Wideband oscillation intelligent monitoring and data transmission method, system and device for power system and storage medium

Abstract

The invention relates to the technical field of electric power system signal processing, in particular to an electric power system broadband oscillation intelligent monitoring and data transmission method, system and device and a storage medium. The method comprises the following steps: carrying out modal decomposition on a broadband oscillation signal by adopting variational modal decomposition, rejecting a noise dominant component and reserving an effective modal component, constructing a self-adaptive fixed-order TLS-ESPRIT algorithm based on a singular value adjacent growth ratio, adaptively determining a signal subspace order, establishing an improved variational auto-encoder to carry out data compression processing, and carrying out data compression processing. Label data is used as an additional condition to be fused into a coding process, the modal aliasing problem of a traditional method in complex signal processing is solved, and accurate and automatic identification of broadband oscillation parameters in a strong noise environment is achieved. The limitation that a traditional auto-encoder is poor in adaptability to new frequency components is overcome, efficient data compression is achieved on the premise that reconstruction precision is guaranteed, and the problem of communication bandwidth limitation under high-speed sampling is solved. And a complete technical chain from signal preprocessing to parameter identification to data transmission is formed.

Description

Technical Field

[0001] This invention relates to the field of power system signal processing technology, and in particular to a method, system, device and storage medium for intelligent monitoring and data transmission of broadband oscillations in power systems. Background Technology

[0002] With the large-scale integration of new energy sources into the power grid, broadband oscillations in the power system are becoming increasingly prominent. These oscillations can range in frequency from a few tenths of a hertz to several hundred hertz or even higher, involving multiple frequency components and exhibiting a wide frequency distribution. Broadband oscillations not only threaten the stable operation of the power grid but can also trigger large-scale power outages. Therefore, accurate monitoring and timely early warning of broadband oscillations are of great significance.

[0003] Traditional oscillation monitoring methods suffer from numerous technical shortcomings. The Prony algorithm cannot effectively capture rapidly decaying oscillation components, resulting in insufficient parameter identification accuracy. Empirical mode decomposition (EMD) methods suffer from mode aliasing and endpoint effects, making it difficult to accurately separate the true oscillation modes. Variational mode decomposition combined with Hilbert transform or Teager-Kaiser energy operators is extremely sensitive to noise interference, and the physical parameters are not fully extracted. Although the TLS-ESPRIT algorithm has strong noise resistance, it relies on manually pre-determining the number of frequency components in the signal. If the setting is inaccurate, it will severely undermine the orthogonality assumption between the signal subspace and the noise subspace, making it impossible to effectively separate the noise.

[0004] On the other hand, existing wide-area measurement systems mainly focus on monitoring signals near the power frequency. When dealing with high-frequency signals in broadband oscillations, the Nyquist sampling theorem requires sampling frequencies of several kilohertz or higher, leading to a surge in data volume that existing communication bandwidth cannot support. Compressed sensing methods heavily rely on data sparsity; insufficient sparsity can cause a significant drop in reconstruction accuracy. Traditional autoencoders use deterministic mapping for compression, resulting in poor adaptability to new frequency components outside of training and weak noise resistance. Summary of the Invention

[0005] In view of the problems existing in the prior art, the present invention is proposed.

[0006] Therefore, the problem to be solved by this invention is how to achieve accurate parameter identification of broadband oscillation signals in a strong noise environment, and how to achieve efficient transmission and accurate reconstruction of broadband oscillation data under limited communication bandwidth conditions, so as to provide a technical solution for broadband oscillation monitoring of new energy power grids that combines high precision, strong noise resistance and low bandwidth dependence.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0008] In a first aspect, embodiments of the present invention provide a method for intelligent monitoring and data transmission of broadband oscillations in a power system, comprising: acquiring a broadband oscillation signal of a power system, and performing mode decomposition on the broadband oscillation signal using variational mode decomposition to obtain multiple intrinsic mode functions;

[0009] The noise-dominant components in the multiple intrinsic mode functions are removed, and the effective mode components are retained;

[0010] An adaptive order-determining TLS-ESPRIT algorithm is constructed to adaptively determine the order of the signal subspace by the singular value adjacency growth ratio.

[0011] The effective modal components are input into the adaptive order-determining TLS-ESPRIT algorithm to identify parameters of each modal component and extract the frequency and amplitude information of the broadband oscillation.

[0012] As a preferred embodiment of the intelligent monitoring and data transmission method for broadband oscillations in power systems according to the present invention, wherein: the removal of noise-dominant components from the plurality of intrinsic mode functions includes:

[0013] Spectral analysis is performed on the multiple intrinsic mode functions; pure noise-dominated intrinsic mode functions are identified based on spectral characteristics; and non-noise-dominated intrinsic mode functions are retained as effective mode components.

[0014] As a preferred embodiment of the intelligent monitoring and data transmission method for broadband oscillations in power systems according to the present invention, wherein: determining the order of the signal subspace based on the singular value adjacency growth ratio includes:

[0015] Construct a Hankel matrix and perform singular value decomposition; calculate the adjacent growth ratio based on the singular value sequence; determine the order of the signal subspace based on the minimum value of the adjacent growth ratio to achieve adaptive partitioning of the signal subspace and noise subspace.

[0016] As a preferred embodiment of the intelligent monitoring and data transmission method for broadband oscillations in power systems according to the present invention, wherein: inputting the effective modal components into the adaptive order-determining TLS-ESPRIT algorithm includes:

[0017] A Hankel matrix is ​​constructed for the effective modal components and singular value decomposition is performed; the signal subspace and noise subspace are divided based on the determined signal subspace order; the eigenvalues ​​are solved by the overall least squares method, and the frequency, damping, amplitude and phase parameters of the oscillation signal are calculated.

[0018] As a preferred embodiment of the power system broadband oscillation intelligent monitoring and data transmission method of the present invention, the method further includes:

[0019] Determine whether a wideband oscillation signal is detected; if a wideband oscillation signal is detected, initiate the data transmission processing procedure to perform data compression processing on the wideband oscillation signal.

[0020] As a preferred embodiment of the intelligent monitoring and data transmission method for broadband oscillations in power systems according to the present invention, the step of performing data compression processing on the broadband oscillation signal includes:

[0021] An improved variational autoencoder is established, comprising an encoder and a decoder; the encoder encodes a wideband oscillation signal into latent variables and label values; the latent variables and label values ​​are concatenated and then input into the decoder to reconstruct the signal.

[0022] As a preferred embodiment of the intelligent monitoring and data transmission method for broadband oscillations in power systems according to the present invention, the construction of the improved variational autoencoder includes:

[0023] A convolutional neural network is used as the encoder, and an exponential linear unit is used as the activation function. A label value processing module is added between the encoder and the decoder to take the label data as additional condition input. A deconvolutional neural network is used as the decoder, and the loss function is calculated through cross-entropy to ensure the authenticity of the output samples. The posterior distribution is made close to the prior distribution by minimizing the KL divergence.

[0024] Secondly, embodiments of the present invention provide a power system broadband oscillation intelligent monitoring and data transmission system, which includes a signal acquisition module for acquiring broadband oscillation signals of the power system and performing mode decomposition on the broadband oscillation signals using variational mode decomposition to obtain multiple intrinsic mode functions;

[0025] The signal preprocessing module is used to remove the noise-dominant components from the multiple intrinsic mode functions and retain the effective mode components.

[0026] The adaptive algorithm module is used to construct the adaptive order-determining TLS-ESPRIT algorithm, which adaptively determines the order of the signal subspace by the singular value adjacency growth ratio;

[0027] The parameter identification module is used to input the effective modal components into the adaptive order determination TLS-ESPRIT algorithm, identify the parameters of each modal component, and extract the frequency and amplitude information of the broadband oscillation.

[0028] Thirdly, embodiments of the present invention provide a computer device, including a memory and a processor, wherein the memory stores a computer program, and the computer program instructions, when executed by the processor, implement the steps of the power system broadband oscillation intelligent monitoring and data transmission method as described in the first aspect of the present invention.

[0029] Fourthly, embodiments of the present invention provide a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program instructions are executed by a processor, they implement the steps of the power system broadband oscillation intelligent monitoring and data transmission method as described in the first aspect of the present invention.

