A method for identifying sub / supersynchronous oscillation frequency based on pre-sequence short window data reconstruction and deep learning

Abstract

The application discloses a method for identifying sub / super synchronous oscillation frequency based on pre-sequence short window data reconstruction and deep learning, which takes the voltage, current, active power and other multi-channel time sequence data collected in the early stage of oscillation as the input of the sub / super synchronous oscillation frequency identification network, firstly reconstructs the missing sampling points through an encoder-decoder with a mask matrix, then extracts the oscillation dynamic characteristics by using a neural network feature extractor, and finally outputs the sub / super synchronous oscillation frequency through a predictor; the method can complete the frequency rapid identification under the condition of a short time window, and improve the identification accuracy and robustness under the condition of missing field data and noise interference.

Description

Technical Field

[0001] This invention belongs to the field of power system oscillation monitoring and artificial intelligence identification technology. More specifically, it relates to a method for identifying sub / supersynchronous oscillation frequencies based on preceding short-window data reconstruction and deep learning. Background Technology

[0002] With the large-scale integration of new energy power generation equipment, power electronic converters, flexible DC devices, and a high proportion of power electronic loads into the power system, the dynamic characteristics of the power grid have undergone significant changes. Complex coupling exists between the control loops of new energy unit converters, phase-locked loops, grid impedance, line parameters, and operating modes. In scenarios such as weak grids, high output, and concentrated integration of new energy sources, the system is prone to subsynchronous or supersynchronous oscillations. Accurately and quickly identifying the frequency of these subsynchronous / supersynchronous oscillations can provide a basis for oscillation type judgment, oscillation source location, control parameter adjustment, additional damping control, and optimization of operation control strategies for new energy power plants.

[0003] Existing oscillation frequency identification methods mainly include those based on Fourier transform, those based on parameter estimation, and those based on data-driven models. Methods based on discrete Fourier transform and interpolated Fourier transform are clear in principle and easy to implement, but they typically rely on long data windows. If the data window is too short, the frequency resolution is insufficient; if the data window is too long, it is difficult to meet the requirement of rapid identification in the early stages of oscillation. For subsynchronous / supersynchronous oscillations in renewable energy grid-connected systems, the oscillation may develop rapidly within hundreds of milliseconds; therefore, frequency analysis methods relying on long time windows cannot support control decisions in a timely manner.

[0004] Methods such as Prony, matrix bundle, ESPRIT, and variational mode decomposition can estimate oscillation frequency parameters to a certain extent. However, these methods are generally sensitive to model order, noise level, signal integrity, and coupling of multiple frequency components. In actual field applications, voltage, current, and power signals may simultaneously contain subsynchronous components, supersynchronous components, power frequency components, and noise components. Furthermore, data acquisition may encounter problems such as communication packet loss, measurement device malfunctions, sampling asynchrony, or local measurement point failures, leading to insufficient reliability and robustness of traditional parameter identification methods.

[0005] Existing data-driven methods can automatically extract nonlinear dynamic features from time-series electrical quantities, but most methods directly use raw measurement data as input and do not fully consider the problem of missing field oscillation data. At the same time, some methods only identify complete data or single oscillation components, which is difficult to adapt to engineering scenarios such as the coexistence of subsynchronous and supersynchronous components, missing data, noise interference, and rapid identification with short windows.

[0006] Therefore, it is necessary to propose a frequency identification method for subsynchronous / supersynchronous oscillation events in power systems. This method can quickly output the oscillation frequency in the early stage of oscillation by using only short-time window data of multiple electrical quantities such as voltage, current, and active power. This is achieved through data reconstruction, deep feature extraction, and frequency prediction, thereby improving the real-time performance, accuracy, and robustness of subsynchronous / supersynchronous oscillation parameter identification. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the prior art and provide a sub / supersynchronous oscillation frequency identification method based on pre-sequence short window data reconstruction and deep learning, which can quickly identify sub / supersynchronous oscillation frequencies and improve the identification accuracy and robustness under the conditions of missing field data and noise interference.

