A method for wideband oscillation disturbance source hierarchical positioning based on compressed sensing and ISTA
By combining compressed sensing and the ISTA algorithm with deep learning, a multi-layer neural network is constructed for hierarchical signal localization, which solves the problem of accuracy in locating broadband oscillation disturbance sources in power systems and achieves fast and accurate oscillation source detection and localization.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-11-16
- Publication Date
- 2026-03-24
AI Technical Summary
Existing broadband oscillation disturbance source localization methods have low accuracy in power system localization tasks, mainly because the master station cannot obtain complete oscillation information and is limited by the Nyquist sampling theorem.
Compressed sensing technology is used for subsampling. Combined with the ISTA algorithm and deep learning methods, a multi-layer neural network is constructed to restore and locate the signal. This includes collecting the power of the power grid generator, extracting typical oscillation feature sub-sequences, and constructing a disturbance region and disturbance source localization network to achieve hierarchical signal localization.
It achieves precise location of broadband oscillation disturbance sources, solves the accuracy problem of positioning tasks, saves computing resources, reduces channel burden, and improves positioning speed and accuracy.
Smart Images
Figure CN116108382B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power system oscillation disturbance source hierarchical localization technology, and in particular to a broadband oscillation disturbance source hierarchical localization method using compressed sensing and ISTA. Background Technology
[0002] Under the development trend of "high-voltage and high-efficiency" power systems, the problem of broadband oscillations is becoming increasingly prominent. The interaction between power electronic equipment and the power grid exhibits strong time-varying and nonlinear characteristics, making accurate location of oscillation disturbance sources difficult. In power systems, limited by the Nyquist sampling theorem, the master station can only obtain oscillation signals within half the frequency range of the transmission frequency from the slave station. Because the broadband oscillation problem covers a wide frequency band, the master station cannot obtain complete oscillation information, significantly affecting the accuracy of the location task.
[0003] Existing broadband oscillation disturbance source localization methods are mostly based on the wide area measurement system (WAMS) of the PMU, which uploads the collected electrical quantities to a remote master station to identify the parameters and realize the disturbance source localization.
[0004] In power systems, limited by the Nyquist sampling theorem, the master station can only obtain oscillation signals within half the frequency range of the transmitted frequency from the slave station. Because the broadband oscillation problem covers a wide frequency band, the master station cannot obtain complete oscillation information, significantly impacting the accuracy of positioning tasks. Summary of the Invention
[0005] The purpose of this invention is to provide a hierarchical localization method for broadband oscillation disturbance sources using compressed sensing and ISTA, aiming to solve the problem of low accuracy in localization tasks of existing broadband oscillation disturbance source localization methods.
[0006] To achieve the above objectives, this invention provides a hierarchical localization method for broadband oscillation disturbance sources using compressed sensing and ISTA, comprising the following steps:
[0007] The system collects the power of the grid generators and stores the generator power as a time series.
[0008] Based on the state of the typical oscillation feature subsequence detection system extracted from the time series, the criterion result is obtained;
[0009] The compressed signal is obtained by subsampling the time series using compressed sensing technology;
[0010] The compressed signal and the criterion result are input into the disturbance area positioning network, and the disturbance area unit number is output.
[0011] The unit number in the disturbance area is input into the ISTA recovery network for recovery, and the oscillation recovery signal is obtained.
[0012] The oscillation recovery signal is input into the disturbance source location network for location, and the serial number of the broadband oscillation disturbance source unit is obtained.
[0013] The power of the power grid generator in the data acquisition system includes:
[0014] The power of the grid generator in the system is collected by the PMU device.
[0015] The state of the detection system for extracting typical oscillation feature subsequences based on the time series, and obtaining the criterion result, includes:
[0016] Based on the time series, typical oscillatory feature subsequences are extracted using the Shapelet algorithm, and these subsequences are then input into a support vector machine classifier to detect the state of the system, thereby obtaining the criterion result.
[0017] The step of inputting the unit number of the disturbed area into the ISTA recovery network for recovery to obtain the oscillation recovery signal includes:
[0018] An ISTA restoration network was constructed using deep learning methods. The network was trained and its parameters were updated to obtain the optimal restoration network.