[0030] The beneficial effects of this invention are as follows: By employing variational mode decomposition to perform mode decomposition on broadband oscillating signals and removing noise-dominant components in the preprocessing step, this invention effectively solves the technical defects of the traditional Prony algorithm in capturing rapidly decaying oscillating components and the mode aliasing problem in the EMD method. This achieves high-quality separation of complex broadband oscillating signals and significantly improves the signal quality of subsequent processing. By constructing an adaptive order-determining TLS-ESPRIT algorithm based on the singular value adjacency growth ratio, this invention overcomes the technical bottleneck of the traditional TLS-ESPRIT algorithm, which relies on manually pre-determining the number of frequency components. It avoids the problem of orthogonality destruction between the signal subspace and noise subspace caused by inaccurate K-value settings, achieving automated and accurate identification of broadband oscillation parameters under strong noise conditions. Compared with traditional methods, it produces unexpected technical improvements in noise resistance and identification accuracy. By establishing an improved variational autoencoder for data compression processing and incorporating tag data as an additional condition into the encoding process, this invention overcomes the limitation of the traditional autoencoder's deterministic mapping in poor adaptability to new frequency components outside the training stage. It achieves efficient data compression while ensuring reconstruction accuracy in the mid-to-high frequency bands, solving the communication bandwidth limitation problem faced by existing WAMS technology under high-speed sampling. The entire technical solution, through the organic combination of three core technologies, forms a complete chain from signal preprocessing to parameter identification and then to data transmission. It provides a systematic solution for broadband oscillation monitoring of new energy power grids that combines high precision, strong noise resistance and low bandwidth dependence, achieving the technical goal of realizing real-time wide-area monitoring of broadband oscillations in complex power grid environments. Attached Figure Description

[0031] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 A flowchart of a method for intelligent monitoring and data transmission of broadband oscillations in power systems;

[0033] Figure 2 A diagram of computer equipment for a broadband oscillation intelligent monitoring and data transmission method for power systems;

[0034] Figure 3A flowchart of the adaptive TLS-ESPRIT oscillation monitoring process for intelligent monitoring and data transmission of broadband oscillations in power systems;

[0035] Figure 4 A schematic diagram of a broadband oscillation monitoring model combining adaptive TLS-ESPRIT and an improved variational autoencoder, which is a method for intelligent monitoring and data transmission of broadband oscillations in power systems;

[0036] Figure 5 A schematic diagram of the mode decomposition results of broadband oscillation signals based on VMD for intelligent monitoring and data transmission of broadband oscillations in power systems;

[0037] Figure 6 A schematic diagram illustrating the broadband oscillation signal identification effect based on TLS-ESPRIT for intelligent monitoring and data transmission of broadband oscillations in power systems;

[0038] Figure 7 This is a comparison chart showing the signal reconstruction effect of different oscillation frequencies for intelligent monitoring and data transmission methods of broadband oscillations in power systems. Detailed Implementation

[0039] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0040] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0041] Secondly, the term "an embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places throughout this specification does not necessarily refer to the same embodiment, nor is it a single embodiment or an embodiment selectively excluded from other embodiments.

[0042] Example 1

[0043] Reference Figures 1-2 This is the first embodiment of the present invention, which provides a method for intelligent monitoring and data transmission of broadband oscillations in power systems, including:

[0044] S100: Acquire the broadband oscillation signal of the power system, and use variational mode decomposition to perform mode decomposition on the broadband oscillation signal to obtain multiple intrinsic mode functions;

[0045] S200: Removes noise-dominant components from multiple intrinsic mode functions and retains effective mode components;

[0046] S300: Construct an adaptive order-determining TLS-ESPRIT algorithm to adaptively determine the order of the signal subspace by the singular value adjacency growth ratio;

[0047] S400: Input the effective modal components into the adaptive order determination TLS-ESPRIT algorithm to identify the parameters of each modal component and extract the frequency and amplitude information of the broadband oscillation.

[0048] Steam pipelines in thermal power plants are thick-walled pipes. During operation, parameters such as temperature, pressure, and medium flow rate inside and outside the pipelines vary significantly, resulting in substantial temperature and stress gradients. The pipe wall thickness also changes due to the backflow of condensate. Operating in a high-temperature, high-pressure environment makes measuring temperature and stress gradients difficult, often leading to delayed safety warnings. Furthermore, the scouring effect of steam within the pipeline continuously thins the pipe walls, jeopardizing pipeline health. The high-temperature, high-pressure environment significantly impacts wall thickness measurement and can also cause fatigue-related damage. Therefore, monitoring and predicting pipeline health is crucial.

[0049] To address the aforementioned issues in operation monitoring and health prediction, a model is constructed through steps S100-S400 to simulate the operating state of the pipeline under typical working conditions. This yields the temperature and stress distribution diagrams and relationships of the pipeline under the influence of wall thickness, enabling accurate estimation of inner wall temperature and stress. Real-time monitoring of pipeline wall thickness is achieved, and the pipeline wall thinning rate is dynamically calculated, providing early warning for abnormal increases in the pipeline thinning rate. Simultaneously, based on fatigue damage and remaining life models, accurate prediction of the pipeline's remaining life is realized.

[0050] This invention combines variational mode decomposition (VMD) with adaptive TLS-ESPRIT. First, VMD is used to eliminate noise-dominated IMF components, improving the signal-to-noise ratio. Then, an adaptive order determination strategy avoids the artificial order determination errors of traditional TLS-ESPRIT, achieving accurate mode-by-mode parameter identification for broadband multimode signals. When a broadband oscillation signal is detected, an improved variational autoencoder is used to compress high-dimensional data into a low-dimensional space, reducing data transmission while maintaining reconstruction accuracy, overcoming the bandwidth limitations of traditional WAMS high-speed sampling. The entire technical solution exhibits a frequency relative error of less than 0.5% and an amplitude relative error of less than 1% under high-intensity noise, significantly improving noise suppression capabilities compared to traditional methods. This provides a high-precision, low-bandwidth-dependent solution for broadband oscillation monitoring in new energy power grids.

[0051] Explanation of English characters:

[0052] VMD (Variational Mode Decomposition) is an adaptive signal processing method that can decompose a signal into multiple modal components with specific center frequencies.

[0053] TLS-ESPRIT (Totalleast Squares-Estimation of Signal Parameters via Rotational Invariance Technique): This technique uses total least squares to estimate signal parameters by dividing the signal into noise and signal spaces.

[0054] IVAE (Improved Variational Autoencoder): An improved variational autoencoder that adds tag value processing to the traditional VAE, improving the encoding and compression effect for specific types of data.

[0055] IMF (Intrinsic Mode Functions): These are signal components with specific frequency characteristics obtained after EMD or VMD decomposition.

[0056] WAMS (Wide Area Measurement System): A power system monitoring network based on synchronous phasor measurement units.

[0057] Example 2

[0058] Reference Figures 3-7 This is the second embodiment of the present invention.

[0059] In this embodiment of the application, step S100, which involves acquiring the broadband oscillation signal of the power system and performing variational mode decomposition, includes the following steps A1-A3:

[0060] A1: Obtain the broadband oscillation signal of the power system

[0061] In power systems, broadband oscillation signals are primarily acquired using synchronous phasor measurement units (PMUs). Considering that the frequency range of broadband oscillations can extend from a few tenths of a hertz to several hundred hertz or even higher, according to the Nyquist sampling theorem, the sampling frequency must be at least twice the highest frequency of the signal. Therefore, this invention sets the sampling frequency to 2 kHz to ensure that high-frequency oscillation components can be captured. The acquired raw signals typically contain electrical quantities such as voltage, current, and power. During transmission, these signals are inevitably affected by factors such as noise from measurement equipment, interference from power electronic devices, and environmental electromagnetic interference, leading to a deterioration in signal quality.

[0062] Specifically, the original broadband oscillation signal can be represented as a superposition of multiple frequency components, containing useful oscillation information and various noise interferences. The signal's time-domain characteristics are characterized by complex amplitude variations and rich frequency components, while its frequency-domain characteristics show multiple peaks distributed at different frequency points. For subsequent processing, the acquired original signal undergoes preprocessing, including outlier removal and power frequency interference filtering, to ensure that the input variational mode decomposition signal has a good quality foundation.

