation coefficients of each gram 2D ​​feature map under each category deep learning dictionary using the sparse representation method, and obtains the reconstructed gram 2D ​​feature map under each category deep learning dictionary using the signal reconstruction algorithm. It then compares the reconstructed gram 2D ​​feature map with the original data and finally outputs the reconstruction residual of each gram 2D ​​feature map under each category deep learning dictionary. The smaller the error, the higher the matching degree with the category.

[0068] Classification Decision Module: The input is the reconstruction residual of each Gram 2D ​​feature map under each category deep learning dictionary. The class with the smallest reconstruction residual is selected as the sample class, and the output is the class label of each Gram 2D ​​feature map.

[0069] Finally, the broadband oscillation classification and recognition model was tested using a test set divided from the two-dimensional image sample dataset. The model parameters were further optimized and adjusted to obtain the broadband oscillation classification and recognition model with the best performance.

[0070] S5. At the main station, a broadband oscillation signal reconstruction model is constructed using compressed sensing. The input is the compressed signal uploaded by each substation, and the output is the reconstructed broadband oscillation signal. The main station performs comprehensive analysis on the reconstructed broadband oscillation signal to achieve wide-area global monitoring of broadband oscillations in the new power system. This includes the following steps:

[0071] S51. Based on the one-dimensional observation signal sample dataset, the main station uses the sparsity-adaptive K-SVD dictionary learning method to design the main station classification sparse basis, that is, the main station classification sparse transform domain.

[0072] S52. The main station uses the random measurement matrix generation model constructed based on the Logistic chaotic mapping method shown in Equation (1) to reconstruct the random measurement matrix, and multiplies it by the main station classification sparse basis to obtain the perception matrix.

[0073] S53. Based on compressed sensing theory and sensing matrix, a broadband oscillation signal reconstruction model is constructed. The model structure is as follows:

[0074] Input layer: The input is a one-dimensional observation signal;

[0075] Signal reconstruction layer: The one-dimensional observed signal is reconstructed using a sparse representation-based signal reconstruction algorithm. The reconstructed sparse signal is obtained using deep learning dictionaries of various categories and sparse representation coefficients. The sparse representation-based signal reconstruction algorithm includes matching pursuit algorithm, orthogonal matching pursuit algorithm, regularized matching pursuit algorithm, basis pursuit algorithm, and alternating direction multiplier method, etc.

[0076] Signal recovery layer: Recovers the reconstructed sparse signal and outputs the recovered wideband oscillation signal, i.e., the reconstructed wideband oscillation signal;

[0077] Finally, the broadband oscillation signal reconstruction model was trained and tested using the training and test sets, and the model parameters were optimized to obtain the broadband oscillation signal reconstruction model with the best performance.

[0078] S54. Utilize the best-performing broadband oscillation signal reconstruction model, with the input being the compressed signal uploaded by each substation and the output being the reconstructed broadband oscillation signal.

[0079] S55. The main station performs comprehensive analysis based on the reconstructed broadband oscillation signal, including time-frequency analysis, wide-area monitoring, mechanism analysis, disturbance source tracing and location, data mining, scheduling decision-making, and prevention, protection and control.

