A method and system for broadband oscillation monitoring based on waveform data
By using multi-node waveform data coherence spectrum and CNN-XGBoost method, the problem of misjudgment in broadband oscillation monitoring of power systems is solved, accurate online monitoring is achieved in strong noise environment, cost is reduced and computational efficiency is improved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NARI NANJING CONTROL SYSTEM CO LTD
- Filing Date
- 2023-08-29
- Publication Date
- 2026-05-15
AI Technical Summary
Existing broadband oscillation monitoring methods for power systems frequently make incorrect judgments in strong noise environments and are sensitive to noise, making it difficult to accurately identify oscillation signals over a wide frequency range.
A method combining multi-node waveform data coherence spectrum with convolutional neural network (CNN) and extreme gradient boosting tree (XGBoost) is adopted. The power grid signal is sampled in real time through a wide-area measurement system. The coherence spectrum is used to distinguish oscillations and noise. CNN is used for unsupervised feature extraction and XGBoost is used for parameter identification to achieve accurate monitoring of oscillation frequency and attenuation factor.
It effectively reduces misjudgments caused by strong noise, lowers monitoring costs, improves the model's generalization ability and computational efficiency, and enables online oscillation signal monitoring over a wide frequency range.
Smart Images
Figure CN117131360B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to power systems, and in particular to a broadband oscillation monitoring method and system based on waveform data. Background Technology
[0002] The new generation of energy power systems is characterized by a high proportion of renewable energy and a high proportion of power electronic equipment. The dynamic process in which electrical quantities, caused by the interaction between power electronic equipment and the power grid, fluctuate periodically over time with a wide frequency range is called broadband oscillation in power systems, ranging from a few Hz to several kHz. In practical cases, broadband oscillations exhibit multiple modes and vary with the source-grid-load relationship. They are also influenced by the strength of disturbances and may be in different stages dominated by linear or nonlinearity, exhibiting strong nonlinear characteristics. The actual oscillation frequency will deviate from the eigenvalue analysis results, and the oscillation frequency will change with the power grid operation mode and the operating point of the power electronic equipment, i.e., frequency drift occurs, showing obvious time-varying characteristics. The strong nonlinearity and strong time-varying characteristics of broadband oscillation signals pose unprecedented challenges to the accurate identification of broadband oscillation parameters.
[0003] Currently, the most prevalent methods for power system oscillation monitoring are signal processing approaches, primarily focusing on low-frequency, subsynchronous / hypersynchronous oscillations. Typical methods include Fast Fourier Transform (FFT), Prony Transform, Hilbert-Huang Transform (HHT), Synchronous Compressed Wavelet Transform (SST), and various improved algorithms. However, each method has its applicable scope and limitations. FFT and Prony Transform can only be used for analyzing linear stationary signals. HHT can analyze nonlinear, non-stationary signals, but its Empirical Mode Decomposition (EMD) suffers from mode aliasing. SST addresses the energy divergence problem of the time spectrum by rearranging the traditional time spectrum and eliminates mode aliasing by changing the support interval, thus resolving the EMD mode aliasing issue. However, the SST method does not yield ideal results for time-varying frequency-enhanced signals. All of the above methods are sensitive to noise; therefore, their practicality in high-noise power system environments requires further investigation. Summary of the Invention
[0004] Purpose of the invention: The purpose of this invention is to provide a broadband oscillation monitoring method and system based on waveform data, thereby solving the problem of online monitoring of oscillation signals over a wide frequency range. Furthermore, it utilizes the coherence spectrum of multi-node waveform data to distinguish between oscillation and noise, effectively reducing misjudgments caused by strong noise, which is of great significance in engineering.
[0005] Technical solution: The broadband oscillation monitoring method based on waveform data described in this invention includes the following steps:
[0006] S1. The power grid system is sampled in real time by the synchronous phasor measurement unit of the wide-area measurement system to obtain the electrical signal waveform sequence of each time period, and the occurrence of oscillation is determined by the multi-channel signal coherence spectrum method.
[0007] S2. Define the electrical signal waveform sequence that was determined to be oscillating by the multi-channel signal coherence spectrum method in S1 as an oscillation sample. Input the oscillation sample into the unsupervised optimized convolutional neural network model (CNN model), compress the amount of data, extract oscillation features, reduce overfitting, and summarize and output the feature dataset D.
