A method and related apparatus for locating broadband oscillators based on LCD-Hilbert subband envelope alignment and directed graph learning model.
By combining LCD-Hilbert subband envelope alignment with a directed graph learning model, the accuracy and adaptability issues in broadband oscillator localization are solved, enabling rapid and accurate localization of broadband oscillators in power systems and improving the safety and stability of the power grid.
Patent Information
- Application Number
- CN202611125839.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-28
- Publication Date
- 2026-11-10
AI Technical Summary
Existing broadband oscillation source localization methods suffer from insufficient accuracy and poor adaptability when processing multi-frequency and multi-mode oscillation information, especially in power systems with a high proportion of new energy sources and power electronic equipment, where it is difficult to accurately identify oscillation sources.
A method based on LCD-Hilbert sub-band envelope alignment and directed graph learning model is adopted. The sub-band envelope sequence of the node is extracted by local feature scale decomposition and Hilbert transform. The node time-frequency matrix and directional propagation matrix are constructed. Combined with the directed graph learning model, the fusion localization of node oscillation characteristics and propagation information is realized.
It enables rapid and accurate localization of broadband oscillation sources, solves the problems of signal mode aliasing and noise sensitivity, improves positioning accuracy and adaptability, and provides technical support for the safe and stable operation of the power grid.
Smart Images

Figure CN123073705A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of broadband oscillation analysis and source tracing technology in power systems, specifically a broadband oscillation source localization method and related apparatus based on LCD-Hilbert sub-band envelope alignment and directed graph learning model. Background Technology
[0002] With the large-scale integration of high-proportion renewable energy sources and high-proportion power electronic equipment, the dynamic behavior of the power system exhibits characteristics of cross-frequency band coupling, multi-device interaction, and complex propagation paths. Wideband oscillations have become a significant issue affecting the safety of grid-connected equipment and the stable operation of the system. Wideband oscillations cover a wide frequency range from ultra-low frequency to mid-high frequency, and their inducing mechanisms are complex. They may manifest as local equipment oscillations or form regional resonances after propagating through the power grid.
[0003] Existing methods for locating broadband oscillation sources generally include those based on mechanistic models, those based on impedance criteria, and those based on deep learning. Mechanistic model methods often rely on device controller parameters and precise system models, which are difficult to obtain comprehensively in practical engineering. While impedance methods offer strong physical interpretability, they often require certain prior conditions or suffer from difficulties in uniformly setting thresholds. In recent years, deep learning models have been introduced into the field of oscillation identification, demonstrating powerful feature learning capabilities. However, while pure deep learning methods have good classification capabilities, they suffer from strong black-box characteristics and insufficient mechanistic explanation, resulting in limited accuracy and generalization ability in locating complex oscillation modes.
[0004] Before oscillation localization, the ability to reliably extract multi-band effective information of broadband oscillations from synchronous waveform measurement signals is fundamental to subsequent propagation mechanism analysis and source localization. Traditional Fourier methods struggle to reflect obvious non-stationary time-frequency characteristics, wavelet methods face difficulties in wavelet basis selection, and EMD-type empirical mode methods may exhibit mode aliasing and endpoint effects.
[0005] In summary, there is an urgent need to develop a broadband oscillation localization method that can deeply integrate signal time-frequency analysis with advanced artificial intelligence models to overcome the shortcomings of existing technologies in processing multi-band, multi-mode oscillation information. By learning feature matrices with clear physical meaning, the advanced artificial intelligence model can be guaranteed to have higher accuracy and stronger adaptability when applied to broadband oscillation localization, providing strong technical support for the safe and stable operation of the power grid. This method has significant theoretical value and promising engineering application prospects. Summary of the Invention
[0006] To address the aforementioned deficiencies in existing technologies, this invention provides a broadband oscillation source localization method and related apparatus based on LCD-Hilbert subband envelope alignment and directed graph learning model, aiming to achieve sensitive, reliable, and accurate localization of broadband oscillation sources in power systems.
[0007] To achieve the above objectives, the present invention provides a broadband oscillation source localization method based on LCD-Hilbert subband envelope alignment and directed graph learning model, comprising:
[0008] S1. Acquire synchronous measurement signals from multiple monitoring nodes in the power system, perform time alignment processing on the measurement signals of each node, and obtain the signal set to be analyzed.
[0009] S2. Using local feature scale decomposition LCD and Hilbert transform, sub-band decomposition is performed on the signal set to be analyzed, the envelope sequence of each monitoring node under different sub-bands is extracted, and the node time-frequency matrix is constructed.
[0010] S3. Based on the sub-band envelope sequence of each monitoring node and the singular value distribution of the node time-frequency matrix, construct a node characteristic matrix that characterizes the local oscillation characteristics of the node.
[0011] S4. At the main station, obtain the adjacency mask matrix that reflects the node relationship based on the power grid topology, and calculate the minimum alignment error of the delay envelope of adjacent nodes in both directions under each sub-frequency band; based on the difference of the minimum alignment error of the delay envelope in both directions, construct a directional propagation matrix that characterizes the direction and intensity of oscillation propagation.
[0012] S5. Establish and train a directed graph learning model, wherein the directed graph learning model takes the node feature matrix and the directed propagation matrix corresponding to the broadband oscillation sample as input, and the oscillation source node label as supervision information, and trains to obtain a directed graph learning model for outputting the node score vector.