[0019] The unit number in the disturbed area is input into the optimal restoration network for restoration, and the oscillation restoration signal is obtained.
[0020] The step of inputting the compressed signal and the criterion result into the disturbance area positioning network and outputting the disturbance area unit number includes:
[0021] A perturbation region localization network is constructed, and the sigmoid function is selected as the activation function of the network output layer of the perturbation region localization network. The improved binary classification cross-entropy function is used as the loss function of the perturbation region localization network.
[0022] The compressed signal and the criterion result are input into the disturbance area positioning network, and the disturbance area unit number is output.
[0023] The step of inputting the oscillation recovery signal into the disturbance source location network for location to obtain the broadband oscillation disturbance source unit serial number includes:
[0024] A disturbance source localization network is constructed based on a convolutional neural network;
[0025] The oscillation recovery signal is input into the disturbance source location network and the data is convolved by a convolution kernel to obtain the broadband oscillation disturbance source unit number.
[0026] This invention discloses a hierarchical localization method for broadband oscillation disturbance sources using compressed sensing and ISTA. The method involves collecting the power output of the grid generators and storing the power output as a time series; extracting typical oscillation feature subsequences from the time series to detect the system's state and obtaining a criterion result; subsampling the time series using compressed sensing technology to obtain a compressed signal; inputting the compressed signal and the criterion result into a disturbance area localization network to output the disturbance area unit number; inputting the disturbance area unit number into an ISTA recovery network for recovery to obtain an oscillation recovery signal; and inputting the oscillation recovery signal into a disturbance source localization network for localization to obtain the broadband oscillation disturbance source unit number. This method achieves precise localization of oscillation disturbance sources and solves the problem of low accuracy in existing broadband oscillation disturbance source localization methods. Attached Figure Description
[0027] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the 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.
[0028] Figure 1 This is a framework diagram of a broadband oscillation disturbance source hierarchical localization method provided by the present invention, which uses compressed sensing and ISTA.
[0029] Figure 2 This is a diagram of the ISTA reconstruction network framework.
[0030] Figure 3 This is a flowchart of a typical oscillator sequence extraction process.
[0031] Figure 4 This is a diagram of a hierarchical localization network for broadband oscillation disturbance sources.
[0032] Figure 5 This is a schematic diagram of the hierarchical location of broadband oscillation disturbance sources.
[0033] Figure 6 This is a flowchart of a broadband oscillation disturbance source hierarchical localization method provided by the present invention, which uses compressed sensing and ISTA. Detailed Implementation
[0034] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.
[0035] Please see Figures 1 to 6 This invention provides a hierarchical localization method for broadband oscillation disturbance sources using compressed sensing and ISTA, comprising the following steps:
[0036] S1 collects the power of the grid generators in the system and stores the generator power as a time series.
[0037] Specifically, the power of the grid generator in the system is collected through the PMU device.
[0038] S2 extracts the state of the typical oscillation feature subsequence detection system based on the time series and obtains the criterion result;
[0039] Specifically, based on the time series, typical oscillatory feature subsequences are extracted using the Shapelet algorithm, and the typical oscillatory feature subsequences are input into a support vector machine classifier to detect the state of the system and obtain the criterion result.
[0040] The specific principle for extracting typical feature subsequences is as follows: The remote substation obtains the current generator power through the PMU device and organizes it into a time series. Due to the high sampling frequency of the PMU, this time series can be approximated as continuous. The randomly generated pre-selected subsequence set is traversed, and the information gain of the segmentation strategy based on each subsequence is calculated:
[0041] I(D)=-p(A)log(p(A))-p(B)log(p(B))
[0042]
[0043] In the formula: D is a time series dataset; A and B are two classes of D; p(A) and p(B) are the proportions of objects in the corresponding classes; D1 and D2 are subsets classified according to the segmentation strategy; f(D1) and f(D2) are the weights of objects in D1 and D2, respectively; I(D) and These represent the entropy of dataset D and the total entropy of D after the segmentation strategy, respectively. During the traversal of the pre-selected subsequence set, the information gain of the segmentation strategy is iteratively updated, gradually realizing feature extraction of the oscillating sequences. The extracted oscillating features are then fed into an SVM classifier to more accurately obtain the current unit status.