[0008] S3. Use the oscillation feature dataset D of the broadband oscillation signal output by the convolutional neural network as the input of the Limiting Gradient Boosting Tree (XGBoost) to identify the oscillation frequency and attenuation factor.
[0009] Step S1 includes the following steps:
[0010] S11. Calculate the power spectral density of the electrical signal waveform sequence and the cross-power spectral density of different nodes;
[0011] S12. Calculate the coherence spectrum between different nodes, and determine whether oscillation has occurred based on the peak value of the coherence spectrum.
[0012] The specific expression for the coherence spectrum described in S12 is as follows:
[0013]
[0014] Where P x1x1 and P x2x2 These are the power spectral densities of signals X1(t) and X2(t), respectively; P x1x2 It is the cross power spectral density of signals X1(t) and X2(t).
[0015] Step S2 includes the following steps:
[0016] S21. Perform unsupervised optimization on the CNN model to eliminate the need for setting the number of labels;
[0017] S22. Input the waveform data that is determined to be oscillating into the CNN model for feature extraction.
[0018] Step S21 includes the following steps:
[0019] The output value after superimposing n sets of single-mode signal time-domain waveforms is:
[0020]
[0021] In the formula, t represents time; A i Let be the amplitude of the i-th modal signal; f represents the phase of the i-th modal signal; i λ is the frequency of the i-th modal signal; i Let be the attenuation factor of the i-th mode signal;
[0022] The loss function is defined as:
[0023]
[0024] Where n is the total number of oscillation samples; g(t) i ) represents the i-th output value; y(t) i Let ) be the i-th expected value;
[0025] Adjust the network weights ω and biases γ using the loss function E of the CNN model's output layer:
[0026]
[0027]
[0028] Where α is the network learning efficiency; ω and γ are the unadjusted network weights and biases; and ω' and γ' are the adjusted network weights and biases using the loss function E.
[0029] Step S22 includes the following steps:
[0030] The raw oscillating data is fed into the CNN model through the input layer;
[0031] Utilizing convolutional layers to extract key features from oscillating data;
[0032] Use pooling layers to reduce the dimensionality of oscillating data;
[0033] The processed oscillating data is classified using a fully connected layer;
[0034] Finally, the output layer outputs the results to the next network model.
[0035] Step S3 includes the following steps:
[0036] S31. Input the sample feature set D into the XGBoost model, set the number of iterations, loss function l, regularization coefficients λ and γ, and set the initial number of iterations to 1; calculate the first derivative of all samples with respect to the current loss function l and G. j And the second derivative and H j Set the default parameter G L =0, H L =0, G R =0, H R =0;
[0037] S32. Calculate the new G after the sample is placed into the left subtree. L With H L After adding it to the right subtree, the new G R With H R The gain parameter score is updated, and features and feature values are divided into subtrees based on the gain parameter score to achieve oscillation parameter prediction.
[0038] The specific expression of the XGBoost model mentioned in step S31 is as follows:
[0039]
[0040] Among them, f p (x i () represents the p-th tree model; p is the tree index; x i It is the i-th oscillating sample; F is the function space composed of decision trees; It is a predicted value;
[0041] The specific expression for the loss function is:
[0042]
[0043] Where, x i It is the i-th oscillation sample; y i These are the actual parameter values.
[0044] The specific expression for the gain parameter score mentioned in step S32 is as follows:
[0045]
[0046] In the formula, G L With H L These are the first and second derivatives of the new loss function after the sample is placed into the left subtree; G R With H R These are the first and second derivatives of the new loss function after the data is placed into the right subtree.
[0047] A broadband oscillation monitoring system based on waveform data includes the following modules:
[0048] Real-time sampling module: The power grid system is sampled in real time through the synchronous phasor measurement unit of the wide-area measurement system to obtain the electrical signal waveform sequence for each time period;
[0049] Oscillation detection module: Detects the occurrence of oscillations using a multi-channel signal coherence spectrum method;
[0050] Sample input module: Inputs oscillating samples into the CNN model after unsupervised optimization;
[0051] Feature dataset output module: Compresses the amount of data through convolutional and pooling layers, extracts oscillatory features, and reduces overfitting. The fully connected layer summarizes and outputs the feature dataset D.