[0013] S6. After the system experiences broadband oscillation, the actual measurement signal is processed according to steps S1 to S4 to obtain the feature matrix and directional propagation matrix of the node to be located. The feature matrix and directional propagation matrix of the node to be located are input into the trained directed graph learning model. The local oscillation features of the node and the directional propagation information between nodes are fused through multi-layer graph convolution to output the node score vector. The node score vector is then input into the Softmax function to obtain the oscillation source probability corresponding to each node. The broadband oscillation source is located based on the oscillation source probability.
[0014] Furthermore, the synchronization measurement signal in S1 is the active power measurement signal of each monitoring node, and the synchronization measurement signal has a unified timestamp to achieve synchronization alignment between different monitoring nodes.
[0015] Furthermore, the process of constructing the node time-frequency matrix in S2 includes:
[0016] S21. The original signal is decomposed into intrinsic scale components (ISC components) using the LCD algorithm. When the number of extreme points in the residual components of the previous decomposition is less than two, or the amplitude of the residual components is less than the threshold, the LCD decomposition is stopped, and each ISC component is filtered using a bandpass filtering algorithm.
[0017] S22. Based on the obtained ISC components, perform a Hibert transform on each ISC component to obtain the instantaneous frequency and instantaneous amplitude signals. For each sampling point, accumulate the amplitude of each ISC at that point to the corresponding sub-band according to its instantaneous frequency, to obtain the envelope sequence of node i in sub-band b. The envelope sequences of each sub-band are then arranged according to frequency band and time to form a node time-frequency matrix A. i :
[0018]
[0019] Where, matrix A i The rows correspond to different sub-frequency bands, and the columns correspond to sampling times. This matrix centrally represents the oscillation evolution information of node i in each frequency band.
[0020] Furthermore, the construction process of the node feature matrix in S3 includes:
[0021] S31. Define Nodes In sub-band The envelope energy is:
[0022]
[0023] in, For nodes In sub-band The envelope sequence below, To analyze data length;
[0024] S32. Define Nodes The total intensity of the local oscillation is:
[0025]
[0026] in, This represents the total number of sub-bands.
[0027] S33. Define Nodes In sub-band The energy percentage below is:
[0028]
[0029] in, It is a very small positive number;
[0030] S34. Time-frequency matrix of nodes Perform singular value decomposition and define nodes. The modal concentration is:
[0031]
[0032] in, The maximum singular value of the node time-frequency matrix. Let be the rank of the node time-frequency matrix;
[0033] S35. Concatenate the feature vectors of each node vertically to construct the node feature matrix. .
[0034]
[0035] Furthermore, the construction process of the directional propagation matrix in step S4 includes:
[0036] S41. Construct an adjacency mask matrix based on the power grid topology. When node With nodes When directly connected, ,otherwise ;
[0037] S42. Normalize the envelope sequences of the nodes in each sub-band to obtain normalized envelope sequences. ;
[0038]
[0039] S43. For adjacent nodes and Define nodes Pointing to node In sub-band Below, time delay is The time delay alignment error is:
[0040]
[0041] in, For nodes In sub-band The normalized envelope sequence below, As a candidate time delay, satisfying , The maximum search latency is preset.
[0042] S44. Determine the optimal forward propagation delay based on the magnitude of each delay alignment error sequence, and thereby determine the forward matching degree, specifically as follows:
[0043] The optimal propagation delay is selected from the data point corresponding to the minimum delay alignment error.
[0044]
[0045] Further define the positive matching degree as:
[0046]
[0047] Similarly, the reverse matching degree can be obtained. ,like Significantly greater than ,illustrate" "current envelope" and " The delayed envelope is easier to align, and therefore more consistent with the node. Located at node The oscillation propagates upstream.
[0048] S45. Define nodes based on the difference between positive and negative matching degrees. To the node The intensity of transmission And construct a directional propagation matrix based on the propagation intensity of each node. .
[0049]
[0050] And order: The directional propagation strength between each node forms the directional propagation matrix:
[0051]
[0052] Furthermore, step S5 includes the training method for the directed graph learning model and the process for obtaining the node score vectors, which includes:
[0053] S51. Construct a broadband oscillation sample set, which includes historical broadband oscillation records, simulated broadband oscillation samples, or a combination of both. For each broadband oscillation sample, construct the corresponding node feature matrix X and directed propagation matrix P according to steps S1 to S4, and use the oscillation source nodes in the broadband oscillation sample as oscillation source node labels. Use the node feature matrix X and directed propagation matrix P as input to the directed graph learning model, and use the oscillation source node labels as supervision information to train the trainable parameters in the directed graph learning model to obtain the trained directed graph learning model.
[0054] S52. The node feature matrix X obtained in S3 and the directional propagation matrix P obtained in S4 are used as inputs to the graph convolutional network. The node feature matrix X represents the local oscillation characteristics of the nodes, while the directional propagation matrix P represents the propagation relationship between nodes. Through graph convolution operations, the node features and the information propagated by neighboring nodes are fused to obtain the score vector H of each node as an oscillation source.
