A method and system for hierarchical positioning of power system wide-area oscillation disturbance sources

Abstract

The present application relates to the technical field of power system, and more particularly to a power system wide frequency oscillation disturbance source hierarchical positioning method and system, comprising calculating the compression ratio of the measurement signal according to the sampling resolution and real-time transmission interval of the original measurement signal to obtain an initial compression coefficient; performing frequency domain decomposition on the original measurement signal to obtain spectral feature data, and compressing the original measurement signal based on the spectral feature data and the initial compression coefficient to obtain a measurement compressed signal; embedding a time sequence position code in the measurement compressed signal to generate time sequence input data; analyzing the time sequence input data using a multi-head attention mechanism to obtain an attention weighted signal; generating a reconstruction signal according to the attention weighted signal, and obtaining a wide frequency oscillation disturbance source hierarchical positioning result based on the reconstruction signal and the disturbance propagation path. The present application realizes fast and accurate hierarchical positioning of the wide frequency oscillation disturbance source through adaptive compression to retain wide frequency oscillation characteristics and topology propagation path analysis and other technologies.

Description

Technical Field

[0001] The present invention relates to the technical field of power systems, and in particular to a hierarchical positioning method and system for broadband oscillation disturbance sources in power systems. Background Art

[0002] During the safe operation of power systems, the location of broadband oscillation disturbance sources is a key link in ensuring the safe and stable operation of the power grid. With the continuous expansion of the scale and increasing complexity of the power grid, how to quickly and accurately locate broadband oscillation disturbance sources under limited bandwidth conditions has become an important issue that needs to be urgently addressed in the operation and maintenance of power systems.

[0003] In practical application scenarios, the location of broadband oscillation disturbance sources faces many challenges. On the one hand, the sampling frequency of measurement signals in power systems is usually high, but due to limited communication bandwidth, the data upload frequency often cannot meet real-time requirements, resulting in a large amount of key information being lost during transmission. Although compression algorithms can be used to reduce the amount of data to adapt to bandwidth limitations, traditional compression methods often inevitably lose key features in the signal while reducing the data volume, making it difficult to restore the true state of the original signal during reconstruction and analysis of the compressed data; on the other hand, the complexity and dynamic nature of the power grid topology further exacerbate the difficulty of fault source location. In the process of locating broadband oscillation disturbance sources, it is often difficult to accurately determine the location of the disturbance source by relying solely on information from local nodes. Due to the differences in dynamic characteristics of different areas in the power grid, the disturbance will be affected by the network topology during propagation, resulting in the need to comprehensively consider global information when locating the disturbance source. Especially in scenarios with multi-source disturbances or distributed energy access, the positioning accuracy and efficiency of existing technologies are difficult to meet actual needs.

[0004] To sum up, the existing technology has many shortcomings in locating the source of broadband oscillation disturbances. How to meet the needs of safe operation of the power grid in the context of a high proportion of new energy access under limited bandwidth conditions, and how to achieve rapid and accurate positioning of broadband oscillation disturbance sources have become technical problems that need to be solved urgently in the current power system field. Summary of the Invention

[0005] In order to solve the above technical problems, the present invention provides a method and system for hierarchical positioning of broadband oscillation disturbance sources in a power system.

[0006] In a first aspect, the present invention provides a method for hierarchically locating a broadband oscillation disturbance source in a power system, the method comprising the following steps:

[0007] Obtaining the original measurement signal of the power system monitoring equipment, calculating the compression ratio of the measurement signal according to the sampling resolution of the original measurement signal and the real-time transmission interval, and obtaining the initial compression coefficient;

[0008] Performing frequency domain decomposition on the original measurement signal to obtain frequency spectrum characteristic data including oscillation frequency distribution characteristics, and compressing the original measurement signal based on the frequency spectrum characteristic data and the initial compression coefficient to obtain a measurement compression signal;

[0009] Embedding a time series position code in the measurement compression signal to generate time series input data with a time evolution trend;

[0010] Utilizing a multi-head attention mechanism to analyze the oscillation amplitude changes in different time windows in the time series input data, and obtaining an attention-weighted signal with a disturbance type label;

[0011] Generate a reconstructed signal through a linear layer transformation according to the attention weighted signal, and determine the scope of the preliminary disturbance source area based on the reconstructed signal and the disturbance propagation path;

[0012] A reconstruction error between the reconstructed signal and the original measurement signal is analyzed within the preliminary disturbance source region to obtain a hierarchical positioning result of a broadband oscillation disturbance source.