[0080] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A novel wide-area monitoring method for broadband oscillations in power systems, characterized in that, Includes the following steps: S1. At each substation, the synchronous measurement unit (PMU) of the wide area measurement system is used to collect electrical signal waveforms of key nodes in the power system, and the collected electrical signal waveforms are stored as time series. S2. Each substation performs mode decomposition and feature extraction on the electrical signal waveform, extracts the oscillation characteristics of the electrical signal waveform, identifies the broadband oscillation type, obtains preliminary prediction results, and saves the corresponding broadband oscillation electrical signal, i.e., the broadband oscillation electrical signal to be compressed. This includes the following steps: S21. The electrical signal waveform is decomposed using an adaptive mode decomposition method to obtain the main mode components of the electrical signal waveform. Then, the main mode components are denoised using correlation analysis, kurtosis criterion and soft thresholding to obtain the denoised main mode components. S22. Use Hilbert transform to extract features from the main modal components after noise reduction to obtain oscillation feature quantities, including oscillation amplitude and oscillation frequency. S23. Based on the preset wideband oscillation frequency threshold and amplitude threshold, determine whether the obtained oscillation amplitude and oscillation frequency are greater than the preset wideband oscillation frequency threshold and amplitude threshold. If both the oscillation amplitude and oscillation frequency are greater than or equal to the preset wideband oscillation frequency threshold and amplitude threshold, then it is identified as wideband oscillation, a preliminary judgment result is obtained, and the corresponding wideband oscillation electrical signal, i.e. the wideband oscillation electrical signal to be compressed, is saved. S3. Each substation uses compressed sensing theory to perform secondary subsampling on the broadband oscillating electrical signal to be compressed, obtaining a compressed signal. The compressed signal and the preliminary prediction results are then uploaded to the main station via the communication network, including the following steps: S31. At the substation, based on the broadband oscillating electrical signal to be compressed, a substation classification sparse basis, i.e., the substation classification sparse transform domain, is designed using the sparsity-adaptive K-SVD dictionary learning method, denoted as: , This represents the set of sparse bases corresponding to various broadband oscillations. For the first Classification of sparse bases corresponding to broadband oscillations Among them, the sparsity-adaptive K-SVD dictionary learning method enables the sparsity, i.e. the number of non-zero coefficients in the representation of each signal in the dictionary, to be adaptively adjusted according to the characteristics of the signal. S32. Construct a random measurement matrix that is uncorrelated with the sparse basis of substation classification using the Logistic chaotic mapping method shown in equation (1). First, determine the matrix dimension as follows: And select appropriate initial values ​​and bifurcation parameters, and then generate a logistic chaotic mapping with a length of using the Logistic chaotic mapping method shown in equation (1). The chaotic sequence is then filled sequentially into... From the matrix, the original random measurement matrix is ​​obtained. Then, for the original random measurement matrix Normalization is performed to obtain a random measurement matrix that is uncorrelated with the sparse basis of substation classification. The random measurement matrix It exhibits good numerical stability and can effectively project high-dimensional signals into low-dimensional space in compressed sensing; （1）； In the formula, , The first , The chaotic variable values ​​of the next iteration. For bifurcation parameters; S33, Based on a substation classification sparse base Random measurement matrix Based on the compressed sensing theory shown in equation (2), the broadband oscillating electrical signal to be compressed is compressed, i.e., subjected to secondary subsampling, to obtain the compressed signal. That is, a one-dimensional observation signal; （2）； In the formula, For length is The original signal, i.e., the broadband oscillating electrical signal to be compressed, This indicates the row number of the observation matrix, i.e., the compressed signal. The dimension that satisfies ; Original signal In a certain subsite, sparse base classification The coefficient vector is sparse, with a sparsity of . ,satisfy ; ,for 3D perception matrix; S4. At the main station, a broadband oscillation classification and recognition model based on deep dictionary learning is used to extract features from the uploaded compressed signal. By comparing the results with the preliminary prediction results uploaded by the sub-station, the final broadband oscillation type is obtained. At the main station, a broadband oscillation classification and recognition model is constructed using deep dictionary learning. This model is then used to extract oscillation features from the uploaded compressed signals to identify the broadband oscillation types. First, the main station normalizes the received compressed one-dimensional observation signals of different broadband oscillation types from each substation and sets classification labels based on the preliminary prediction results of the broadband oscillation types, forming a one-dimensional observation signal sample dataset. Then, the Gram corner field method is used to convert the one-dimensional observation signals in the one-dimensional observation signal sample dataset into two-dimensional Gram feature maps, and a corresponding classification label is assigned to each two-dimensional Gram feature map. The two-dimensional Gram feature maps and corresponding classification labels are used to construct a two-dimensional image sample dataset, which is then divided into a training set and a test set. S5. At the main station, a broadband oscillation signal reconstruction model is constructed using compressed sensing. The input is the compressed signal uploaded by each substation, and the output is the reconstructed broadband oscillation signal. The main station performs comprehensive analysis on the reconstructed broadband oscillation signal to achieve global monitoring of broadband oscillation in the new power system.

2. The novel wide-area monitoring method for broadband oscillations in power systems according to claim 1, characterized in that, In step S1, the electrical signal waveform refers to the electrical signal waveforms of each time period obtained by real-time sampling by the synchronous measurement unit (PMU) of the wide-area measurement system. The electrical signal includes line current, node voltage, active power, reactive power, and rotor angle.