[0052] Oscillation monitoring module: The oscillation characteristics of the wideband oscillation signal output by the convolutional neural network are used as the input of XGBoost to identify the oscillation frequency and attenuation factor.
[0053] The specific expression for the coherence spectrum is:
[0054]
[0055] Where P x1x1 and P x2x2 These are the power spectral densities of signals X1(t) and X2(t), respectively; P x1x2 It is the cross power spectral density of signals X1(t) and X2(t).
[0056] The sample input module operates as follows:
[0057] The output value after superimposing n sets of time-domain waveforms of single-mode signals is:
[0058]
[0059] In the formula, t represents time; A i Let be the amplitude of the i-th modal signal; f represents the phase of the i-th modal signal; i λ is the frequency of the i-th modal signal; i Let be the attenuation factor of the i-th mode signal;
[0060] The loss function is defined as:
[0061]
[0062] Where n is the total number of oscillation samples; g(t) i ) represents the i-th output value; y(t) i Let ) be the i-th expected value;
[0063] Adjust the network weights ω and biases γ using the loss function E of the CNN model's output layer:
[0064]
[0065]
[0066] Where α is the network learning efficiency; ω and γ are the unadjusted network weights and biases; and ω' and γ' are the adjusted network weights and biases using the loss function E.
[0067] The feature dataset output module operates as follows:
[0068] The raw oscillating data is fed into the CNN model through the input layer;
[0069] Utilizing convolutional layers to extract key features from oscillating data;
[0070] Use pooling layers to reduce the dimensionality of oscillating data;
[0071] The processed oscillating data is classified using a fully connected layer;
[0072] Finally, the output layer outputs the results to the next network model.
[0073] The oscillation monitoring module operates as follows:
[0074] Input the sample feature set D into the XGBoost model, set the number of iterations, loss function l, regularization coefficients λ and γ, and set the initial number of iterations to 1; calculate the first derivative of all samples with respect to the current loss function l and G. j And the second derivative and H j Set the default parameter G L =0, H L =0, G R =0, H R =0;
[0075] Calculate the new G after the sample is placed into the left subtree L With H L After adding it to the right subtree, the new G R With H R The gain parameter score is updated, and features and feature values are divided into subtrees based on the gain parameter score to achieve oscillation parameter prediction.
[0076] A computer storage medium storing a computer program that, when executed by a processor, implements the aforementioned broadband oscillation monitoring method based on waveform data.
[0077] A computer device includes a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the above-described broadband oscillation monitoring method based on waveform data.
[0078] Beneficial effects: Compared with the prior art, the present invention has the following advantages:
[0079] This invention eliminates the need for additional equipment installation. It utilizes data from existing synchronous phasor measurement devices widely installed in the power grid to monitor oscillations over a wide frequency range, significantly reducing monitoring costs and solving the problem of online monitoring of oscillation signals over a wide frequency range. Furthermore, it uses the coherence spectrum of multi-node waveform data to distinguish between oscillations and noise, effectively reducing misjudgments caused by strong noise. By constructing the CNN in an unsupervised manner, it eliminates the need for label setting, improving the model's generalization ability and computational efficiency, which is of great significance in engineering. Attached Figure Description
[0080] Figure 1 This is a flowchart of the steps of the method described in this invention;
[0081] Figure 2 This is a schematic diagram of a convolutional neural network structure;
[0082] Figure 3 This is a schematic diagram of the XGBoost algorithm structure;
[0083] Figure 4 This is a flowchart of a broadband oscillation monitoring method based on waveform data. Detailed Implementation
[0084] The technical solution of the present invention will be further described below with reference to the accompanying drawings.
[0085] like Figure 1 As shown, a broadband oscillation monitoring method based on waveform data includes the following steps:
[0086] S1. Coherence spectrum calculation based on waveform data
[0087] The standard coherence spectrum is defined as shown in equation (1), where P x1x1 and P x2x2 P represents the power spectral density of signals X1(t) and X2(t), respectively. x1x2 It is the cross power spectral density of signals X1(t) and X2(t).