[0055] The directed graph learning model is preferably a directed weighted graph convolutional network, whose... +1 layer convolution is represented as:
[0056]
[0057] in, This represents the node feature representation of the l-th layer; This is the node feature matrix of the input layer, i.e., the node feature matrix obtained in step 3; This represents the information that is aggregated from the propagation information between nodes to the current node; This represents information propagated outward from the current node; and These are the in-degree matrix and the out-degree matrix, respectively, used to normalize the results of graph convolution operations; and This is a trainable weight matrix used to adjust the influence weights of node features; For non-linear activation functions, the ReLU function is typically used.
[0058] Through graph convolution operations at each layer, the local oscillation characteristics of a node are fused with the propagation information of its neighboring nodes layer by layer. Under the action of the multi-layer graph convolutional network, the score matrix of each node gradually fuses from local information to global propagation information, ultimately obtaining a node score phasor H that can reflect whether a node is an oscillation source.
[0059] Furthermore, the method for obtaining the oscillation source probability of each node in step S6 is as follows:
[0060] After broadband oscillation occurs, the actual measurement signal is processed according to steps S1 to S4. The obtained node feature matrix and directional propagation matrix are input into the trained graph convolutional model, and the node score vector output by the graph convolutional network is... The value is passed as input to the Softmax function to obtain the probability k of each node acting as an oscillation source. i for:
[0061]
[0062] in, It refers to the output features of the last layer of a graph convolutional network. Corresponding node eigenvalues, The total number of nodes is represented by 0. The score matrix is mapped to the range of 0 to 1. Based on the obtained oscillation source probabilities, the node with the highest oscillation source probability is the broadband oscillation source, thus realizing the broadband oscillation source localization.
[0063] On the other hand, the present invention also provides a broadband oscillator localization device based on LCD-Hilbert sub-band envelope alignment and directed graph learning model, for performing the broadband oscillator localization method based on LCD-Hilbert sub-band envelope alignment and directed graph learning model as described above, including:
[0064] The signal acquisition module is used to acquire synchronous measurement signals from multiple monitoring nodes in the power system, perform time alignment processing on the measurement signals from each node, and obtain the signal set to be analyzed.
[0065] The time-frequency decomposition module is used to decompose the LCD and Hilbert transform using local feature scale decomposition, perform sub-frequency band decomposition on the signal set to be analyzed, extract the envelope sequence of each monitoring node under different sub-frequency bands, and construct the node time-frequency matrix of each monitoring node.
[0066] The node feature construction module is used to perform singular value decomposition based on the sub-band envelope sequence of each monitoring node and the node time-frequency matrix to construct a node feature matrix that characterizes the local oscillation characteristics of the node.
[0067] The directional propagation construction module is used to construct an adjacency mask matrix according to the power grid topology, calculate the minimum alignment error of the delay envelope of adjacent nodes in both directions under each sub-frequency band based on the sub-frequency band envelope sequence of each monitoring node, and construct a directional propagation matrix characterizing the oscillation propagation direction and propagation intensity based on the difference between the minimum alignment errors of the delay envelope in both directions.
[0068] The model training module is used to establish and train a directed graph learning model. It takes the node feature matrix and the directed propagation matrix corresponding to the broadband oscillation sample as input and the oscillation source node label as supervision information to train a directed graph learning model for outputting node score vectors.
[0069] The oscillation source localization module is used to process the actual measurement signal through the time-frequency decomposition module, node feature construction module, and directional propagation construction module after the system experiences broadband oscillation, to obtain the node feature matrix and directional propagation matrix to be located. The node feature matrix and directional propagation matrix to be located are then input into the trained directed graph learning model to output node score vectors. The node score vectors are then input into the Softmax function to obtain the oscillation source probability corresponding to each node. Broadband oscillation source localization is achieved based on the oscillation source probability.
[0070] On the other hand, the present invention also provides a broadband oscillation source localization system based on LCD-Hilbert subband envelope alignment and directed graph learning model, comprising: a computer-readable storage medium and a processor;
[0071] The computer-readable storage medium is used to store executable instructions;
[0072] The processor is used to read executable instructions stored in the computer-readable storage medium and execute the above method.
[0073] On the other hand, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the substation hard plate status monitoring method as described in any one of claims 1-7.
[0074] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects:
[0075] This invention, for the first time, combines the signal extraction method of local feature scale decomposition and Hilbert transform algorithm with a directed graph learning model. It constructs two feature matrices that can simultaneously reflect the oscillation intensity of the node itself, the power grid topology, and the oscillation propagation direction between nodes as inputs to the graph learning model, thus building a novel broadband oscillation source localization method. This method solves the problems of signal mode mixing, noise sensitivity, poor mechanism interpretability, and strong black box nature of existing technologies. It can achieve rapid and accurate localization of broadband oscillation sources, providing strong technical support for the safe and stable operation of the power grid, and has significant theoretical value and engineering application prospects. Attached Figure Description
[0076] Figure 1 The flowchart illustrates the implementation of a broadband oscillation source localization method based on LCD-Hilbert sub-band envelope alignment and directed graph learning model, as provided in this embodiment of the invention.
[0077] Figure 2 This is a schematic diagram of the 12-node wideband oscillation simulation system used in this embodiment of the invention.