[0063] A2: Constructing a variational mode decomposition constrained variational model

[0064] Variational mode decomposition (VMD) is based on the theoretical foundations of Wiener filtering, Hilbert transform, and frequency-domain hybrid demodulation, transforming the signal decomposition problem into a constrained variational problem. Its core idea is to adaptively decompose the original signal into K eigenmode functions with finite bandwidths, each mode function being compactly distributed around its center frequency. The goal of the constrained variational model is to minimize the sum of the estimated bandwidths of all modes, while ensuring that the sum of all modes can reconstruct the original signal.

[0065] Constrained variational models comprise two key elements: an objective function and constraints. The objective function achieves efficient signal decomposition by minimizing the bandwidth of each modal component, while the constraints ensure the integrity of the decomposition results. By introducing a penalty factor α and the Lagrange multiplication operator λ, the constrained variational problem is transformed into an unconstrained variational problem, facilitating subsequent numerical solutions. The penalty factor α controls the balance between data fidelity and signal reconstruction error; a larger α value helps obtain fewer modes but may lose some signal details, while a smaller α value may generate too many modal components.

[0066] A3: Solving variational problems using the alternating direction multiplier method.

[0067] The Alternating Direction Multiplier Method (ADMM) is an efficient numerical method for solving constrained optimization problems, particularly suitable for separable optimization problems. In VMD, ADMM is used to iteratively update the modal components uk and the center frequency ωk. The algorithm first initializes each modal component and the center frequency, and then alternately updates each modal component and its corresponding center frequency through frequency domain transformation and inverse transformation until the convergence condition is met.

[0068] VMD Principles

[0069] Considering that broadband oscillation signals in actual power grids have characteristics such as wide frequency domain, strong time variation, strong nonlinearity, and multiple modes, this invention considers first using VMD to perform mode decomposition and feature extraction on the monitored electrical quantity signals.

[0070] VMD (Variable Frequency Decomposition) has significant advantages in processing complex nonlinear time series, effectively improving signal stationarity. This method decomposes the original signal into multiple stationary sub-components with different frequency characteristics, exhibiting good adaptability and frequency resolution. The specific form of the constrained variational model is as follows:

[0071]

[0072] In the formula, u k ={u1,...,u k} represents each modal component; ω k ={ω1,...,ω k} represents the frequency of each component; f(t) is the initial signal; δ(t) is the Dirac function.

[0073] By introducing a penalty factor α and a Lagrange multiplier λ, the constrained problem is transformed into an unconstrained variational problem.

[0074]

[0075] The unconstrained problem is solved using the alternating direction multiplier method, thereby achieving effective separation of signal frequencies. The iterative update formulas for the intrinsic mode components and the center frequency are as follows:

[0076]

[0077] In the formula: This refers to the k-th eigenmode component with center frequency ω at the (n+1)-th iteration. It is the center frequency of the k-th intrinsic mode component in the (n+1)-th iteration, which is the centroid of the power spectrum of the current mode component.

[0078] Modal component updates are achieved through Wiener filtering in the frequency domain, optimizing the modal components using the current center frequency estimate and Lagrange multipliers. The center frequency is updated by calculating the power spectral centroid of the corresponding modal component, ensuring that each mode has the most compact spectral distribution around its center frequency. During iteration, the Lagrange multipliers act as a binding constraint, ensuring that the sum of all modes equals the original signal.

[0079] In an optional implementation, the variational mode decomposition in step S100 can also optimize the decomposition effect by adaptively selecting the number of modes K. Traditional VMD requires pre-setting the number of modes, while the adaptive method can dynamically determine the optimal number of modes based on the complexity of the signal. Specifically, the optimal value of K can be determined by analyzing the correlation of continuous decomposition results, monitoring the changing trend of decomposition residuals, or using information theory criteria, thus avoiding over-decomposition or under-decomposition.

[0080] In this embodiment of the application, step S200, which involves removing the dominant noise component and retaining the effective modal components, includes the following steps B1-B3:

[0081] B1: Removes noise-dominant components from multiple intrinsic mode functions, including:

[0082] Spectral analysis is performed on multiple intrinsic mode functions; pure noise-dominated intrinsic mode functions are identified based on spectral characteristics; and non-noise-dominated intrinsic mode functions are retained as effective mode components.

[0083] The intrinsic mode functions (IMFs) obtained after VMD decomposition have different frequency characteristics and energy distributions. To identify and eliminate noise-dominant components, a detailed spectral analysis is first performed on each IMF. The spectral analysis uses Fast Fourier Transform (FFT) to convert the time-domain signal to the frequency domain and calculates the amplitude and phase information at each frequency point. By analyzing the spectral characteristics, key parameters such as the dominant frequency, bandwidth, and energy concentration of each IMF can be identified.

[0084] Useful oscillation modes typically exhibit significant energy concentration around specific frequencies, displaying sharp peaks in their spectrum, while noise-dominated modes exhibit a wide-bandgap flat distribution or random characteristics. Furthermore, it is necessary to analyze the time-domain characteristics of each IMF, including signal continuity, periodicity, and amplitude variation patterns. Realistic oscillation modes generally possess good periodicity and continuity, while noise components exhibit strong randomness and lack obvious periodicity.

[0085] B2: Identification of Noise-Dominated Eigenmode Functions Based on Spectral Features

[0086] Noise-dominated IMF identification is primarily based on the following criteria: spectral flatness, energy concentration, correlation analysis, and time-domain feature analysis. Spectral flatness reflects the uniformity of signal energy distribution in the frequency domain; noisy signals typically have high spectral flatness, while the energy of useful signals is mainly concentrated around specific frequencies. Energy concentration is measured by calculating the ratio of the dominant frequency component to the total energy; useful signals usually have high energy concentration.

[0087] Correlation analysis assesses the importance of each mode by calculating the correlation coefficient between each IMF and the original signal. Noise-dominated components have low correlation with the original signal, while modes containing useful information have high correlation. Time-domain characteristic analysis mainly examines the signal's stationarity, periodicity, and amplitude variation characteristics. The nature of the modes can be comprehensively judged by calculating the signal's autocorrelation function, power spectral density, and statistical characteristics (such as mean, variance, skewness, kurtosis, etc.).

[0088] Specifically, a comprehensive judgment index is established: multiple criteria are set, such as spectral flatness threshold, energy concentration threshold, and correlation coefficient threshold. When multiple indicators of an IMF simultaneously meet the noise characteristics, it is identified as the dominant noise component. To improve the accuracy of identification, machine learning methods can be used to train a classifier and a feature model of noise and useful signals can be established using historical data.

[0089] B3: Retain non-noise-dominated eigenmode functions as effective modal components.

[0090] After noise identification, non-noise-dominated IMFs are retained as effective modal components for subsequent processing. The selection of effective modal components must consider not only the characteristics of individual IMFs but also the interrelationships between modes and the overall reconstruction effect. The retained effective modes should be able to reconstruct the main features of the original signal and possess good physical meaning and engineering value.

[0091] When selecting effective modes, reconstruction error analysis is necessary to ensure that the retained mode components accurately reproduce the key information of the original signal. Reconstruction error can be assessed by calculating metrics such as the mean square error (MSE) and correlation coefficient between the reconstructed signal and the original signal. If the reconstruction error is too large, it may be necessary to re-evaluate the noise identification criteria or adjust the VMD parameters.

[0092] In an optional implementation, the noise removal in step S200 can be combined with other signal processing methods such as wavelet denoising and morphological filtering to further improve the denoising effect. Wavelet denoising utilizes the multi-resolution characteristics of wavelet transform to remove small-scale noise components through thresholding; morphological filtering uses opening and closing operations in mathematical morphology to remove impulse noise and smooth signals. These methods can complement VMD to form a multi-level noise suppression system.

[0093] In this embodiment of the application, the construction of the adaptive order-determining TLS-ESPRIT algorithm in step S300 includes the following steps C1-C4:

[0094] C1: Determining the order of the signal subspace based on the singular value adjacency growth ratio, including:

[0095] Construct the Hankel matrix and perform singular value decomposition; calculate the adjacent growth ratio based on the singular value sequence; determine the order of the signal subspace based on the minimum value of the adjacent growth ratio, and realize the adaptive partitioning of the signal subspace and the noise subspace.