[0088]
[0089]
[0090] C x1x2 (f) can well reflect the degree of linear correlation between X1(t) and X2(t) at frequency f. If X1(t) is a sine function with frequency f1 and X2(t) is a sine function with frequency f2, then C x1x2 (f1) and C x1x2 The relationship of (f2) is:
[0091]
[0092] Wideband oscillations exhibit wide-area propagation characteristics, so oscillations detected on different lines are often linearly similar to each other, represented by high peak values in the coherence spectrum. Mathematically, if the fluctuations in waveform data originate from oscillations, then multiple waveform data points can be linearly represented, and the value of the coherence spectrum function is close to 1. If the fluctuations in waveform data originate from noise rather than oscillations, then multiple waveform data points cannot be linearly represented, and the value of the coherence spectrum function is close to 0. In practical engineering applications, a reasonable threshold is set to determine the degree of linear correlation of signals; a threshold of 0.7 is generally used.
[0093] S2. Wideband Oscillation Parameter Identification Based on CNN-XGBoost
[0094] S2.1, Convolutional Neural Network (CNN)
[0095] The waveform data identified as oscillating in the previous section is input into a Convolutional Neural Network (CNN) for feature extraction. CNN models excel in feature extraction, compensating for the shortcomings of non-CNN models. The accuracy of oscillation data feature extraction directly impacts the accuracy of oscillation identification. A CNN model consists of five basic functional modules: an input layer handles the raw oscillating data; convolutional layers extract key features from the oscillating data; pooling layers reduce the dimensionality of the oscillating data; fully connected layers classify the processed oscillating data; and finally, the results are output to the next network model. The structure of a CNN is as follows: Figure 2 As shown.
[0096] To avoid the need for label setting, the CNN is first subjected to unsupervised optimization. The output value after superimposing the time-domain waveforms of n single-modal signals is:
[0097]
[0098] In the formula: t is time; A i Let be the amplitude of the i-th modal signal; f represents the phase of the i-th modal signal; i λ is the frequency of the i-th modal signal; i is the attenuation factor of the i-th mode signal.
[0099] The loss function is defined as:
[0100]
[0101] Where: n is the total number of oscillation samples; g(t) i ) represents the i-th output value; y(t) i Let be the i-th expected value.
[0102] Adjust the network weights ω and biases γ using the loss function E of the CNN model's output layer:
[0103]
[0104]
[0105] Where α is the network learning efficiency; ω and γ are the unadjusted network weights and biases; and ω' and γ' are the adjusted network weights and biases using the loss function E.
[0106] S2.2, XGBoost model
[0107] like Figure 3 As shown, the XGBoost algorithm is used to identify the feature information extracted by the CNN model. The core idea of the XGBoost algorithm is to continuously split the data features to build new trees to fit the residual between the previous predicted value and the actual value, and then sum the results of all trees to obtain the final prediction result. The expression formula of the ensemble model is as follows:
[0108]
[0109] f p (x i () represents the p-th tree model; p is the tree index; x i It is the i-th oscillating sample; F is the function space composed of decision trees; These are predicted values. For N t A set of arbitrarily different oscillation characteristics:
[0110]
[0111] In the formula, D is about x i y i The dataset; y i It is the i-th set x i The label is ; R is the set of real numbers.
[0112] The formula for calculating the function space F above is:
[0113]
[0114] In the formula q(x i ) represents sample x i A leaf node; Represents the leaf node q(x) i The weight of ).
[0115] The tree structure q maps the oscillation sets of different features to leaf nodes through attribute discrimination. Each decision tree model corresponds to a unique tree structure q and a corresponding leaf node weight vector ω. For example... Figure 4 As shown, for the oscillation feature set D, the XGBoost model obtains the final predicted value by mapping the oscillation feature set to the corresponding leaf node on each decision tree, and then superimposing the weights of the P leaf nodes corresponding to the oscillation sample.