[0078] Figure 3 This is a graph showing the changes in training loss value and validation loss value for the method of this invention.
[0079] Figure 4 This is a probability output diagram of oscillation sources at each monitoring node for a typical test sample using the method of the present invention.
[0080] Figure 5 This is a confusion matrix diagram for locating broadband oscillation sources on a test set using the method of the present invention. Detailed Implementation
[0081] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0082] This invention provides a broadband oscillation source localization method based on LCD-Hilbert sub-band envelope alignment and directed graph learning model, such as... Figure 1 As shown, the proposed method specifically includes:
[0083] S1. Acquire synchronous measurement signals from multiple monitoring nodes in the power system, perform time alignment processing on the measurement signals of each node, and obtain the signal set to be analyzed.
[0084] Specifically, the synchronization measurement signal is the active power measurement signal of each monitoring node, and the synchronization measurement signal has a unified timestamp to achieve synchronization alignment between different monitoring nodes.
[0085] S2. Using local feature scale decomposition LCD and Hilbert transform, sub-band decomposition is performed on the signal set to be analyzed, the envelope sequence of each monitoring node under different sub-bands is extracted, and the node time-frequency matrix is constructed.
[0086] Specifically, the process of constructing the node time-frequency matrix in S2 includes:
[0087] S21. The original signal is decomposed into intrinsic scale components (ISC components) using the LCD algorithm. When the number of extreme points in the residual components of the previous decomposition is less than two, or the amplitude of the residual components is less than the threshold, the LCD decomposition is stopped, and each ISC component is filtered using a bandpass filtering algorithm.
[0088] S22. Based on the obtained ISC components, perform a Hibert transform on each ISC component to obtain the instantaneous frequency and instantaneous amplitude signals. For each sampling point, accumulate the amplitude of each ISC at that point to the corresponding sub-band according to its instantaneous frequency, to obtain the envelope sequence of node i in sub-band b. The envelope sequences of each sub-band are then arranged according to frequency band and time to form a node time-frequency matrix A. i :
[0089]
[0090] Where, matrix A iThe rows correspond to different sub-frequency bands, and the columns correspond to sampling times. This matrix centrally represents the oscillation evolution information of node i in each frequency band.
[0091] S3. Based on the sub-band envelope sequence of each monitoring node and the singular value distribution of the node time-frequency matrix, construct a node characteristic matrix that characterizes the local oscillation characteristics of the node.
[0092] Specifically, the construction process of the node feature matrix in S3 includes:
[0093] S31. Define Nodes In sub-band The envelope energy is:
[0094]
[0095] in, For nodes In sub-band The envelope sequence below, To analyze data length;
[0096] S32. Define Nodes The total intensity of the local oscillation is:
[0097]
[0098] in, This represents the total number of sub-bands.
[0099] S33. Further, define the nodes. In sub-band The energy percentage below is:
[0100]
[0101] in, It is a very small positive number;
[0102] S34. Time-frequency matrix of nodes Perform singular value decomposition and define nodes. The modal concentration is:
[0103]
[0104] in, The maximum singular value of the node time-frequency matrix. Let be the rank of the node time-frequency matrix;
[0105] S35. Concatenate the feature vectors of each node vertically to construct the node feature matrix. .
[0106]
[0107] S4. At the main station, obtain the adjacency mask matrix that reflects the node relationship based on the power grid topology, and calculate the minimum alignment error of the delay envelope of adjacent nodes in both directions under each sub-frequency band; based on the difference in the minimum alignment error in both directions, construct a directional propagation matrix that characterizes the direction and intensity of oscillation propagation.
[0108] Specifically, the construction process of the directional propagation matrix in step S4 includes:
[0109] S41. Construct an adjacency mask matrix based on the power grid topology. When node With nodes When directly connected, ,otherwise ;
[0110] S42. Normalize the envelope sequences of the nodes in each sub-band to obtain normalized envelope sequences. ;
[0111]
[0112] S43. For adjacent nodes and Define nodes Pointing to node In sub-band Below, time delay is The time delay alignment error is:
[0113]
[0114] in, For nodes In sub-band The normalized envelope sequence below, As a candidate time delay, satisfying , The maximum search latency is preset.
[0115] S44. Determine the optimal forward propagation delay based on the magnitude of each delay alignment error sequence, and thereby determine the forward matching degree, specifically as follows:
[0116] The optimal propagation delay is selected from the data point corresponding to the minimum delay alignment error.
[0117]
[0118] Further define the positive matching degree as:
[0119]
[0120] Similarly, the reverse matching degree can be obtained. ,like Significantly greater than ,illustrate" "current envelope" and " The delayed envelope is easier to align, and therefore more consistent with the node. Located at node The oscillation propagates upstream.
[0121] S45. Define nodes based on the difference between positive and negative matching degrees. To the node The intensity of transmission And construct a directional propagation matrix based on the propagation intensity of each node. .
[0122]
[0123] And order: The directional propagation strength between each node forms the directional propagation matrix:
[0124]
[0125] S5. Establish and train a directed graph learning model, wherein the directed graph learning model takes the node feature matrix and the directed propagation matrix corresponding to the broadband oscillation sample as input, and the oscillation source node label as supervision information, and trains to obtain a directed graph learning model for outputting the node score vector.