[0096] The first step of the TLS-ESPRIT algorithm is to construct a Hankel matrix based on the input time series data. A Hankel matrix is ​​a special matrix structure where all elements on the antidiagonal are equal, effectively embedding the time series delay information. For time series data of length N, the constructed Hankel matrix has a dimension of M×L, where the choices of M and L must satisfy the constraint M + L - 1 = N. The choice of matrix dimension has a significant impact on algorithm performance; generally, choosing M ≈ N / 3 yields good results.

[0097] a) TLS-ESPRIT algorithm

[0098] The TLS-ESPRIT algorithm is based on the ESPRIT algorithm, but it uses Total Least Squares (TLS) instead of the Least Squares (LS) technique in ESPRIT. In Least Squares, it is assumed that noise exists only in the observations, while the signal itself is accurate. Therefore, when the signal-to-noise ratio (SNR) is low, Least Squares may lead to significant estimation bias, affecting the algorithm's performance. In contrast, Total Least Squares assumes that noise exists not only in the observations but also potentially in the signal itself. Therefore, Total Least Squares can better handle situations with low SNR, thus improving the algorithm's performance. Its implementation process is as follows:

[0099] 1) Construct the Hankel matrix X based on the sampled signal:

[0100]

[0101] In the formula, N is the number of sampled oscillation data, M = N / 2, and L = N - M + 1.

[0102] 2) Perform singular value decomposition on the Hankel matrix X above:

[0103] X=UΣV H (6)

[0104] In the formula, U is an L×L matrix; V is an M×M matrix; Σ is an L×M matrix with diagonal elements being the singular values ​​of matrix X.

[0105] 3) Based on the magnitude of the singular values, V is divided into V1 and V2, and U is divided into U... s and U n Assuming the order of the signal subspace is p, equation (6) can be rewritten as:

[0106]

[0107] In the formula, Σ s Σ n All are diagonal matrices; Σs It consists of the p singular values ​​with the largest magnitudes in matrix X, Σ n It consists of the remaining singular values.

[0108] 4) Construct a new matrix [V1, V2] and perform singular value decomposition on it:

[0109]

[0110] In the formula, V1 is V s Remove the last row of the matrix; V2 is V s Remove the first row of the matrix.

[0111] Will Decomposed into a matrix of 4:

[0112]

[0113] Then we have:

[0114]

[0115] φ tls The eigenvalue is λ i (i = 1, 2, 3, ..., p), the frequency f of the oscillation can be obtained from the eigenvalues. i and damping ζ i :

[0116] f i =arctan(Im(λ) i ) / Re(λ i (11)

[0117]

[0118] 5) The amplitude and phase are determined using the total least squares method. First, φ is used... tls eigenvalues ​​λ i Construct a new Z matrix:

[0119]

[0120] Calculate the fitting coefficient b:

[0121] b = (Z H Z) -1 Z H x (14)

[0122] Calculate the signal amplitude A i and initial phase θ i :

[0123] A i =2|b| (15)

[0124]

[0125] b) Adaptive order determination algorithm

[0126] Determining the order of the Hankel matrix X involves dividing the signal subspace V1 and the noise subspace V2. For an unfamiliar signal, it's difficult to predetermine the number of its actual frequency components. Assuming K represents the number of actual frequency components, if K is set too small, some actual frequency components will be missed, causing bias in the acquisition of the signal containing parameter information; if K is set too large, spurious frequency components will be mixed into the acquired signal, which will significantly impact subsequent processing. Therefore, the adjacency growth ratio ΔS of singular values ​​is used here to determine the number of components K. Assuming M is an m×n matrix, and all elements of M belong to the same domain, there exists a decomposition such that:

[0127] M=UΣV * (17)

[0128] In the formula, U is an m×m matrix; Σ is an m×n matrix; V * It is an n×n unitary matrix. The elements on the diagonal are Σ. i , where Σ i That is, the singular value of .

[0129] A common practice is to arrange the singular values ​​in descending order so that Σ can be uniquely determined by M.

[0130] A discrete-time series x is obtained by sampling, and the Hankel matrix X is obtained by spatial reconstruction of it.

[0131]

[0132] In the formula, N is the number of sampled oscillation data, M = N / 2, and L = N - M + 1.

[0133] Perform singular value decomposition on X and take its diagonal matrix ∑:

[0134]

[0135] Arrange the elements on the diagonal from largest to smallest, that is:

[0136] σ1≥σ2≥…≥σ K ≥…≥σ i ≥…≥σ n (20)

[0137] In the formula, K is the number of useful frequency components in the singular values.

[0138]

[0139] In the formula, P K ΔS represents the percentage of singular value sequence operations; ΔS represents the singular value adjacency growth ratio.

[0140] The number of valid signals depends on the actual number of frequency components, where K represents the number of valid components, and the number following it represents the noise signal. Since the cumulative percentage increments on either side of the K value differ by orders of magnitude, there must exist a minimum value ΔS in the distribution of ΔS. min The minimum value corresponding to K is the number of true frequency components in the signal. When there is no noise, P after order K... K It will tend to 1; for noise-free cases, add a judgment P. K -P K-1 <10 -6 Then the adjacent growth ratio will not be calculated.

[0141] The entire adaptive TLS-ESPRIT algorithm oscillation monitoring process is as follows: Figure 3 As shown.

[0142] After constructing the Hankel matrix, we perform singular value decomposition (SVD) on it. Singular value decomposition is an important tool in linear algebra, capable of decomposing any matrix into a product of three matrices. Through SVD decomposition, we obtain a sequence of singular values ​​of the matrix. These singular values ​​are arranged in descending order, reflecting the energy magnitude of different components in the signal. Larger singular values ​​correspond to the signal subspace, and smaller singular values ​​correspond to the noise subspace.

[0143] The mathematical significance of Singular Value Decomposition (SVD) lies in decomposing the original signal space into orthogonal subspaces, where the main subspaces contain the primary information of the signal, while the remaining subspaces mainly contain noise information. This decomposition provides the theoretical basis for subsequent signal-to-noise separation. In practical calculations, SVD has high computational complexity, especially for large matrices, requiring efficient numerical algorithms to ensure computational efficiency.

[0144] C2: Calculating the adjacent growth ratio based on singular value sequences

[0145] The traditional TLS-ESPRIT algorithm requires manually setting the dimension of the signal subspace. This method is highly subjective and prone to subspace partitioning errors, affecting parameter identification accuracy. To address this issue, this invention employs an adaptive order determination method based on the singular value adjacency growth ratio. The adjacency growth ratio is defined as the ratio of adjacent singular values, reflecting the changing trend of the singular value sequence.

[0146] Specifically, for a singular value sequence σ1≥σ2≥...≥σr, the adjacent growth ratio ri=σi+1 / σi is calculated. In an ideal noise-free environment, the singular values ​​corresponding to the signal components are larger and relatively stable, while the singular values ​​corresponding to the noise components are smaller and fluctuate drastically. Therefore, the adjacent growth ratio exhibits a significant jump at the boundary between the signal subspace and the noise subspace. By analyzing the variation pattern of the adjacent growth ratio, the dimension of the signal subspace can be automatically determined.

[0147] To improve the robustness of the adjacent growth ratio method, the impact of noise on the singular value distribution needs to be considered. In noisy conditions, small singular values ​​will not be strictly equal to zero, but will fluctuate within a small range. Therefore, an appropriate threshold needs to be set to determine significant changes in the adjacent growth ratio. The threshold can be selected based on an estimate of the noise level or through methods such as cross-validation.

[0148] C3: Determine the order of the signal subspace based on the minimum value of the adjacent growth ratio.

[0149] The minimum position in the adjacent growth ratio sequence typically corresponds to the boundary between the signal subspace and the noise subspace. This is because at the boundary, singular values ​​jump from relatively large signal component values ​​to relatively small noise component values, resulting in a minimum adjacent growth ratio. By finding the minimum position of the adjacent growth ratio, the order p of the signal subspace can be automatically determined.

[0150] After determining the order of the signal subspace, the singular value decomposition result is divided into two parts: a signal subspace and a noise subspace. The signal subspace consists of the left and right singular vectors corresponding to the first p singular values, containing the main information of the signal; the noise subspace consists of the vectors corresponding to the remaining singular values, mainly containing noise information. This division provides a reliable foundation for subsequent parameter estimation.