[0116] The corresponding objective function is defined as:
[0117]
[0118]
[0119]
[0120] In the formula, l is the loss function, representing the target predicted value. Compared with the target true value y i The error between them; Ω represents the regularization function, used to control the complexity of the tree and prevent overfitting; T is the total number of leaf nodes in each tree; ω represents the weight of each leaf node in each tree; λ is the regularization penalty term parameter; γ is the penalty parameter of T.
[0121] During the training of the prediction model, it is necessary to fix the already learned trees and add a new tree to them each time in order to minimize the objective function:
[0122]
[0123] In the formula This is the predicted value obtained at step t; during model training, a new function f is added in each round to minimize the objective function. t (x i The objective function for round t is:
[0124]
[0125] in
[0126]
[0127] Finally, the optimal leaf node score is calculated.
[0128]
[0129] In the above formula, j represents a leaf node; I j ={i|q(x i )=j}, which represents the set of nodes to which the sample is mapped.
[0130] The final objective function is in the form of
[0131]
[0132] S3, Wideband Oscillation Identification Process
[0133] S3.1. The power grid system is sampled in real time by the synchronous phasor measurement unit of the wide-area measurement system to obtain the electrical signal waveform sequence for each time period;
[0134] S3.2. By calculating the coherence spectrum of waveform data samples at different nodes, the data samples in which oscillations occur are determined.
[0135] S3.3. Unsupervised optimization of the CNN model is performed by inputting oscillating samples into the CNN model, compressing the amount of data through convolutional and pooling layers, extracting oscillating features, and reducing overfitting. The fully connected layers summarize and output the feature dataset D.
[0136] S3.4. Build an XGBoost model using the feature set D output by the CNN model, and use the training samples as input to train the recognition model. Through continuous learning and parameter tuning, obtain the optimal recognition model.
[0137] S3.5. Using the test samples as input to XGBoost, XGBoost is used to mine the time series information of feature set D to obtain its internal variation patterns, and finally the XGBoost recognition result is output. That is, output the frequency and attenuation factor of each oscillation.
[0138] A broadband oscillation monitoring system based on waveform data includes the following modules:
[0139] Real-time sampling module: The power grid system is sampled in real time through the synchronous phasor measurement unit of the wide-area measurement system to obtain the electrical signal waveform sequence for each time period;
[0140] Oscillation detection module: Detects the occurrence of oscillations using a multi-channel signal coherence spectrum method;
[0141] Sample input module: Inputs oscillating samples into the CNN model after unsupervised optimization;
[0142] Feature dataset output module: Compresses the amount of data through convolutional and pooling layers, extracts oscillatory features, and reduces overfitting. The fully connected layer summarizes and outputs the feature dataset D.
[0143] Oscillation monitoring module: The oscillation characteristics of the wideband oscillation signal output by the convolutional neural network are used as the input of XGBoost to identify the oscillation frequency and attenuation factor.
[0144] The specific expression for the coherence spectrum is:
[0145]
[0146] Where P x1x1 and P x2x2 These are the power spectral densities of signals X1(t) and X2(t), respectively; P x1x2 It is the cross power spectral density of signals X1(t) and X2(t).
[0147] The sample input module operates as follows:
[0148] The output value after superimposing n sets of time-domain waveforms of single-mode signals is:
[0149]
[0150] In the formula, t represents time; A i Let be the amplitude of the i-th modal signal; f represents the phase of the i-th modal signal; i λ is the frequency of the i-th modal signal; i Let be the attenuation factor of the i-th mode signal;
[0151] The loss function is defined as:
[0152]
[0153] Where n is the total number of oscillation samples; g(t) i ) represents the i-th output value; y(t) i Let ) be the i-th expected value;
[0154] Adjust the network weights ω and biases γ using the loss function E of the CNN model's output layer:
[0155]
[0156]
[0157] Where α is the network learning efficiency; ω and γ are the unadjusted network weights and biases; and ω' and γ' are the adjusted network weights and biases using the loss function E.
[0158] The feature dataset output module operates as follows:
[0159] The raw oscillating data is fed into the CNN model through the input layer;
[0160] Utilizing convolutional layers to extract key features from oscillating data;
[0161] Use pooling layers to reduce the dimensionality of oscillating data;
[0162] The processed oscillating data is classified using a fully connected layer;
[0163] Finally, the output layer outputs the results to the next network model.