[0013] In a further embodiment, the step of calculating the compression ratio of the measurement signal according to the sampling resolution of the original measurement signal and the real-time transmission interval to obtain the initial compression coefficient includes:

[0014] Obtaining a sampling resolution and a sampling point density of the original measurement signal, and calculating a ratio of the sampling resolution to the sampling point density to obtain a density ratio parameter;

[0015] Obtaining a data upload frequency and a real-time transmission interval of the original measurement signal, and calculating a ratio of the data upload frequency to the real-time transmission interval to obtain a frequency control factor;

[0016] Calculating an initial compression ratio value based on the density ratio parameter and the frequency control factor, and comparing the initial compression ratio value with a preset compression upper limit threshold;

[0017] If the initial compression ratio value exceeds a preset compression upper limit threshold, the sampling resolution is reversely adjusted using the initial compression ratio value to obtain a sampling resolution update value;

[0018] The density ratio parameter is recalculated based on the sampling resolution update value to obtain a density ratio parameter update value, and an initial compression coefficient is obtained based on the density ratio parameter update value and a real-time transmission interval.

[0019] In a further embodiment, the step of performing frequency domain decomposition on the original measurement signal to obtain frequency spectrum feature data including oscillation frequency distribution characteristics comprises:

[0020] The original measurement signal is converted into a measurement frequency domain signal by fast Fourier transform, and the measurement frequency domain signal is divided into frequency bands by using a bandpass filter based on the wide-band oscillation frequency distribution range to obtain a normalized frequency band spectrum;

[0021] A multi-head attention mechanism is used to extract the spatiotemporal distribution patterns of oscillation energy in different frequency bands from the normalized frequency band spectrum to obtain corresponding frequency band feature vectors;

[0022] Performing dimensionality reduction processing on the frequency band feature vector to obtain a frequency band reduced dimensionality feature vector, and inputting the frequency band reduced dimensionality feature vector into a pre-trained autoencoder to reconstruct the frequency band spectrum to obtain a reconstructed feature vector;

[0023] The reconstructed feature vectors of different frequency bands are arranged into row vectors in ascending order of frequency to construct a two-dimensional spectrum feature matrix, and the two-dimensional spectrum feature matrix is ​​normalized to obtain spectrum feature data.

[0024] In a further embodiment, the step of compressing the original measurement signal based on the spectrum characteristic data and the initial compression coefficient to obtain the measurement compression signal includes:

[0025] According to the energy proportion of each frequency band in the spectrum feature data and the initial compression coefficient, a frequency band feature weight vector of each frequency band is obtained, and a feedforward neural network including three hidden layers is constructed;

[0026] The frequency band feature weight vector is input as the bias condition of the hidden layer, the original measurement signal is input into the trained feedforward neural network, and the signal dimension is compressed through the nonlinear transformation of the hidden layer to generate a measurement compression signal.

[0027] In a further embodiment, the number of nodes in each hidden layer of the feedforward neural network decreases exponentially according to the initial compression coefficient.