[0151] In practical applications, the adjacent growth ratio method may be affected by strong noise, causing the minimum value location to shift. To improve the stability of the method, multiple criteria can be combined, such as information theory criteria (AIC, MDL, etc.) and cross-validation methods, to help determine the optimal order. In addition, the rationality of the selected order can be verified through statistical analysis of multiple independent experiments.

[0152] C4: Establishing the complete process of adaptive order determination

[0153] The adaptive order determination algorithm TLS-ESPRIT transforms the traditional manual order determination process into a data-driven, automated process. The complete adaptive order determination workflow includes: data preprocessing, Hankel matrix construction, singular value decomposition, adjacent growth ratio calculation, order determination, and subspace partitioning. Each step has corresponding parameter settings and quality control measures to ensure the algorithm's stability and reliability.

[0154] To further improve the algorithm's adaptability, an online learning mechanism can be introduced to dynamically adjust algorithm parameters based on historical processing experience. For example, a threshold adaptive adjustment mechanism can be established to automatically select the optimal judgment threshold based on different signal characteristics and noise levels. Furthermore, uncertainty quantification methods can be introduced to evaluate the confidence level of the order estimate, providing a reliability indicator for subsequent parameter identification.

[0155] In this embodiment of the application, the parameter identification and data transmission processing in step S400 includes the following steps D1-D4:

[0156] D1: Input the effective modal components into the adaptive order-determining TLS-ESPRIT algorithm, including:

[0157] Construct Hankel matrices for the effective modal components and perform singular value decomposition; divide the signal subspace and noise subspace based on the determined signal subspace order; solve the eigenvalues ​​using the overall least squares method to calculate the frequency, damping, amplitude, and phase parameters of the oscillating signal.

[0158] For each valid mode component retained after noise removal, its corresponding Hankel matrix is ​​constructed. Since each mode component has different frequency characteristics and signal length, the dimensions of the Hankel matrix need to be appropriately adjusted according to the characteristics of each mode component. Generally, the number of rows M of the matrix should be large enough to contain sufficient time delay information, but not too large to avoid introducing excessive computational complexity.

[0159] When performing singular value decomposition on the Hankel matrix of each modal component, special attention must be paid to numerical stability. Since the amplitudes of the modal components can vary significantly, appropriate normalization is necessary to avoid numerical overflow or underflow. Furthermore, for some modal components with very small amplitudes, it may be necessary to improve numerical precision or employ special numerical algorithms to ensure the accuracy of the decomposition.

[0160] When performing singular value decomposition (SVD), computational efficiency must also be considered. For real-time monitoring applications, the algorithm's computation time must be kept within a reasonable range. Techniques such as fast SVD algorithms, parallel computing, or GPU acceleration can be used to improve computational efficiency. Furthermore, for highly periodic signals, the periodicity of the signal can be utilized to reduce the computational load.

[0161] Divide the signal subspace and noise subspace based on a determined signal subspace order.

[0162] Using the signal subspace order determined in step C3, the singular value decomposition results for each modal component are divided into subspaces. The signal subspace contains the main frequency information of the modal component, while the noise subspace contains the remaining noise and interference information. The quality of the subspace division directly affects the accuracy of subsequent parameter estimation; therefore, the rationality of the division results needs to be carefully verified.

[0163] After subspace partitioning, it is necessary to check the orthogonality of the signal subspace and the noise subspace. Theoretically, these two subspaces should be orthogonal, but in actual calculations, due to numerical errors, a certain degree of non-orthogonality may occur. The degree of orthogonality can be evaluated by calculating the angle between the subspaces or the projection error. If the non-orthogonality is too strong, it may be necessary to readjust the order or use other subspace estimation methods.

[0164] To improve the robustness of subspace estimation, methods such as subspace averaging or subspace projection can be used. Subspace averaging reduces the impact of random errors by averaging multiple independent estimation results; subspace projection suppresses noise by projecting the signal onto the estimated signal subspace. These methods can improve the stability of parameter estimation to some extent.

[0165] Eigenvalues ​​are solved using the total least squares method.

[0166] The core of the TLS-ESPRIT algorithm is to solve for the eigenvalues ​​in rotation-invariant relations using total least squares. Unlike traditional least squares methods, total least squares considers the errors of all elements in the data matrix, not just the errors of the observation vectors. This method can provide more accurate and stable estimation results even with low signal-to-noise ratios.

[0167] Specifically, the total least squares method solves the overdetermined system of equations by constructing an augmented matrix and performing singular value decomposition. The right singular vector corresponding to the smallest singular value of the augmented matrix gives the direction of the optimal solution. In TLS-ESPRIT, it is necessary to construct an augmented matrix consisting of upper and lower submatrices and solve its generalized eigenvalue problem. The calculation of eigenvalues ​​requires numerically stable algorithms, such as QZ decomposition or Schur decomposition.

[0168] In the process of eigenvalue calculation, special attention needs to be paid to the ordering and selection of eigenvalues. Due to numerical errors, some spurious eigenvalues ​​may appear in the results. Valid eigenvalues ​​need to be selected based on their magnitude, the relationship between their real and imaginary parts, and their physical meaning. Generally speaking, eigenvalues ​​corresponding to the true signal components should lie on or inside the unit circle and have a reasonable physical interpretation.

[0169] Calculate the frequency, damping, amplitude, and phase parameters of the oscillating signal.

[0170] Based on the eigenvalues ​​obtained from the TLS-ESPRIT algorithm, various parameters of the oscillation signal can be calculated. The frequency parameter is calculated from the amplitude of the eigenvalues, and the damping parameter is calculated from the magnitude of the eigenvalues. Specifically, for the eigenvalue λ = re^(jθ), the frequency f = θ / (2πΔt), and the damping δ = -ln(r) / Δt, where Δt is the sampling interval.

[0171] The calculation of amplitude and phase parameters needs to be performed in the time domain. First, a Vandermonde matrix is ​​constructed using the calculated frequency and damping parameters. Then, the complex amplitude is solved using the global least squares method. The magnitude of the complex amplitude gives the signal amplitude, and the phase angle gives the initial phase of the signal. During the calculation, attention must be paid to the numerical condition number, especially when there are multiple closely spaced frequency components, as the Vandermonde matrix may become ill-conditioned.

[0172] To improve the accuracy of parameter estimation, an iterative refinement method can be employed. First, initial parameter estimates are obtained using TLS-ESPRIT, and then the parameters are further optimized using nonlinear least squares. Furthermore, methods such as multiple signal classification (MUSIC) can be used to verify and supplement the TLS-ESPRIT results, thereby improving the reliability of parameter estimation through cross-validation of multiple methods.

[0173] D2: The method also includes:

[0174] Determine whether a wideband oscillation signal is detected; if a wideband oscillation signal is detected, initiate the data transmission processing procedure to compress the wideband oscillation signal.

[0175] After parameter identification is completed, it is necessary to determine whether a real broadband oscillation signal has been detected. The criteria for this determination include: whether the oscillation frequency is within the expected range, whether the oscillation amplitude exceeds the set threshold, whether the oscillation duration meets the requirements, and whether the signal-to-noise ratio of the oscillation signal is sufficiently high. These criteria need to be set according to the specific application scenario and engineering requirements.

[0176] Wideband oscillations typically range from 0.1 Hz to 1000 Hz, but different types of oscillations may be concentrated in specific frequency bands. For example, low-frequency oscillations are mainly in the range of 0.1–2 Hz, subsynchronous oscillations are in the range of 10–50 Hz, while high-frequency oscillations may be in the range of several hundred hertz. A reasonable frequency range needs to be set based on the specific characteristics of the power system.

[0177] Determining the oscillation amplitude requires considering both the fundamental component of the signal and its relative change. Generally, an oscillation amplitude should exceed the fundamental amplitude by a certain percentage (e.g., 1% or 2%) to be considered a meaningful oscillation. Simultaneously, the duration of the oscillation must be considered; momentary disturbances should not be considered broadband oscillations. A minimum duration threshold (e.g., 100ms or 1s) can be set to filter out brief interference signals.

[0178] D3: Performs data compression processing on the broadband oscillation signal, including:

[0179] An improved variational autoencoder is established, which includes an encoder and a decoder. The encoder encodes the broadband oscillating signal into latent variables and tag values. The latent variables and tag values ​​are concatenated and then input into the decoder to reconstruct the signal.