[0164] The oscillation monitoring module operates as follows:
[0165] Input the sample feature set D into the XGBoost model, set the number of iterations, loss function l, regularization coefficients λ and γ, and set the initial number of iterations to 1; calculate the first derivative of all samples with respect to the current loss function l and G. j And the second derivative and H j Set the default parameter G L =0, H L =0, G R =0, H R =0;
[0166] Calculate the new G after the sample is placed into the left subtree L With H L After adding it to the right subtree, the new G R With H R The gain parameter score is updated, and features and feature values are divided into subtrees based on the gain parameter score to achieve oscillation parameter prediction.
[0167] A computer storage medium storing a computer program that, when executed by a processor, implements the aforementioned broadband oscillation monitoring method based on waveform data.
[0168] A computer device includes a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the above-described broadband oscillation monitoring method based on waveform data.
[0169] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented in various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.
[0170] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0171] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0172] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0173] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.
[0174] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.
Claims
1. A broadband oscillation monitoring method based on waveform data, characterized in that, Includes the following steps: S1. The power grid system is sampled by the synchronous phasor measurement unit of the wide-area measurement system to obtain the electrical signal waveform sequence of each time period, and the occurrence of oscillation is determined by the multi-channel signal coherence spectrum method. S2. Define the electrical signal waveform sequence that was determined to be oscillating by the multi-channel signal coherence spectrum method in S1 as an oscillation sample. Input the oscillation sample into the unsupervised optimized convolutional neural network (CNN) model, compress the amount of data, extract oscillation features, reduce overfitting, and summarize and output the oscillation feature dataset D. Unsupervised optimization of CNN models includes the following steps: The output value after superimposing n sets of time-domain waveforms of single-mode signals is: In the formula, t represents time; Let be the amplitude of the i-th modal signal; Let be the phase of the i-th modal signal; Let be the frequency of the i-th modal signal; Let be the attenuation factor of the i-th mode signal; The loss function is defined as: Where n is the total number of oscillation samples; This is the i-th output value; Let i be the expected value; Adjusting network weights using the loss function E of the CNN model's output layer and bias : in, To improve the efficiency of online learning; and These are the unadjusted network weights and biases; and It is the network weights and biases adjusted using the loss function E; S3. Based on the oscillation feature dataset D of the broadband oscillation signal output by the convolutional neural network, establish and train the limit gradient boosting tree XGBoost recognition model, and use the trained recognition model to identify the oscillation frequency and attenuation factor.
2. The broadband oscillation monitoring method based on waveform data according to claim 1, characterized in that, Step S1 includes the following steps: S11. Calculate the power spectral density of the electrical signal waveform sequence and calculate the cross-power spectral density at different nodes; S12. Calculate the coherence spectrum between different nodes, and determine whether oscillation has occurred based on the peak value of the coherence spectrum.
3. The broadband oscillation monitoring method based on waveform data according to claim 2, characterized in that, The specific expression for the coherence spectrum described in S12 is as follows: in and These are signals and The power spectral density; It is a signal and The cross-power spectral density.
4. The broadband oscillation monitoring method based on waveform data according to claim 1, characterized in that, Step S2 includes the following steps: S21. Perform unsupervised optimization on the CNN model to eliminate the need for setting the number of labels; S22. Input the waveform data that is determined to be oscillating into the CNN model for feature extraction.
5. The broadband oscillation monitoring method based on waveform data according to claim 4, characterized in that, Step S22 includes the following steps: The raw oscillating data is fed into the CNN model through the input layer; Utilizing convolutional layers to extract key features from oscillating data; Use pooling layers to reduce the dimensionality of oscillating data; The processed oscillation data is classified using a fully connected layer, and an oscillation feature dataset D is established. Finally, the oscillation feature dataset D is output to the next network model through the output layer.
6. The broadband oscillation monitoring method based on waveform data according to claim 1, characterized in that, Step S3 includes the following steps: S31. Input the oscillation feature dataset D into the XGBoost model, and set the number of iterations, loss function l, and regularization coefficient. and The initial number of iterations is set to 1; calculate the first derivative of all samples with respect to the current loss function l and G. j And the second derivative and H j Set default parameters , , , ; S32, Calculate the new sample after it is placed in the left subtree. and After adding to the right subtree, the new and The gain parameter score is updated, and features and feature values are divided into subtrees based on the gain parameter score to achieve oscillation parameter prediction.