[0044] S3 subsamples the time series based on compressed sensing technology to obtain a compressed signal;
[0045] Compressed sensing theory states that if a signal has a sparse representation in a certain transform domain, the signal can be sampled at a frequency much lower than the Nyquist sampling frequency and the signal fidelity can be maintained during reconstruction. If there exists an N-dimensional original signal... The M-dimensional compressed signal y can then be obtained by sampling through the observation matrix Φ, where ε = M / N is the signal compression ratio. The sampling process is as follows:
[0046]
[0047] In the above formula Since N >> M, the signal reconstruction problem is essentially solving a system of underdetermined equations. Under the condition that the observation matrix satisfies the Restricted Isometry Property (RIP), the system of equations has a feasible solution, enabling accurate signal reconstruction. In power systems, the electrical forces generated by the interaction between power electronic devices and the power grid are a continuous sequence in the time domain and have a sparse representation in the frequency domain, satisfying the conditions for compressed sensing.
[0048] S4 inputs the compressed signal and the criterion result into the disturbance area positioning network and outputs the disturbance area unit number;
[0049] Specifically, a disturbance area positioning network is constructed, the sigmoid function is selected as the activation function of the network output layer of the disturbance area positioning network, and the improved binary cross-entropy function is used as the loss function of the disturbance area positioning network; the compressed signal and the criterion result are input into the disturbance area positioning network, and the disturbance area unit number is output.
[0050] S5 inputs the unit number of the disturbance area into the ISTA recovery network for recovery, and obtains the oscillation recovery signal;
[0051] Specifically, an ISTA restoration network is constructed using deep learning methods. The ISTA restoration network is trained and its parameters are updated through a neural network to obtain the optimal restoration network. The unit number of the disturbance area is input into the optimal restoration network for restoration to obtain the oscillation restoration signal.
[0052] The ISTA algorithm belongs to the convex optimization algorithm family and is a classic algorithm for solving the inverse problem of sparse coding, often used in compressed sensing reconstruction. The traditional l1-norm compressed sensing reconstruction model is as follows:
[0053]
[0054] In the formula: Ψx is the representation of x in the sparse domain. In ISTA, given low-dimensional compressed data y, equation (3) is solved through gradient iteration and signal iteration. The result of the k-th gradient iteration is r (k) The result of the kth signal iteration x (k) The expressions are as shown in equation (3) and equation (4).
[0055] r (k) =x (k-1) -ρΦ T (Φx(k-1) -y) (3)
[0056]
[0057] In the above formula, k is the number of ISTA iterations; ρ is the iteration step size; and ||Ψx||1 is a regularization penalty term to ensure the sparsity of the reconstructed signal.
[0058] Based on equations (4) and (5), this invention employs a deep learning method to construct an ISTA restoration network, updating parameters through neural network training to adapt to the problem of broadband oscillation signal restoration. Compared to the original ISTA, ρ and λ in equations (4) and (5) are trainable parameters, updated iteratively by the neural network. Compared to manually set transformation Ψ, this invention utilizes the strong nonlinear representation capability of convolutional neural networks, replacing the Ψ transformation with a convolutional layer of learnable parameters, denoted as χ(·).
[0059] If the N-dimensional initial signal is [x1, x2, ..., x...], then... N-1 ,x N Let r be a normally distributed random variable. (k) With χ(r) (k) ) are the average values of the restored signal x and χ(x), respectively. Then, in equation (5), ||xr (k) ||2 and α||χ(x)-χ(r) (k) If ||2 is approximated, and α is a constant, then equation (4) is equivalent to equation (5):
[0060]
[0061] In the formula: θ=λ / α, This is the inverse change of χ(·); It is a soft thresholding function. The threshold is a soft threshold, and its expression is shown in equation (6).