[0008] To achieve the above-mentioned objectives, this invention provides a method for identifying sub / supersynchronous oscillation frequencies based on preceding short-window data reconstruction and deep learning, characterized by comprising:

[0009] (1) Construct a sub / supersynchronous oscillation frequency identification network based on the reconstruction of preceding short window data;

[0010] (2) Collect subsynchronous / supersynchronous oscillation sample data and construct training samples;

[0011] (3) Train the sub / hypersynchronous oscillation frequency identification network using training samples until convergence;

[0012] (4) Use the trained sub / supersynchronous oscillation frequency identification network to identify oscillations.

[0013] The objective of this invention is achieved as follows:

[0014] This invention presents a sub / supersynchronous oscillation frequency identification method based on short-window data reconstruction and deep learning. It uses multi-channel time-series data (voltage, current, active power, etc.) acquired in the early stages of oscillation as input to the sub / supersynchronous oscillation frequency identification network. First, a masked encoder-decoder reconstructs the missing sampling points. Then, a neural network feature extractor extracts the dynamic features of the oscillation. Finally, a predictor outputs the sub / supersynchronous oscillation frequency. This method enables rapid frequency identification under short time window conditions and improves the accuracy and robustness of identification under conditions of missing field data and noise interference.

[0015] Meanwhile, the sub / supersynchronous oscillation frequency identification method based on preceding short window data reconstruction and deep learning in this invention also has the following beneficial effects:

[0016] (1) The present invention clearly distinguishes between valid sampling positions and missing sampling positions through a mask matrix, and uses an encoder-decoder reconstruction module to fill in only the missing positions, avoiding the direct discarding of incomplete samples or the use of zero filling or mean filling to destroy the oscillation dynamic characteristics, thereby improving the usability of incomplete waveform data recorded on site.

[0017] (2) The present invention constructs a feature extractor by using a convolutional layer, a residual block, a non-local feature extraction block and an effective channel attention layer. It simultaneously captures local waveform slope, periodic changes, long-distance time-series dependence and correlation between voltage, current and active power channels within a short window, thereby improving the accuracy of sub / supersynchronous oscillation frequency identification.

[0018] (3) The present invention directly outputs the continuous dominant frequency value through the predictor, avoiding the simplification of frequency identification into a fixed category classification problem, which is beneficial to adapting to scenarios where different operating points, different oscillation frequencies and sub / supersynchronous components coexist.

[0019] (4) In the online stage, the present invention only needs to input the preceding short window data after the oscillation starts to directly output the dominant frequency, which can provide more timely parameter basis for early oscillation alarm, oscillation source location and suppression control. Attached Figure Description

[0020] Figure 1 This is a flowchart of the sub / supersynchronous oscillation frequency identification method based on preceding short window data reconstruction and deep learning of the present invention;

[0021] Figure 2 This is the overall structure diagram of the sub / supersynchronous oscillation frequency identification network based on the reconstruction of preceding short window data in this invention;

[0022] Figure 3 yes Figure 1 The encoder structure diagram shown is shown below;

[0023] Figure 4 yes Figure 1 The decoder structure diagram shown below;

[0024] Figure 5 yes Figure 1 The feature extractor structure diagram is shown below;

[0025] Figure 6 yes Figure 1 The predictor structure diagram is shown below;

[0026] Figure 7 This is a simulation diagram verifying the sub / supersynchronous oscillation frequency identification network based on the reconstruction of preceding short window data in this invention. Detailed Implementation

[0027] The specific embodiments of the present invention will now be described with reference to the accompanying drawings to enable those skilled in the art to better understand the invention. It should be particularly noted that in the following description, detailed descriptions of known functions and designs that might obscure the main content of the invention will be omitted here.

[0028] Example

[0029] In this embodiment, as Figure 1 As shown, the present invention provides a method for identifying sub / supersynchronous oscillation frequencies based on preceding short-window data reconstruction and deep learning, comprising:

[0030] (1) Construct a sub / supersynchronous oscillation frequency identification network based on the reconstruction of preceding short window data;

[0031] In this embodiment, as Figure 2 As shown, the sub / supersynchronous oscillation frequency identification network reconstructed based on preceding short window data includes a reconstruction module consisting of an encoder and a decoder, a feature extractor, and a predictor.