7. The broadband oscillation monitoring method based on waveform data according to claim 6, characterized in that, The specific expression of the XGBoost model mentioned in step S31 is as follows: in, This is the model of the p-th tree; p is the tree index; x i It is the i-th oscillating sample; F is the function space composed of decision trees; These are the predicted frequency and attenuation factor of each oscillation; The specific expression for the loss function is: Where, x i It is the i-th oscillation sample; These are the actual parameter values.
8. The broadband oscillation monitoring method based on waveform data according to claim 6, characterized in that, The specific expression for the gain parameter score mentioned in step S32 is as follows: In the formula, and These are the first and second derivatives of the new loss function after the sample is placed into the left subtree; and These are the first and second derivatives of the new loss function after the data is placed into the right subtree.
9. A broadband oscillation monitoring system based on waveform data, characterized in that, Includes the following modules: Real-time sampling module: The power grid system is sampled in real time through the synchronous phasor measurement unit of the wide-area measurement system to obtain the electrical signal waveform sequence for each time period; Oscillation discrimination module: The oscillation is determined by the multi-channel signal coherence spectrum method. The electrical signal waveform sequence that is determined to be oscillating by the multi-channel signal coherence spectrum method is defined as an oscillation sample. Sample input module: Inputs oscillating samples into the unsupervised optimized convolutional neural network model; Unsupervised optimization of convolutional neural network models includes the following steps: The output value after superimposing n sets of time-domain waveforms of single-mode signals is: In the formula, t represents time; Let be the amplitude of the i-th modal signal; Let be the phase of the i-th modal signal; Let be the frequency of the i-th modal signal; Let be the attenuation factor of the i-th mode signal; The loss function is defined as: Where n is the total number of oscillation samples; This is the i-th output value; Let i be the expected value; Adjusting network weights using the loss function E of the CNN model's output layer and bias : in, To improve the efficiency of online learning; and These are the unadjusted network weights and biases; and It is the network weights and biases adjusted using the loss function E; Feature dataset output module: Compresses the amount of data through convolutional and pooling layers, extracts oscillation features, and reduces overfitting. The fully connected layer summarizes and outputs the oscillation feature dataset D. Oscillation monitoring module: Based on the oscillation feature dataset D of the broadband oscillation signal output by the convolutional neural network, a limit gradient boosting tree (XGBoost) recognition model is established and trained, and the trained recognition model is used to identify the oscillation frequency and attenuation factor.
10. A broadband oscillation monitoring system based on waveform data according to claim 9, characterized in that, The specific expression for the coherence spectrum is: in and These are signals and The power spectral density; It is a signal and The cross-power spectral density.
11. A broadband oscillation monitoring system based on waveform data according to claim 9, characterized in that, The feature dataset output module operates as follows: The raw oscillating data is fed into the CNN model through the input layer; Utilizing convolutional layers to extract key features from oscillating data; Use pooling layers to reduce the dimensionality of oscillating data; The processed oscillation data is classified using a fully connected layer, and an oscillation feature dataset D is established. Finally, the oscillation feature dataset D is output to the next network model through the output layer.
12. A broadband oscillation monitoring system based on waveform data according to claim 9, characterized in that, The oscillation monitoring module operates as follows: The oscillation feature dataset D is input into the XGBoost model, and the number of iterations, loss function l, and regularization coefficient are set. and The initial number of iterations is set to 1; calculate the first derivative of all samples with respect to the current loss function l and G. j And the second derivative and H j Set default parameters , , , ; Calculate the new sample after it is placed in the left subtree and After adding to the right subtree, the new and The gain parameter score is updated, and features and feature values are divided into subtrees based on the gain parameter score to achieve oscillation parameter prediction.
13. A computer storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements a broadband oscillation monitoring method based on waveform data as described in any one of claims 1-8.
14. A computer device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements a broadband oscillation monitoring method based on waveform data as described in any one of claims 1-8.