[0180] A variational autoencoder (VAE) is a generative model that combines the ideas of autoencoders and variational inference. Its network structure mainly consists of two parts: an encoder and a decoder. The encoder maps the input data x to random variable parameters in the latent space that follow a probability distribution, including the expected value μ and the standard deviation σ. 2 The decoder then reconstructs the input data based on the latent variable z sampled from this distribution.

[0181] In a VAE, the latent variable z and the input data x can form a joint probability density distribution p(z,x), and the KL divergence is used to measure the posterior distribution p. θ (z,x) and prior distribution p φ The difference between (z,x) is minimized by the KL divergence to ensure that the posterior distribution approximates the prior distribution. The formula is as follows:

[0182]

[0183] Traditional Visual Encoding Engines (VAEs), as unsupervised learning generative models, rely on the probability distribution of the latent space during the generation process. However, without additional conditions, this distribution may not be able to generate targeted data for specific types. Therefore, this invention considers labeling broadband oscillatory data, changing the traditional VAE structure by inputting labeled data as additional conditions into the encoder, thereby enabling the model to generate corresponding types based on these labels.

[0184] Compared to traditional VAEs, the Improved Variational Autoencoder (IVAE) adds a label value y between the encoder and decoder. During training, real samples are passed through the encoder to generate label values ​​y in an unsupervised manner. The latent variable z(μ′,σ,ε) can be calculated from y, μ,σ,ε, where ε is a noise term conforming to a standard normal distribution. Here, μ′ = μ - y, meaning that feature information is removed from the latent variable while retaining it in the label value y. Finally, the latent variable and the label value y are concatenated and input into the decoder to generate samples. Similar to traditional VAEs, its training objective is to make the posterior distribution p... θ (z,x,y) and prior distribution q φ Minimize the KL divergence between (z,x,y):

[0185] L(θ,φ)=KL(p θ (z,x,y)||q φ (z,x,y)) (24)

[0186] By replacing the joint probability density function with the conditional probability density function and rearranging it, and removing the constant term, we obtain equation (25):

[0187]

[0188] This invention employs a Convolutional Neural Network (CNN) as the data encoding network. The encoding network needs to extract features from the input samples. Furthermore, to avoid the data prematurely entering the nonlinear saturation region ("dead zone") when mapped to the activation function's domain, this invention uses an Exponential Linear Unit (ELU) as the activation function to activate the data. For the decoder, to ensure that the data processed by the decoding network can accurately restore the original dimensional space before the encoding operation, this invention adopts a deconvolutional neural network architecture that is symmetrical to the encoder, and uses cross-entropy to calculate the loss function to guarantee the authenticity of the output samples.

[0189] Framework of WAMS-based wideband oscillation monitoring method

[0190] First, the noisy broadband oscillation raw signal is acquired through a PMU. Variational Mode Decomposition (VMD) is then used to decompose it into multiple Intrinsic Mode Functions (IMFs), and noise-dominated IMFs are removed to complete signal preprocessing. Subsequently, for the retained effective IMF components, an adaptive TLS-ESPRIT algorithm is used for parameter identification. By constructing a Hankel matrix and performing singular value decomposition, an adaptive order determination strategy (based on the singular value adjacency growth ratio to determine the signal subspace order) is used to identify parameters such as frequency and amplitude of the oscillation signal mode by mode. This overcomes the shortcomings of traditional methods that require manual order determination and enhances noise immunity. If a broadband oscillation signal is identified, an improved variational autoencoder (IVAE) is used to encode and compress the signal, mapping high-dimensional data to a low-dimensional latent space to reduce data volume. After the compressed data is uploaded to the main station, it is decoded and reconstructed using IVAE to restore the original signal. This achieves efficient transmission and accurate restoration of the broadband oscillation signal under limited communication bandwidth, providing support for the main station's analysis and oscillation suppression.

[0191] In summary, the basic architecture of the broadband oscillation monitoring method based on the combination of adaptive TLS-ESPRIT and improved variational autoencoder is as follows: Figure 4 As shown:

[0192] D4: Construct an improved variational autoencoder, including:

[0193] A convolutional neural network is used as the encoder, and an exponential linear unit is used as the activation function. A label value processing module is added between the encoder and the decoder to take the label data as additional condition input. A deconvolutional neural network is used as the decoder, and the loss function is calculated through cross-entropy to ensure the authenticity of the output samples. The posterior distribution is made close to the prior distribution by minimizing the KL divergence.

[0194] Once a broadband oscillation signal is detected, the data transmission process is initiated. Because the broadband oscillation signal has a wide frequency range, a high sampling frequency is required according to the Nyquist sampling theorem, resulting in a massive data volume. Traditional communication networks cannot support such a large data transmission volume; therefore, data compression techniques are needed to reduce the amount of data transmitted.

[0195] The improved variational autoencoder (IVAE) is an enhanced version of the traditional variational autoencoder (VAE). While traditional VAEs are primarily used for unsupervised learning and generative modeling, IVAEs improve encoding performance by introducing label information, making them particularly suitable for compressing specific types of data. In broadband oscillatory data compression, features such as the frequency and amplitude of the oscillations can be used as label information to guide the encoder in learning more effective data representations.

[0196] The IVAE network structure includes an encoder, a decoder, and a label processing module. The encoder employs a convolutional neural network (CNN) structure, which effectively extracts local and temporal features of the signal. To prevent data from prematurely entering the saturation region when mapped to the activation function, an exponential linear unit (ELU) is used as the activation function. The ELU function has exponential properties on the negative half-axis and linear properties on the positive half-axis, which can maintain the effective propagation of gradients.

[0197] The decoder employs a deconvolutional neural network structure symmetrical to the encoder, reconstructing the low-dimensional latent representation into a signal of the original dimension through upsampling and deconvolution operations. To ensure the quality of the reconstructed signal, the loss function uses a weighted combination of reconstruction error and KL divergence. The reconstruction error measures the difference between the reconstructed signal and the original signal, while the KL divergence constrains the distribution of latent variables, ensuring good continuity and interpretability of the latent space.

[0198] The tag processing module is a major improvement of IVAE over traditional VAE. During the encoding process, in addition to encoding the input signal into a latent variable z, a tag variable y is generated to represent the signal's category or characteristic information. In broadband oscillation applications, the tag can contain information such as the oscillation's frequency range, amplitude level, and duration. By incorporating this prior knowledge into the encoding process, IVAE can learn a more compact and meaningful data representation.

[0199] During training, IVAE employs a variational inference framework to optimize network parameters. The objective function consists of two parts: a reconstruction loss and a regularization term. The reconstruction loss uses mean squared error or cross-entropy loss to measure reconstruction quality, while the regularization term uses KL divergence to constrain the distribution of latent variables. By minimizing the objective function, the network learns an effective compressed representation of the input data while maintaining reconstruction accuracy.

[0200] Specifically, the training process of IVAE is optimized using stochastic gradient descent or its variants (such as Adam, RMSprop, etc.). To improve training efficiency and stability, regularization techniques such as batch normalization and dropout can be employed. The training data needs to contain a large number of wideband oscillation samples, covering different frequencies, amplitudes, noise levels, etc., to ensure the model's generalization ability.

[0201] In the data compression process, the input broadband oscillating signal is first compressed by an encoder into a low-dimensional latent representation and label information. The dimension of the latent representation is usually much smaller than the dimension of the original signal, thus achieving the purpose of data compression. The compression ratio can be controlled by adjusting the dimension of the latent space, and can generally reach a compression ratio of 3:1 to 10:1, depending on the complexity of the signal and the reconstruction accuracy requirements.

[0202] The compressed data is transmitted to the dispatch center via the existing communication network. Due to the significant reduction in data volume, both transmission time and bandwidth requirements are significantly reduced. At the dispatch center, the received compressed data is reconstructed using an IVAE decoder to recover the original broadband oscillation signal. The reconstruction process requires the same network structure and parameters as the compression process; therefore, a corresponding IVAE model needs to be deployed at the dispatch center.

[0203] To evaluate the effectiveness of compression and reconstruction, corresponding performance metrics need to be defined. Key metrics include: reconstruction error (such as mean square error and signal-to-noise ratio), compression ratio, and processing time. Reconstruction error reflects the similarity between the reconstructed signal and the original signal and is the most important quality metric. Compression ratio reflects the efficiency of data compression, and processing time reflects the real-time performance of the algorithm.