[0126] Specifically, step S5 includes the training method for the graph learning model and the process of obtaining the node score vectors, which includes:
[0127] S51. Construct a broadband oscillation sample set, which includes historical broadband oscillation records, simulated broadband oscillation samples, or a combination of both. For each broadband oscillation sample, construct the corresponding node feature matrix X and directed propagation matrix P according to steps S1 to S4, and use the oscillation source nodes in the broadband oscillation sample as oscillation source node labels. Use the node feature matrix X and directed propagation matrix P as input to the directed graph learning model, and use the oscillation source node labels as supervision information to train the trainable parameters in the directed graph learning model to obtain the trained directed graph learning model.
[0128] S52. The node feature matrix X obtained in S3 and the directional propagation matrix P obtained in S4 are used as inputs to the graph convolutional network. The node feature matrix X represents the local oscillation characteristics of the nodes, while the directional propagation matrix P represents the propagation relationship between nodes. Through graph convolution operations, the node features and the information propagated by neighboring nodes are fused to obtain the score vector H of each node as an oscillation source.
[0129] The directed graph learning model is preferably a directed weighted graph convolutional network, whose... +1 layer convolution is represented as:
[0130]
[0131] in, This represents the node feature representation of the l-th layer; This is the node feature matrix of the input layer, i.e., the node feature matrix obtained in step 3; This represents the information that is aggregated from the propagation information between nodes to the current node; This represents information propagated outward from the current node; and These are the in-degree matrix and the out-degree matrix, respectively, used to normalize the results of graph convolution operations; and This is a trainable weight matrix used to adjust the influence weights of node features; For non-linear activation functions, the ReLU function is typically used.
[0132] Through graph convolution operations at each layer, the local oscillation characteristics of a node are fused with the propagation information of its neighboring nodes layer by layer. Under the action of the multi-layer graph convolutional network, the score matrix of each node gradually fuses from local information to global propagation information, ultimately obtaining a node score phasor H that can reflect whether a node is an oscillation source.
[0133] S6. After the system experiences broadband oscillation, the actual measurement signal is processed according to steps S1 to S4 to obtain the feature matrix and directional propagation matrix of the node to be located. The feature matrix and directional propagation matrix of the node to be located are input into the trained directed graph learning model. The local oscillation features of the node and the directional propagation information between nodes are fused through multi-layer graph convolution to output the node score vector. The node score vector is then input into the Softmax function to obtain the oscillation source probability corresponding to each node. The broadband oscillation source is located based on the oscillation source probability.
[0134] Specifically, the method for obtaining the oscillation source probability of each node in step S6 is as follows:
[0135] After broadband oscillation occurs, the actual measurement signal is processed according to steps S1 to S4. The obtained node feature matrix and directional propagation matrix are input into the trained graph convolutional model, and the node score vector output by the graph convolutional network is... The value is passed as input to the Softmax function to obtain the probability k of each node acting as an oscillation source. i for:
[0136]
[0137] in, It is the node score vector output by the last layer of the graph convolutional network. Corresponding node The score, The total number of nodes is represented by 0. The score matrix is mapped to the range of 0 to 1. Based on the obtained oscillation source probabilities, the node with the highest oscillation source probability is the broadband oscillation source, thus realizing the broadband oscillation source localization.
[0138] Taking a specific simulation model as an example, the following methods are used: Figure 2 A 12-node broadband oscillation simulation system was used for case study verification. The simulation system comprises two regions: Region 1 includes generators G1 and G2 and transformers T1 and T2; Region 2 includes generators G3 and G4, a permanent magnet direct-drive wind turbine (PMSG), and transformers T3, T4, and T5. The system has 12 nodes in total, and the regions are connected via tie lines, with loads 1 and 2 connected to them. This simulation model can effectively simulate the propagation characteristics of broadband oscillations in a scenario where conventional power sources and new energy sources are mixed, making it suitable for verifying the effectiveness of the method presented in this invention.
[0139] In this embodiment, the frequency range of the broadband oscillation samples is set to 0.02–2000 Hz, the sampling rate is set to 10 kHz, and the duration of a single sample is 2 seconds. Various operating condition samples are constructed, including low-frequency oscillation, subsynchronous oscillation, supersynchronous oscillation, mid-to-high frequency oscillation, and multi-frequency coupled oscillation. 10 dB white noise is superimposed on the samples to improve the method's adaptability to noise disturbance conditions. All samples are divided into training, validation, and test sets in a 7:2:1 ratio. For each set of samples, the node feature matrix and directed propagation matrix are extracted according to steps S1 to S4, and the directed graph learning model is trained using the corresponding oscillation source node labels as supervision information.
[0140] Figure 3 This shows the curves showing the changes in training and validation loss values. (From...) Figure 3 As can be seen, with the increase of training rounds, both types of loss gradually decrease and tend to stabilize, indicating that the constructed directed graph learning model has good convergence and stability.
[0141] The verification was performed using a typical test sample from node 4, representing a real oscillation source. Figure 4 As can be seen, the oscillation source probability of node 4 is 81.2%, which is significantly higher than that of other nodes, indicating that the method of the present invention can accurately identify the actual oscillation source node.