3. A novel wide-area monitoring method for broadband oscillations in power systems according to claim 2, characterized in that, In step S4, a deep dictionary learning method is proposed based on deep learning theory and dictionary learning theory to construct a broadband oscillation classification and recognition model. The model structure includes: Input module: The input consists of two-dimensional Gram feature maps of each category, providing basic data for subsequent deep dictionary learning; The classification deep learning dictionary module contains multiple categories of deep learning dictionary modules. The input is the two-dimensional gram feature map of each category in the input module, and the output is the updated deep learning dictionary for each category. Each category's deep learning dictionary module consists of multiple cascaded deep dictionary learning modules. Each deep dictionary learning module includes a sparse coding layer and a dictionary update layer. It progressively learns and extracts local and global features from the two-dimensional gram feature map through multiple dictionary layers, improving feature extraction capabilities. The input of each deep dictionary learning module is the two-dimensional gram feature map, and the output is the updated deep learning dictionary for each category. Specifically, the input of the sparse coding layer is the output data of the previous layer (i.e., the input module or the previous deep dictionary learning module) and the dictionary matrix of this layer. It optimizes and solves the output sparse representation coefficients to extract key features of the input signal. The input of the dictionary update layer is the sparse representation coefficients output by the sparse coding layer. It uses a dictionary learning algorithm based on nonlinear transformation to update the dictionary matrix of this layer, obtaining the updated dictionary for this layer, thereby improving nonlinear expressive power and suppressing noise and insignificant features in the signal. The sparse representation coefficient calculation module takes as input the original data of each gram 2D ​​feature map and the updated deep learning dictionary for each category. It calculates the sparse representation coefficients of each gram 2D ​​feature map under each category deep learning dictionary using the sparse representation method, and obtains the reconstructed gram 2D ​​feature map under each category deep learning dictionary using the signal reconstruction algorithm. It then compares the reconstructed gram 2D ​​feature map under each category deep learning dictionary with the original data. Finally, it outputs the reconstruction residual of each gram 2D ​​feature map under each category deep learning dictionary. The smaller the error, the higher the matching degree with the category. Classification Decision Module: The input is the reconstruction residual of each Gram 2D ​​feature map under each category deep learning dictionary. The class with the smallest reconstruction residual is selected as the sample class, and the output is the class label of each Gram 2D ​​feature map. Finally, the broadband oscillation classification and recognition model was tested using a test set divided from the two-dimensional image sample dataset. The model parameters were further optimized and adjusted to obtain the broadband oscillation classification and recognition model with the best performance.

4. A novel wide-area monitoring method for broadband oscillations in power systems according to claim 3, characterized in that, Step S5 includes the following steps: S51. Based on the one-dimensional observation signal sample dataset, the main station uses the sparsity-adaptive K-SVD dictionary learning method to design the main station classification sparse basis, that is, the main station classification sparse transform domain. S52. The main station uses the random measurement matrix generation model constructed based on the Logistic chaotic mapping method shown in Equation (1) to reconstruct the random measurement matrix, and multiplies it by the main station classification sparse basis to obtain the perception matrix. S53. Based on compressed sensing theory and sensing matrix, a broadband oscillation signal reconstruction model is constructed. The model structure is as follows: Input layer: The input is a one-dimensional observation signal; Signal reconstruction layer: The one-dimensional observed signal is reconstructed using a signal reconstruction algorithm based on sparse representation. The reconstructed sparse signal is obtained by using deep learning dictionaries of various categories and sparse representation coefficients. Signal recovery layer: Recovers the reconstructed sparse signal and outputs the recovered wideband oscillation signal, i.e., the reconstructed wideband oscillation signal; Finally, the broadband oscillation signal reconstruction model was trained and tested using the training and test sets, and the model parameters were optimized to obtain the broadband oscillation signal reconstruction model with the best performance. S54. Utilize the best-performing broadband oscillation signal reconstruction model, with the input being the compressed signal uploaded by each substation and the output being the reconstructed broadband oscillation signal. S55. The main station performs comprehensive analysis based on the reconstructed broadband oscillation signal, including time-frequency analysis, wide-area monitoring, mechanism analysis, disturbance source tracing and location, data mining, scheduling decision-making, and prevention, protection and control.

5. A novel wide-area monitoring method for broadband oscillations in power systems according to claim 4, characterized in that, The signal reconstruction algorithms based on sparse representation include matching pursuit algorithm, orthogonal matching pursuit algorithm, regularized matching pursuit algorithm, basis pursuit algorithm, and alternating direction multiplier method.

Images