[0204] Performance Verification of a Broadband Oscillation Detection Method Based on VMD and TLS-ESPRIT

[0205] To verify the performance of the broadband oscillation monitoring method based on the combination of VMD and TLS-ESPRIT proposed in this invention, a broadband oscillation signal as shown in equation (26) was designed in this example:

[0206] f(t)=cos(2π·2.5t)+41·cos(2π·50t)+

[0207] 161·cos(2π·93.4t)+641·cos(2π·288.6t)+

[0208] 2561·cos(2π·899.5t)+0.2·N(0,1)(26)

[0209] Considering the wide frequency distribution range of the broadband oscillation and the possibility of multiple oscillation components, and the fact that actual power grid measurement data usually contains a certain amount of noise, the oscillation signal is mainly composed of 2.5Hz, 50Hz, 93.4Hz, 288.6Hz, and 899.5Hz oscillation signal components and Gaussian white noise with an amplitude coefficient of 0.2. The signal oscillation amplitude is randomly generated, the sampling frequency is set to 2kHz, and the monitoring data time length is set to 1s.

[0210] To describe the monitoring accuracy of the method designed in this invention for multimodal broadband oscillation signals, the frequency relative error (FRE) and amplitude relative error (ARE) of the oscillation signal are used as evaluation indicators:

[0211]

[0212]

[0213] In the formula: f and a are the theoretical values ​​of frequency and amplitude, respectively; f′ and a′ are the monitored values ​​of frequency and amplitude.

[0214] To describe the effectiveness of the designed method under high-intensity random noise interference, this invention uses relative root mean square error (RRMSE), signal-to-noise ratio (SNR), and correlation coefficient as evaluation indicators:

[0215]

[0216] In the formula: S is the original signal without random noise interference; S′ is the effective signal after noise interference suppression; Cov(S,S′) is the covariance of S and S′, and Var(S) and Var(S′) are their variances. The smaller the RRMSE, the closer the signal after noise interference suppression is to the original signal; the larger the SNR, the lower the noise content and the better the noise reduction effect; ρ S The closer the value is to 1, the better the signal waveform reconstruction effect.

[0217] This indicates that the better the signal waveform reconstruction effect, the better.

[0218] According to the method of the present invention, firstly, the broadband oscillation signal designed by equation (26) is subjected to mode decomposition using VMD to remove the pure noise IMF, resulting in the following: Figure 5 The decomposition results are shown.

[0219] In one alternative implementation, data compression can be combined with other compression techniques to further improve compression efficiency. For example, lossless compression algorithms (such as Huffman coding, arithmetic coding, etc.) can be used to further compress the latent representation based on IVAE compression. This cascaded compression method can further improve the compression ratio while ensuring reconstruction quality.

[0220] In another alternative implementation, an adaptive compression strategy can be employed, dynamically adjusting compression parameters based on the complexity and importance of the signal. For signal segments containing significant oscillation information, a lower compression ratio can be used to ensure reconstruction accuracy; for relatively stable signal segments, a higher compression ratio can be used to save transmission bandwidth. This adaptive strategy requires establishing a mechanism for evaluating the importance of the signal, which can be based on the signal's energy distribution, spectral characteristics, or expert knowledge.

[0221] It should be noted that the entire technical solution organically combines three core technologies: VMD preprocessing, adaptive order determination TLS-ESPRIT parameter identification, and IVAE data compression, forming a complete broadband oscillation monitoring and data transmission solution. VMD effectively suppresses noise interference and improves signal quality; the adaptive order determination strategy avoids the subjectivity of manual order determination and improves the automation and accuracy of parameter identification; IVAE data compression solves the bandwidth limitation problem under high-speed sampling, making wide-area monitoring of broadband oscillations possible.

[0222] From the perspective of technological innovation, the main contributions of this invention are: proposing an adaptive order determination method based on the singular value adjacency growth ratio, which solves the problem that traditional TLS-ESPRIT requires manual parameter setting; introducing an improved variational autoencoder into the field of broadband oscillation data compression, achieving efficient data compression and accurate signal reconstruction; and establishing a complete technical chain from signal preprocessing to parameter identification to data transmission, providing a systematic solution for broadband oscillation monitoring of new energy power grids.

[0223] From an engineering application perspective, this invention has good practicality and promotional value. The algorithm's computational complexity is moderate, meeting the requirements of real-time monitoring; data compression technology effectively solves the limitation of communication bandwidth, reducing system construction and maintenance costs; adaptive characteristics enhance the algorithm's robustness and reduce the need for manual intervention. These advantages make this invention particularly suitable for broadband oscillation monitoring applications in large-scale new energy power grids.

[0224] In a further alternative embodiment, the invention can be extended to multi-point collaborative monitoring applications. By deploying multiple monitoring devices at key nodes of the power grid, spatial location and propagation path tracing of broadband oscillations can be achieved. After processing by the present invention, the data from each monitoring point can be comprehensively analyzed at the dispatch center to identify the source and scope of the oscillations, providing more accurate information support for oscillation suppression.

[0225] In another alternative implementation, the present invention can also be integrated with other monitoring and control systems of the power system to form a more comprehensive power grid security protection system. For example, it can be linked with a relay protection system to automatically trigger protection actions when severe broadband oscillations are detected; it can be combined with a load forecasting system to analyze the impact of load changes on broadband oscillations; and it can be coordinated with a renewable energy power forecasting system to assess the impact of renewable energy output fluctuations on power grid stability.

[0226] From a development trend perspective, with the digital transformation of power systems and the development of artificial intelligence technology, the technical approach adopted in this invention has promising prospects. Variational mode decomposition, as an emerging signal processing technology, has unique advantages in the analysis of non-stationary signals; adaptive algorithms can reduce manual intervention and improve the intelligence level of the system; deep learning technology demonstrates powerful capabilities in data compression and pattern recognition. Further development of these technologies will provide more advanced technical means for broadband oscillation monitoring.

[0227] In summary, this invention provides an efficient, accurate, and practical solution for monitoring broadband oscillations in power systems through technological innovation and system integration. This solution not only addresses key issues in current technologies but also lays a solid foundation for future technological development. As the proportion of new energy sources in power systems continues to increase, the broadband oscillation problem will become more prominent, further enhancing the technological value and application prospects of this invention.

[0228] Example 3

[0229] Reference Figures 3-7 This is the third embodiment of the present invention.

[0230] Depend on Figure 5 It can be seen that VMD can remove the noise-dominated IMF after decomposing the broadband oscillation signal designed by equation (26), and obtain five IMF mode components related to the frequency of the oscillation signal. Spectral analysis of each IMF mode shows that the original oscillation signals of 2.5Hz, 50Hz, 93.4Hz, 288.6Hz, and 899.5Hz are all distributed within the frequency bands of each IMF mode. To further accurately realize broadband oscillation monitoring, this invention uses TLS-ESPRIT to identify the oscillation information of each IMF component. The identification results are compared with the original signal as follows. Figure 6 As shown in Table 1.

[0231] Table 1 Evaluation of TLS-ESPRIT identification effectiveness of each IMF

[0232]

[0233] pass Figure 3 As shown in Table 1, using VMD to achieve mode decomposition can effectively eliminate pure noise IMFs, achieving partial noise reduction. However, noise still exists in the frequency bands contained in the effective IMFs. Building upon this, further using the TLS-ESPRIT algorithm to identify the oscillation characteristics of each IMF can further eliminate noise within the frequency bands and accurately obtain the frequency and amplitude of each IMF. By comparing the monitoring indicators with the original oscillation signal, the monitoring results show that both FRE and ARE are very small, verifying that the proposed method can accurately identify broadband oscillation information in each frequency band after suppressing noise interference.

[0234] To further verify the performance of the proposed method, the method of the present invention is compared and analyzed with the traditional wavelet transform method, Prony identification method, EMD and VMD. The monitoring effect of each method on the oscillation signal with high intensity random noise designed for this example is shown in Table 2 below.