[0142] Statistical analysis was performed on the localization results of all samples in the test set to obtain... Figure 5 The confusion matrix shown is row-normalized to percentages, with the main diagonal representing the percentage of correct positioning and the off-diagonal representing the percentage of incorrect positioning. Figure 5 As can be seen, the diagonal elements are significantly higher than the off-diagonal elements, and the overall positioning accuracy is 98.10%, indicating that the method of the present invention has high positioning accuracy.
[0143] To further illustrate the superiority of the method of this invention, it is compared with the Convolutional Neural Network (CNN) method, the Long Short-Term Memory (LSTM) method, and the Graph Convolutional Network (GCN) method without the introduction of a directional propagation matrix. Experiments were conducted on the same training, validation, and test sets for each method, and the average of five independent experimental results was taken. The results are shown in Table 1.
[0144] Table 1 Performance comparison between the method of this invention and existing methods
[0145]
[0146] As shown in Table 1, compared with other broadband oscillation localization methods, the proposed broadband oscillation source localization method based on LCD-Hilbert sub-band envelope alignment and directed graph learning model has the highest accuracy, verifying the superiority of the method proposed in this invention.
[0147] This invention also provides a broadband oscillator localization device based on LCD-Hilbert sub-band envelope alignment and directed graph learning model, used to execute the broadband oscillator localization method based on LCD-Hilbert sub-band envelope alignment and directed graph learning model as described above, including:
[0148] The signal acquisition module is used to acquire synchronous measurement signals from multiple monitoring nodes in the power system, perform time alignment processing on the measurement signals from each node, and obtain the signal set to be analyzed.
[0149] The time-frequency decomposition module is used to decompose the LCD and Hilbert transform using local feature scale decomposition, perform sub-frequency band decomposition on the signal set to be analyzed, extract the envelope sequence of each monitoring node under different sub-frequency bands, and construct the node time-frequency matrix of each monitoring node.
[0150] The node feature construction module is used to perform singular value decomposition based on the sub-band envelope sequence of each monitoring node and the node time-frequency matrix to construct a node feature matrix that characterizes the local oscillation characteristics of the node.
[0151] The directional propagation construction module is used to construct an adjacency mask matrix according to the power grid topology, calculate the minimum alignment error of the delay envelope of adjacent nodes in both directions under each sub-frequency band based on the sub-frequency band envelope sequence of each monitoring node, and construct a directional propagation matrix characterizing the oscillation propagation direction and propagation intensity based on the difference between the minimum alignment errors of the delay envelope in both directions.
[0152] The model training module is used to establish and train a directed graph learning model. It takes the node feature matrix and the directed propagation matrix corresponding to the broadband oscillation sample as input and the oscillation source node label as supervision information to train a directed graph learning model for outputting node score vectors.
[0153] The oscillation source localization module is used to process the actual measurement signal through the time-frequency decomposition module, node feature construction module, and directional propagation construction module after the system experiences broadband oscillation, to obtain the node feature matrix and directional propagation matrix to be located. The node feature matrix and directional propagation matrix to be located are then input into the trained directed graph learning model to output node score vectors. The node score vectors are then input into the Softmax function to obtain the oscillation source probability corresponding to each node. Broadband oscillation source localization is achieved based on the oscillation source probability.
[0154] This invention also provides a broadband oscillation source localization system based on LCD-Hilbert subband envelope alignment and directed graph learning model, comprising: a computer-readable storage medium and a processor;
[0155] The computer-readable storage medium is used to store executable instructions;
[0156] The processor is used to read executable instructions stored in the computer-readable storage medium and execute the broadband oscillation source localization method based on LCD-Hilbert sub-band envelope alignment and directed graph learning model as described above.
[0157] This invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the substation hard plate status monitoring method as described above.
[0158] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied 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.
[0159] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. 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 illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0160] 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.
[0161] 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.
[0162] This invention deeply integrates LCD-Hilbert time-frequency decomposition with directed graph learning, and has the following characteristics and effects:
[0163] (1) By extracting the sub-band envelope sequence of each node and constructing the time-frequency matrix through LCD-Hilbert, the mode mixing and endpoint effect in traditional mode decomposition can be effectively overcome, and the multi-band time-frequency features of broadband oscillation can be stably extracted.
[0164] (2) Based on the envelope energy ratio and the singular values of the time-frequency matrix, the node characteristic matrix is constructed. The physical meaning is clear and can accurately characterize the local oscillation intensity and mode concentration of each node.
[0165] (3) A directional propagation matrix based on the minimum alignment error of the forward and reverse delay envelope is proposed. The propagation direction and intensity of oscillation between adjacent nodes can be quantified without the need for a preset threshold, providing a clear physical prior for the graph model.
[0166] (4) Construct a directed weighted graph convolutional network that integrates local node features and directional propagation information, combining physical interpretability with the strong fitting ability of deep learning, significantly improving the positioning accuracy and generalization performance under complex oscillation modes.