[0032] The reconstruction module includes an encoder E and a decoder D;

[0033] like Figure 3 As shown, the encoder consists of three convolutional layers and max-pooling layers connected alternately. The first convolutional layer has 3 input channels, 16 output channels, a kernel length of 3, and padding of 1. The first max-pooling layer has a kernel length of 2 and a stride of 2. The second convolutional layer has 16 input channels, 8 output channels, a kernel length of 3, and padding of 1. The second max-pooling layer has a kernel length of 2 and a stride of 2. The third convolutional layer has 8 input channels, 3 output channels, a kernel length of 3, and padding of 1. The third max-pooling layer has a kernel length of 2 and a stride of 2.

[0034] like Figure 4 As shown, the decoder consists of four convolutional layers and upsampling layers connected in series with adjacent convolutional layers. The first convolutional layer has 3 input channels, 16 output channels, a kernel length of 3, and 1 padding. The second convolutional layer has 16 input channels, 32 output channels, a kernel length of 3, and 1 padding. The third convolutional layer has 32 input channels, 16 output channels, a kernel length of 3, and 1 padding. The fourth convolutional layer has 16 input channels, 3 output channels, a kernel length of 3, and 1 padding. All upsampling layers have a scale factor of 2 and use nearest-neighbor interpolation.

[0035] like Figure 5As shown, the feature extractor consists of, in sequence, a first one-dimensional convolutional layer with 3 input channels, 16 output channels, a kernel length of 3, padding of 1, and a stride of 1; a first batch normalization layer with 16 channels; an effective channel attention layer with 16 channels; a first residual block with 16 input channels, 8 output channels, and a stride of 2; a non-local feature extraction block with 8 channels; a second batch normalization layer with 8 channels; a second residual block with 8 input channels, 4 output channels, and a stride of 1; a third residual block with 4 input channels, 1 output channel, and a stride of 1; a flattening layer; and a fully connected feature layer with a 128-dimensional deep oscillatory feature output. The fully connected feature layer is followed by a Dropout layer with a dropout rate of 0.1 and a ReLU activation function.

[0036] like Figure 6 As shown, the predictor consists of a first fully connected predictive layer, a ReLU activation function, a Dropout layer, and a second fully connected predictive layer connected in series. The first fully connected predictive layer maps 128-dimensional deep oscillation features to 64-dimensional latent features. The dropout rate of the Dropout layer is 0.1. The second fully connected predictive layer maps the 64-dimensional latent features to a 1-dimensional frequency output, which is the predicted value of the dominant frequency of the sub / supersynchronous oscillation corresponding to the short window sample.

[0037] (2) Collect subsynchronous / supersynchronous oscillation sample data and construct training samples;

[0038] When a power system experiences subsynchronous / supersynchronous oscillations, acquire multi-channel electrical quantity data, including voltage, during the early stages of the oscillation. Current Active power reactive power ,frequency and phase angle ;

[0039] set up In each acquisition phase, the window length is set to [value]. , No. The multi-channel electrical quantity data collected in each acquisition phase are represented as follows:

[0040] , ;

[0041] In this embodiment, preferably When the sampling frequency is 2kHz and a short window of 0.2s after oscillation is used as the preceding window, T=400.

[0042] Therefore, a data acquisition step of length is obtained in each acquisition stage. Short window data To reduce the impact of different operating points and different measurement amplitudes on training, The electrical quantity data of each channel are normalized.

[0043] In actual field operations, oscillation records may contain missing sampling points. Reasons for this include communication packet loss, sampling device malfunction, storage failure, synchronization errors, or partial measurement point failure. To address this issue, according to... Construct a mask matrix from the electrical quantity data of each channel:

[0044]

[0045] in, For the first Line 1 Column elements, when For effective data sampling, ;otherwise, ;

[0046] Will Acquired in each collection stage and the corresponding mask matrix This constitutes the training sample set.