[0235] Table 2 Comparison of the effects of different broadband oscillation signal monitoring methods

[0236] method RRMSE / (%) SNR / (dB) Correlation coefficient Wavelet Transform 19.12 15.57 0.9122 Prony 12.69 21.33 0.9488 EMD 8.73 23.91 0.9637 VMD 4.54 29.47 0.9882 Method of the present invention 1.24 36.86 0.9915

[0237] As shown in Table 2, the method of this invention has a high signal-to-noise ratio, a relatively small root mean square error, and a relatively large correlation coefficient. Compared with traditional methods, this invention achieves pure noise IMF suppression through the VMD method, which significantly improves the IMF signal-to-noise ratio (SNR) after mode decomposition. Furthermore, it uses the TLS-ESPRIT algorithm to achieve high-precision parameter identification and broadband oscillation feature extraction. While suppressing noise, it can accurately monitor broadband oscillation signal information, thus verifying that the method of this invention is more suitable for practical application in the monitoring of high-noise multimodal broadband oscillation signals.

[0238] Performance verification of wideband oscillatory data compression transmission based on IVAE

[0239] Upon detecting a broadband oscillation signal, the original broadband oscillation data needs to be encoded, compressed, and uploaded. The compressed data is then reconstructed at the dispatch center to accurately recover the broadband oscillation signal, facilitating subsequent analysis and processing by the dispatch master station. Considering that existing literature typically uses the traditional AE model for signal compression and reconstruction, to verify the advantages of the IVAE model used in this invention, a comparative analysis of reconstruction error and reconstruction time was conducted using both the traditional AE model and the IVAE model of this invention for original oscillation signals of different frequencies. The results are as follows. Figure 7 As shown in Table 3.

[0240] Table 3 Comparison of data compression performance at different oscillation frequencies

[0241]

[0242]

[0243] Depend on Figure 7It can be seen that the oscillation signal reconstructed by the IVAE model is significantly better than that of the AE model, and it is basically consistent with the original oscillation signal. Moreover, this advantage becomes more significant with the increase of oscillation frequency. Furthermore, as shown in Table 3, although the reconstruction time of the two models is not significantly different at different frequency bands, basically within 5 seconds, the mean square error of the reconstructed signal of the IVAE model is smaller than that of the AE model. In addition, the signal-to-noise ratio of the two models differs significantly under different oscillation signals. The signal-to-noise ratio of the reconstructed signal of the IVAE model is basically 1.5 to 2.2 times that of the AE model. This advantage shows a significant increasing trend with the increase of oscillation signal frequency. Therefore, this invention not only verifies that the IVAE model of this invention has better signal recovery capability than traditional methods, but also shows that the IVAE model proposed in this invention will have a significant advantage for the recovery of mid-to-high frequency oscillation signals, fully verifying the effectiveness of the data compression method designed in this invention.

[0244] Example 4

[0245] The above is a schematic scheme for a method of intelligent monitoring and data transmission of broadband oscillations in power systems. It should be noted that the technical solution of this system for intelligent monitoring and data transmission of broadband oscillations in power systems is based on the same concept as the technical solution of the aforementioned method. Details not described in detail in this embodiment can be found in the description of the aforementioned method.

[0246] This embodiment also provides a power system broadband oscillation intelligent monitoring and data transmission system, including:

[0247] The signal acquisition module is used to acquire the broadband oscillation signal of the power system and perform mode decomposition on the broadband oscillation signal using variational mode decomposition to obtain multiple intrinsic mode functions;

[0248] The signal preprocessing module is used to remove noise-dominant components from multiple intrinsic mode functions and retain effective mode components.

[0249] The adaptive algorithm module is used to construct the adaptive order-determining TLS-ESPRIT algorithm, which adaptively determines the order of the signal subspace by the singular value adjacency growth ratio;

[0250] The parameter identification module is used to input the effective modal components into the adaptive order determination TLS-ESPRIT algorithm, identify the parameters of each modal component, and extract the frequency and amplitude information of the broadband oscillation.

[0251] This embodiment also provides an electronic device suitable for intelligent monitoring and data transmission of broadband oscillations in power systems, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the method for intelligent monitoring and data transmission of broadband oscillations in power systems as proposed in the above embodiment.

[0252] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, it implements the method for intelligent monitoring and data transmission of broadband oscillations in power systems as proposed in the above embodiments.

[0253] The storage medium proposed in this embodiment and the method for realizing intelligent monitoring and data transmission of broadband oscillations in power systems proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0254] Based on the above description of the implementation methods, those skilled in the art will clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0255] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for intelligent monitoring and data transmission of broadband oscillations in power systems, characterized in that: This includes acquiring a broadband oscillation signal from a power system and performing mode decomposition on the broadband oscillation signal using variational mode decomposition to obtain multiple intrinsic mode functions; The noise-dominant components in the multiple intrinsic mode functions are removed, and the effective mode components are retained; An adaptive order-determining TLS-ESPRIT algorithm is constructed to adaptively determine the order of the signal subspace by the singular value adjacency growth ratio. The effective modal components are input into the adaptive order-determining TLS-ESPRIT algorithm to identify parameters of each modal component and extract the frequency and amplitude information of the broadband oscillation.

2. The method for intelligent monitoring and data transmission of broadband oscillations in power systems as described in claim 1, characterized in that: The removal of noise-dominant components from the plurality of intrinsic mode functions includes: Spectral analysis is performed on the multiple intrinsic mode functions; pure noise-dominated intrinsic mode functions are identified based on spectral characteristics; and non-noise-dominated intrinsic mode functions are retained as effective mode components.

3. The method for intelligent monitoring and data transmission of broadband oscillations in power systems as described in claim 2, characterized in that: The determination of the signal subspace order based on the singular value adjacency growth ratio includes: Construct a Hankel matrix and perform singular value decomposition; calculate the adjacent growth ratio based on the singular value sequence; determine the order of the signal subspace based on the minimum value of the adjacent growth ratio to achieve adaptive partitioning of the signal subspace and noise subspace.

4. The method for intelligent monitoring and data transmission of broadband oscillations in power systems as described in claim 3, characterized in that: Inputting the effective modal components into the adaptive order-determining TLS-ESPRIT algorithm includes: A Hankel matrix is ​​constructed for the effective modal components and singular value decomposition is performed; the signal subspace and noise subspace are divided based on the determined signal subspace order; the eigenvalues ​​are solved by the overall least squares method, and the frequency, damping, amplitude and phase parameters of the oscillation signal are calculated.

5. The method for intelligent monitoring and data transmission of broadband oscillations in power systems as described in claim 4, characterized in that: The method further includes: Determine whether a wideband oscillation signal is detected; if a wideband oscillation signal is detected, initiate the data transmission processing procedure to perform data compression processing on the wideband oscillation signal.

6. The method for intelligent monitoring and data transmission of broadband oscillations in power systems as described in claim 5, characterized in that: The data compression processing of the broadband oscillation signal includes: An improved variational autoencoder is established, comprising an encoder and a decoder; the encoder encodes a wideband oscillation signal into latent variables and label values; the latent variables and label values ​​are concatenated and then input into the decoder to reconstruct the signal.

7. The method for intelligent monitoring and data transmission of broadband oscillations in power systems as described in claim 6, characterized in that: The construction of the improved variational autoencoder includes: A convolutional neural network is used as the encoder, and an exponential linear unit is used as the activation function. A label value processing module is added between the encoder and the decoder to take the label data as additional condition input. A deconvolutional neural network is used as the decoder, and the loss function is calculated through cross-entropy to ensure the authenticity of the output samples. The posterior distribution is made close to the prior distribution by minimizing the KL divergence.

8. A power system broadband oscillation intelligent monitoring and data transmission system, based on the power system broadband oscillation intelligent monitoring and data transmission method according to any one of claims 1 to 7, characterized in that: It also includes a signal acquisition module, used to acquire the broadband oscillation signal of the power system, and to perform mode decomposition on the broadband oscillation signal using variational mode decomposition to obtain multiple intrinsic mode functions; The signal preprocessing module is used to remove the noise-dominant components from the multiple intrinsic mode functions and retain the effective mode components. The adaptive algorithm module is used to construct the adaptive order-determining TLS-ESPRIT algorithm, which adaptively determines the order of the signal subspace by the singular value adjacency growth ratio; The parameter identification module is used to input the effective modal components into the adaptive order determination TLS-ESPRIT algorithm, identify the parameters of each modal component, and extract the frequency and amplitude information of the broadband oscillation.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the power system broadband oscillation intelligent monitoring and data transmission method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the power system broadband oscillation intelligent monitoring and data transmission method according to any one of claims 1 to 7.

Images