[0167] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A broadband oscillation source localization method based on LCD-Hilbert sub-band envelope alignment and directed graph learning model, characterized in that, Includes the following steps: S1. Acquire synchronous measurement signals from multiple monitoring nodes in the power system, perform time alignment processing on the measurement signals of each node, and obtain the signal set to be analyzed; S2. Using local feature scale decomposition (LCD) and Hilbert transform, sub-band decomposition is performed on the signal set to be analyzed, the envelope sequence of each monitoring node under different sub-bands is extracted, and the node time-frequency matrix is constructed. S3. Based on the sub-band envelope sequence of each monitoring node and the node time-frequency matrix, perform singular value decomposition to construct a node feature matrix characterizing the local oscillation characteristics of the node; S4. At the main station, obtain the adjacency mask matrix that reflects the node relationship based on the power grid topology, and calculate the minimum alignment error of the delay envelope of adjacent nodes in both directions under each sub-frequency band; based on the difference of the minimum alignment error of the delay envelope in both directions, construct a directional propagation matrix that characterizes the direction and intensity of oscillation propagation. S5. Establish and train a directed graph learning model, wherein the directed graph learning model takes the node feature matrix and the directed propagation matrix corresponding to the broadband oscillation sample as input, and the oscillation source node label as supervision information, and trains to obtain a directed graph learning model for outputting node score vectors; S6. After the system experiences broadband oscillation, the actual measurement signal is processed according to steps S1 to S4 to obtain the feature matrix and directional propagation matrix of the node to be located. The feature matrix and directional propagation matrix of the node to be located are input into the trained directed graph learning model. The local oscillation features of the node and the directional propagation information between nodes are fused through multi-layer graph convolution to output the node score vector. The node score vector is then input into the Softmax function to obtain the oscillation source probability corresponding to each node. The broadband oscillation source is located based on the oscillation source probability.
2. The broadband oscillation source localization method based on LCD-Hilbert sub-band envelope alignment and directed graph learning model according to claim 1, characterized in that, The synchronization measurement signal in S1 is the active power measurement signal of each monitoring node. The synchronization measurement signal has a unified timestamp to achieve synchronization alignment between different monitoring nodes.
3. The broadband oscillation source localization method based on LCD-Hilbert sub-band envelope alignment and directed graph learning model according to claim 1, characterized in that, The process of constructing the node time-frequency matrix in S2 includes: S21. Use the LCD algorithm to decompose the original signal into ISC components. Stop the LCD decomposition when the number of extreme points in the residual components of the previous decomposition is less than two, or the amplitude of the residual components is less than the threshold. Then use the bandpass filtering algorithm to filter each ISC component. S22. Based on the obtained ISC components, perform a Hibert transform on each ISC component to obtain the instantaneous frequency and instantaneous amplitude signals. For each sampling point, accumulate the amplitude of each ISC component to the corresponding sub-band according to its instantaneous frequency at that point to obtain the envelope sequence of node i in sub-band b. The envelope sequences of each sub-band are then arranged according to frequency band and time to form a node time-frequency matrix A. i : ; Where, matrix A i The rows correspond to different sub-frequency bands, and the columns correspond to sampling times. This matrix centrally represents the oscillation evolution information of node i in each frequency band.
4. The broadband oscillation source localization method based on LCD-Hilbert sub-band envelope alignment and directed graph learning model according to claim 1, characterized in that, The construction process of the node feature matrix in S3 includes: S31. Define Nodes In sub-band The envelope energy is: ; in, For nodes In sub-band The envelope sequence below, To analyze data length; S32. Define Nodes The total intensity of the local oscillation is: ; in, This represents the total number of sub-bands. S33. Define Nodes In sub-band The energy percentage below is: ; in, It is a very small positive number; S34. Time-frequency matrix of nodes Perform singular value decomposition and define nodes. The modal concentration is: ; in, The maximum singular value of the node time-frequency matrix. Let be the rank of the node time-frequency matrix; S35. Concatenate the feature vectors of each node vertically to construct the node feature matrix. ; 。 5. The broadband oscillation source localization method based on LCD-Hilbert sub-band envelope alignment and directed graph learning model according to claim 1, characterized in that, The construction process of the directional propagation matrix in step S4 includes: S41. Construct an adjacency mask matrix based on the power grid topology. When node With nodes When directly connected, ,otherwise ; S42. Normalize the envelope sequences of the nodes in each sub-band to obtain normalized envelope sequences. ; ; S43. For adjacent nodes and Define nodes Pointing to node In sub-band Below, time delay is The time delay alignment error is: ; in, For nodes In sub-band The normalized envelope sequence below, As a candidate time delay, satisfying , The maximum search latency is preset. S44. Determine the optimal forward propagation delay based on the magnitude of each delay alignment error sequence, and thereby determine the forward matching degree, specifically as follows: The optimal propagation delay is selected from the data point corresponding to the minimum delay alignment error. ; Further define the positive matching degree as: ; Similarly, the reverse matching degree is obtained. ,like Significantly greater than ,illustrate" "current envelope" and " The delayed envelope is easier to align, and therefore more suitable for nodes. Located at node The oscillation propagates upstream; S45. Define nodes based on the difference between positive and negative matching degrees. To the node The intensity of transmission And construct a directional propagation matrix based on the propagation intensity of each node. : ; And order: The directional propagation strength between each node forms the directional propagation matrix: 。 