[0062]
[0063] An ISTA restoration network is built based on equations (4) and (6), and the specific framework is as follows: Figure 2 .Depend on Figure 2 As can be seen, the restoration network consists of two parts: initial generation and iterative optimization. Since the compressed signal cannot intuitively represent the characteristics of the original signal, the initialization part uses multiple fully connected layers. The network adaptively trains the parameters to learn the most suitable conditions for use as input to the iterative optimization part. The iterative optimization part is built based on equation (6), where χ(·) and All are constructed from convolutional networks, and the parameters are adaptively learned to achieve corresponding changes.
[0064] S6 inputs the oscillation recovery signal into the disturbance source location network for location, and obtains the broadband oscillation disturbance source unit serial number.
[0065] Specifically, a disturbance source localization network is constructed based on a convolutional neural network; the oscillation recovery signal is input into the disturbance source localization network and data convolution is performed through a convolution kernel to obtain the unit number of the broadband oscillation disturbance source.
[0066] The method for determining the unique oscillation source of broadband oscillation is as follows: construct m one-dimensional convolutional binary classification networks with identical parameters for each of the m generating units, input the restored oscillation signal of each generating unit into the classification network, judge each generating unit synchronously, and obtain the unique oscillation source by combining the judgment results of all generating units.
[0067] In the problem of locating broadband oscillation disturbance sources, compressed oscillation signals have low dimensionality, resulting in fewer network parameters and shorter inference time. However, due to the strong nonlinearity of compressed signals, they are difficult to fit, leading to poor positioning accuracy. Positioning models based on global oscillation recovery signals achieve considerable accuracy, but the large number of parameters severely impacts inference speed. This invention combines the characteristics of both models, proposing a hierarchical positioning network for broadband oscillation sources. This network inputs compressed and recovered signals hierarchically, ensuring positioning accuracy while maintaining overall network inference speed.
[0068] This invention constructs an oscillation source area localization network, using compressed electrical quantity data of oscillating units as input, and outputs the probability that each unit is an oscillation source, thus transforming the localization problem into a regression prediction problem. The network output layer selects the sigmoid function as the activation function to ensure that the output values of each unit are compressed between 0 and 1 and are relatively independent. The expression for the sigmoid function is shown in equation (7).
[0069]
[0070] In the design of the loss function, since the goal of this network is to locate the disturbance area, the ideal output is the disturbance area containing the oscillating source unit. Therefore, the network has a certain tolerance for erroneous predictions of non-oscillating source units; conversely, erroneous predictions of oscillating source units are unacceptable. After analyzing and comparing several common loss functions, this invention improves the binary classification cross-entropy function as the loss function of the disturbance area location network. The modified loss function is shown in equation (8).
[0071]
[0072] In the above formula, y represents the actual label; The result is the prediction result; α and β are constants, set manually. When α > 1 and β < 1, the penalty for misjudging non-oscillating sources can be reduced and the penalty for misjudging oscillating sources can be increased.
[0073] The precise positioning network is based on a convolutional neural network. By performing convolution operations on the data through convolution kernels, it can effectively extract data features. The convolution process is shown in equation (9).
[0074]
[0075] In the above equation, h(x) and f(x) are bounded and integrable functions. In convolutional neural networks (CNNs), data is locally connected through convolutional kernels of a specified size, thus enabling better perception of local features of the data while reducing the amount of parameter training. For time series problems, CNNs can effectively perceive local features of the sequence and quickly achieve data dimensionality reduction.
[0076] If the oscillation area contains m generator units, the positioning process ends only when one generator unit is identified as the source of the oscillation disturbance, and the remaining generator units are all non-disturbance sources; otherwise, the positioning network is restarted. If 'a' represents the accuracy of a single judgment, the probability χ(a) of the disturbance source not being unique in the judgment result and the probability of positioning error can be obtained. The expressions are as shown in equation (10) and equation (11).
[0077] χ(a)=1-σ[(m-1)(1-a) 2 -a 2 ]×a m-2 -(1-σ)(m-1)(1-a)a m-1 (10)
[0078]
[0079] In the above formula, σ represents the accuracy of the disturbance region division. By setting the upper limit of the number of model restarts to T, the positioning success rate δ(a) can be obtained as shown in formula (12).
[0080]
[0081] This restart mechanism effectively improves the accuracy of oscillation disturbance source localization. By setting the disturbance region division range and the upper limit of the number of model restarts, the model localization speed and accuracy can be adjusted to meet the needs of different scenarios in actual engineering.