[0047] (3) Train the sub / hypersynchronous oscillation frequency identification network using training samples until convergence;

[0048] (3.1) From Multi-channel electrical quantity data Selected from One data point is used as a batch of training data;

[0049] (3.2) will indivual With the corresponding The input is fed into a reconstruction module consisting of an encoder and a decoder, and is mapped into latent features by the encoder E. :

[0050] ;

[0051] Then the latent features are decoded by decoder D. Reconstruct complete electrical quantity timing data :

[0052] ;

[0053] Input data and reconstructing data According to the mask matrix To merge:

[0054]

[0055] In the formula, This indicates element-wise multiplication. This represents the merged electrical data;

[0056] (3.3) The merged electrical data The input is fed into the feature extractor F to obtain deep oscillation features:

[0057] ;

[0058] (3.4) Deep oscillation characteristics The input is fed into the predictor to predict the dominant frequency of the subsynchronous / supersynchronous oscillation:

[0059] ;

[0060] (3.5) Calculate the loss function value after this round of training;

[0061] ;

[0062] ;

[0063] ;

[0064] in, Let represent the dominant frequency label of the i-th training sample. Reconstruct loss weights for the data;

[0065] (3.6) Based on the loss function value The parameters of the subsynchronous / supersynchronous oscillation frequency identification network are updated using the gradient descent method, and then the process returns to step (3.1) to perform the next round of training until the subsynchronous / supersynchronous oscillation frequency identification network converges.

[0066] (4) Use the trained sub / supersynchronous oscillation frequency identification network to identify oscillations.

[0067] In this embodiment, a set of short-window data is collected according to step (2). For short window data After preprocessing, a mask matrix is ​​constructed. Then the short window data and the corresponding mask matrix The input is fed into the trained sub / hypersynchronous oscillation frequency identification network to predict the dominant sub / hypersynchronous oscillation frequency;

[0068] To verify the effectiveness of the method of the present invention, in this example, a subsynchronous / supersynchronous oscillation simulation sample is constructed in a grid-connected system containing new energy units, as shown below. Figure 7As shown, voltage, current, and active power are selected as the model input channels, with a sampling frequency of 2kHz. Data from 0.2s after oscillation start-up is used as the input samples. Therefore, each sample dimension is: During the training phase, offline spectral analysis was used to determine the oscillation frequency label for each sample. In the online identification phase, only a 0.2s short window of data was input, and the predictor directly output the oscillation frequency. To verify the model's adaptability to missing data, different missing data rates were set in the test set, including 20%, 30%, and 40%.

[0069] Based on the experimental results, the oscillation frequency identification index of the method of the present invention under different data missing rates is shown in Table 1.

[0070] 20% 0.85 0.01 30% 0.87 0.01 40% 0.99 0.02

[0071] Table 1. Frequency identification index of the method of the present invention under different data missing rates.

[0072] As shown in Table 1, even when the test data missing rate reaches 40%, the dominant frequency identification MAPE of the method of the present invention can still be controlled within 1%, indicating that the joint training of the encoder-decoder reconstruction module, feature extractor and predictor can effectively alleviate the impact of missing field data on frequency identification accuracy and maintain high robustness of oscillation frequency identification.

[0073] Although the illustrative specific embodiments of the present invention have been described above to enable those skilled in the art to understand the invention, it should be understood that the invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the invention as defined and determined by the appended claims, and all inventions utilizing the concept of the present invention are protected.

Claims

1. A method for identifying sub / supersynchronous oscillation frequencies based on preceding short-window data reconstruction and deep learning, characterized in that, include: (1) Construct a sub / supersynchronous oscillation frequency identification network based on the reconstruction of preceding short window data; (2) Collect subsynchronous / supersynchronous oscillation sample data and construct training samples; (3) Train the sub / hypersynchronous oscillation frequency identification network using training samples until convergence; (4) Use the trained sub / supersynchronous oscillation frequency identification network to identify oscillations.