6. The broadband oscillation source localization method based on LCD-Hilbert sub-band envelope alignment and directed graph learning model according to claim 1, characterized in that, The training method for the directed graph learning model and the process of obtaining the node score vectors in step S5 include: S51. Construct a broadband oscillation sample set, which includes historical broadband oscillation records, simulated broadband oscillation samples, or a combination of both; for each broadband oscillation sample, construct the corresponding node feature matrix X and directed propagation matrix P according to steps S1 to S4, and use the oscillation source nodes in the broadband oscillation sample as oscillation source node labels; use the node feature matrix X and directed propagation matrix P as input to the directed graph learning model, and use the oscillation source node labels as supervision information to train the trainable parameters in the directed graph learning model to obtain the trained directed graph learning model; S52. The node feature matrix X obtained in S3 and the directional propagation matrix P obtained in S4 are used as inputs to the graph convolutional network. The node feature matrix X represents the local oscillation characteristics of the nodes, while the directional propagation matrix P represents the propagation relationship between nodes. Through graph convolution operation, the node features and the information propagated by neighboring nodes are fused to obtain the score vector H of each node as an oscillation source; The directed graph learning model is preferably a directed weighted graph convolutional network, whose... +1 layer convolution is represented as: ; in, This represents the node feature representation of the l-th layer; This is the node feature matrix of the input layer, i.e., the node feature matrix obtained in step 3; This represents the information that is aggregated from the propagation information between nodes to the current node; This represents information propagated outward from the current node; and These are the in-degree matrix and the out-degree matrix, respectively, used to normalize the results of graph convolution operations; and This is a trainable weight matrix used to adjust the influence weights of node features; For non-linear activation functions, the ReLU function is typically used; Through graph convolution operations at each layer, the local oscillation characteristics of a node are fused with the propagation information of its neighboring nodes layer by layer. Under the action of the multi-layer graph convolutional network, the score matrix of each node gradually fuses from local information to global propagation information, ultimately yielding a node score phasor H that reflects whether a node is an oscillation source.
7. The broadband oscillation source localization method based on LCD-Hilbert sub-band envelope alignment and directed graph learning model according to claim 1, characterized in that, The method for obtaining the probability of each node's oscillation source in step S6 is as follows: After broadband oscillation occurs, the actual measurement signal is processed according to steps S1 to S4. The obtained node feature matrix and directional propagation matrix are input into the trained graph convolutional model, and the node score vector output by the graph convolutional network is... The value is passed as input to the Softmax function to obtain the probability k of each node acting as an oscillation source. i for: ; in, It refers to the output features of the last layer of a graph convolutional network. Corresponding node eigenvalues, The total number of nodes is represented by 0. The score matrix is mapped to the range of 0 to 1. Based on the obtained oscillation source probabilities, the node with the highest oscillation source probability is the broadband oscillation source, thus realizing the broadband oscillation source localization.
8. A broadband oscillator localization device based on LCD-Hilbert sub-band envelope alignment and directed graph learning model, used to execute the broadband oscillator localization method based on LCD-Hilbert sub-band envelope alignment and directed graph learning model as described in any one of claims 1-7, characterized in that, include: The signal acquisition module is used to acquire synchronous measurement signals from multiple monitoring nodes in the power system, perform time alignment processing on the measurement signals from each node, and obtain the signal set to be analyzed. The time-frequency decomposition module is used to decompose the LCD and Hilbert transform using local feature scale decomposition, perform sub-frequency band decomposition on the signal set to be analyzed, extract the envelope sequence of each monitoring node under different sub-frequency bands, and construct the node time-frequency matrix of each monitoring node. The node feature construction module is used to perform singular value decomposition based on the sub-band envelope sequence of each monitoring node and the node time-frequency matrix to construct a node feature matrix that characterizes the local oscillation characteristics of the node. The directional propagation construction module is used to construct an adjacency mask matrix according to the power grid topology, calculate the minimum alignment error of the delay envelope of adjacent nodes in both directions under each sub-frequency band based on the sub-frequency band envelope sequence of each monitoring node, and construct a directional propagation matrix characterizing the oscillation propagation direction and propagation intensity based on the difference between the minimum alignment errors of the delay envelope in both directions. The model training module is used to establish and train a directed graph learning model. It takes the node feature matrix and the directed propagation matrix corresponding to the broadband oscillation sample as input and the oscillation source node label as supervision information to train a directed graph learning model for outputting node score vectors. The oscillation source localization module is used to process the actual measurement signal through the time-frequency decomposition module, node feature construction module, and directional propagation construction module after the system experiences broadband oscillation, to obtain the node feature matrix and directional propagation matrix to be located. The node feature matrix and directional propagation matrix to be located are then input into the trained directed graph learning model to output node score vectors. The node score vectors are then input into the Softmax function to obtain the oscillation source probability corresponding to each node. Broadband oscillation source localization is achieved based on the oscillation source probability.
9. A broadband oscillation source localization system based on LCD-Hilbert sub-band envelope alignment and directed graph learning model, characterized in that, include: Computer-readable storage media and processors; The computer-readable storage medium is used to store executable instructions; The processor is used to read executable instructions stored in the computer-readable storage medium and execute the broadband oscillation source localization method based on LCD-Hilbert sub-band envelope alignment and directed graph learning model as described in any one of claims 1-7.
10. A non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the substation hard plate status monitoring method as described in any one of claims 1-7.