[0082] This invention discloses a hierarchical localization method for broadband oscillation disturbance sources based on compressed sensing and ISTA. Its key features include: proposing a localization initiation criterion for oscillation disturbance sources; employing compressed sensing technology to achieve high compression ratio sampling of oscillation signals, circumventing the limitations of the Nyquist sampling theorem while preserving the signal's frequency domain characteristics; and proposing a hierarchical oscillation source localization network based on ISTA, incorporating a restart mechanism to automatically restart the localization network when the localization result is not unique. The beneficial effects of this invention are: enabling rapid and accurate detection of broadband oscillation phenomena with strong time-varying and nonlinear characteristics, saving computational resources; allowing current transmission channels to transmit compressed signals containing complete oscillation information, reducing channel load; and enabling the master station to receive compressed data and input it into the hierarchical network to obtain accurate localization results.
[0083] The above-disclosed embodiments are merely preferred embodiments of the broadband oscillation disturbance source hierarchical localization method of compressed sensing and ISTA according to the present invention. Of course, they should not be construed as limiting the scope of the present invention. Those skilled in the art can understand that all or part of the above-described embodiments can be implemented, and equivalent changes made in accordance with the claims of the present invention are still within the scope of the invention.
Claims
1. A hierarchical localization method for broadband oscillation disturbance sources using compressed sensing and ISTA, characterized in that, Includes the following steps: The system collects the power of the grid generators and stores the generator power as a time series. Based on the state of the typical oscillation feature subsequence detection system extracted from the time series, the criterion result is obtained; The compressed signal is obtained by subsampling the time series using compressed sensing technology; The compressed signal and the criterion result are input into the disturbance area positioning network, and the disturbance area unit number is output. The unit number in the disturbance area is input into the ISTA recovery network for recovery, and the oscillation recovery signal is obtained. The oscillation recovery signal is input into the disturbance source location network for location, and the serial number of the broadband oscillation disturbance source unit is obtained.
2. The broadband oscillation disturbance source hierarchical localization method of compressed sensing and ISTA as described in claim 1, characterized in that, The power of the power grid generator in the data acquisition system includes: The power of the grid generator in the system is collected by the PMU device.
3. The broadband oscillation disturbance source hierarchical localization method of compressed sensing and ISTA as described in claim 2, characterized in that, The state of the detection system based on the time series extraction of typical oscillation feature subsequences yields the following criterion results: Based on the time series, typical oscillatory feature subsequences are extracted using the Shapelet algorithm, and these subsequences are then input into a support vector machine classifier to detect the state of the system, thereby obtaining the criterion result.
4. The broadband oscillation disturbance source hierarchical localization method of compressed sensing and ISTA as described in claim 3, characterized in that, The step of inputting the unit number of the disturbed area into the ISTA recovery network for recovery to obtain the oscillation recovery signal includes: An ISTA restoration network was constructed using deep learning methods. The network was trained and its parameters were updated to obtain the optimal restoration network. The unit number in the disturbed area is input into the optimal restoration network for restoration, and the oscillation restoration signal is obtained.
5. The broadband oscillation disturbance source hierarchical localization method of compressed sensing and ISTA as described in claim 4, characterized in that, The step of inputting the compressed signal and the criterion result into the disturbance area positioning network and outputting the disturbance area unit number includes: A perturbation region localization network is constructed, and the sigmoid function is selected as the activation function of the network output layer of the perturbation region localization network. The improved binary classification cross-entropy function is used as the loss function of the perturbation region localization network. The compressed signal and the criterion result are input into the disturbance area positioning network, and the disturbance area unit number is output.
6. The broadband oscillation disturbance source hierarchical localization method of compressed sensing and ISTA as described in claim 5, characterized in that, The step of inputting the oscillation recovery signal into the disturbance source location network for location, and obtaining the broadband oscillation disturbance source unit serial number, includes: A disturbance source localization network is constructed based on a convolutional neural network; The oscillation recovery signal is input into the disturbance source location network and the data is convolved by a convolution kernel to obtain the broadband oscillation disturbance source unit number.