2. The sub / supersynchronous oscillation frequency identification method based on pre-sequence short window data reconstruction and deep learning according to claim 1, characterized in that, The sub / supersynchronous oscillation frequency identification network based on the reconstruction of preceding short window data includes a reconstruction module consisting of an encoder and a decoder, a feature extractor, and a predictor. The reconstruction module includes an encoder E and a decoder D; The encoder consists of three convolutional layers and max-pooling layers connected alternately. The first convolutional layer has 3 input channels, 16 output channels, a kernel length of 3, and padding of 1. The first max-pooling layer has a kernel length of 2 and a stride of 2. The second convolutional layer has 16 input channels, 8 output channels, a kernel length of 3, and padding of 1. The second max-pooling layer has a kernel length of 2 and a stride of 2. The third convolutional layer has 8 input channels, 3 output channels, a kernel length of 3, and padding of 1. The third max-pooling layer has a kernel length of 2 and a stride of 2. The decoder consists of four convolutional layers and upsampling layers connected in series with adjacent convolutional layers. The first convolutional layer has 3 input channels, 16 output channels, a kernel length of 3, and 1 padding. The second convolutional layer has 16 input channels, 32 output channels, a kernel length of 3, and 1 padding. The third convolutional layer has 32 input channels, 16 output channels, a kernel length of 3, and 1 padding. The fourth convolutional layer has 16 input channels, 3 output channels, a kernel length of 3, and 1 padding. All upsampling layers have a scale factor of 2 and use nearest-neighbor interpolation. The feature extractor consists of, in sequence, a first one-dimensional convolutional layer with 3 input channels, 16 output channels, a kernel length of 3, padding of 1, and a stride of 1; a first batch normalization layer with 16 channels; an effective channel attention layer with 16 channels; a first residual block with 16 input channels, 8 output channels, and a stride of 2; a nonlocal feature extraction block with 8 channels; a second batch normalization layer with 8 channels; a second residual block with 8 input channels, 4 output channels, and a stride of 1; a third residual block with 4 input channels, 1 output channel, and a stride of 1; a flattening layer; and a fully connected feature layer with a 128-dimensional deep oscillatory feature output. Each fully connected feature layer is followed by a Dropout layer with a dropout rate of 0.1 and a ReLU activation function. The predictor consists of a first fully connected predictive layer, a ReLU activation function, a Dropout layer, and a second fully connected predictive layer connected in series. The first fully connected predictive layer maps 128-dimensional deep oscillation features to 64-dimensional latent features. The dropout rate of the Dropout layer is 0.

1. The second fully connected predictive layer maps the 64-dimensional latent features to a 1-dimensional frequency output. Its output is the predicted value of the dominant frequency of the sub / supersynchronous oscillation corresponding to the short window sample.

3. The sub / supersynchronous oscillation frequency identification method based on pre-sequence short window data reconstruction and deep learning according to claim 1, characterized in that, The process of constructing training samples in step (2) is as follows: When a power system experiences subsynchronous / supersynchronous oscillations, acquire multi-channel electrical quantity data, including voltage, during the early stages of the oscillation. Current Active power reactive power ,frequency and phase angle ; set up The first collection phase, the... The multi-channel electrical quantity data collected in each acquisition phase are represented as follows: ， ； right The electrical quantity data of each channel are normalized. according to Construct a mask matrix from the electrical quantity data of each channel: ； in, For the first Line 1 Column elements, when For effective data sampling, ;otherwise, ; Will Acquired in each collection stage and the corresponding mask matrix This constitutes the training sample set.

4. The sub / supersynchronous oscillation frequency identification method based on pre-sequence short window data reconstruction and deep learning according to claim 1, characterized in that, The training process of the sub / supersynchronous oscillation frequency identification network in step (3) is as follows: (3.1) From Multi-channel electrical quantity data Selected from One data point is used as a batch of training data; (3.2) will indivual With the corresponding The input is fed into a reconstruction module consisting of an encoder and a decoder, and is mapped into latent features by the encoder E. : ； Then the latent features are decoded by decoder D. Reconstruct complete electrical quantity timing data : ； Input data and reconstructing data According to the mask matrix To merge: ； In the formula, This indicates element-wise multiplication. This represents the merged electrical data; (3.3) The merged electrical data The input is fed into the feature extractor F to obtain deep oscillation features: ； (3.4) Deep oscillation characteristics The input is fed into the predictor to predict the dominant frequency of the subsynchronous / supersynchronous oscillation: ； (3.5) Calculate the loss function value after this round of training; ； ； ； in, Let represent the dominant frequency label of the i-th training sample. Reconstruct loss weights for the data; (3.6) Based on the loss function value The parameters of the subsynchronous / supersynchronous oscillation frequency identification network are updated using the gradient descent method, and then the process returns to step (3.1) to perform the next round of training until the subsynchronous / supersynchronous oscillation frequency identification network converges.

Images