[0028] In a further embodiment, the step of analyzing the oscillation amplitude changes in different time windows in the time series input data using a multi-head attention mechanism to obtain an attention weighted signal with a disturbance type label includes:

[0029] According to the broadband oscillation characteristics of the power system, the dominant mode period of electromechanical oscillation is taken as the basic oscillation period;

[0030] Using a multi-head attention mechanism to process the time series input data in parallel, and using the basic oscillation period to divide the time series input data into multiple time window segments;

[0031] Calculate the amplitude change rate, energy accumulation and phase difference absolute value in each time window segment based on the time series input data, and generate an attention key value matrix for each time window segment;

[0032] Based on the pre-acquired load disturbance feature template, the attention key value matrix of each time window segment is matched with the disturbance feature template to obtain the attention weight vector of each time window segment;

[0033] The time series input data is weightedly fused according to the attention weight vector, and an attention weighted signal with a disturbance type label is output.

[0034] In a further embodiment, the length of each time window segment is set to an integer multiple of the basic oscillation period.

[0035] In a further embodiment, the step of determining the scope of the preliminary disturbance source region based on the reconstructed signal and the disturbance propagation path includes:

[0036] The electrical nodes of the power grid topology are analyzed through a graph attention network to extract node feature vectors that combine electrical coupling strength and power transmission direction.

[0037] Calculate the propagation path weight of each electrical node in the disturbance propagation path based on the node eigenvector and the oscillation propagation delay characteristics of the reconstructed signal;

[0038] Screening out a set of key nodes whose propagation path weight is greater than a preset area division threshold from the electrical nodes in the power grid topology structure, and removing isolated nodes and merging adjacent node groups based on the spatial distribution density of the key node set to form a candidate key node cluster;

[0039] Integrating the edge connection weights within the candidate key node cluster to calculate the disturbance propagation path density per unit area, and dividing the candidate key node cluster into different disturbance propagation levels according to the disturbance propagation path density;

[0040] According to the principle of shortest electrical distance, the connected domain analysis algorithm is used to merge the overlapping areas of different disturbance propagation levels to form a continuous preliminary disturbance source area range.

[0041] In a further embodiment, the step of analyzing the reconstruction error between the reconstructed signal and the original measurement signal within the preliminary disturbance source region to obtain a hierarchical positioning result of the broadband oscillation disturbance source comprises:

[0042] Calculating a mean square error between the reconstructed signal and the original measurement signal in a preset frequency band within the preliminary disturbance source region, and screening out high-error electrical nodes from electrical nodes in the power grid topology structure based on the mean square error;

[0043] Based on the principle of shortest electrical distance, the boundary of the disturbance propagation level area is expanded to the minimum connected area covering all high-error electrical nodes to obtain the correction range of the disturbance source area;

[0044] Extracting attenuation characteristic parameters of high-error electrical nodes within the correction range of the disturbance source region, and matching the attenuation characteristic parameters with a pre-acquired device-level attenuation feature library to obtain an attenuation characteristic matching degree;

[0045] Screening out high-error electrical nodes whose attenuation characteristic matching degree is higher than a preset attenuation threshold as core candidate nodes, and performing spatial density clustering on the core candidate nodes to generate a core disturbance area;

[0046] Analyzing the spatial consistency of the phase propagation direction and the amplitude attenuation gradient direction of the reconstructed signal within the core perturbation region to generate a consistency score;

[0047] The core disturbance area is divided according to the consistency score to obtain hierarchical positioning areas, and the hierarchical positioning areas are associated with a device coordinate library of a power grid topology structure to obtain hierarchical positioning results of broadband oscillation disturbance sources.

[0048] In a second aspect, the present invention provides a hierarchical positioning system for broadband oscillation disturbance sources in a power system, the system comprising:

[0049] The compression analysis module is used to obtain the original measurement signal of the power system monitoring equipment, calculate the compression ratio of the measurement signal according to the sampling resolution of the original measurement signal and the real-time transmission interval, and obtain the initial compression coefficient;

[0050] a signal compression module, configured to perform frequency domain decomposition on the original measurement signal to obtain frequency spectrum characteristic data including oscillation frequency distribution characteristics, and compress the original measurement signal based on the frequency spectrum characteristic data and the initial compression coefficient to obtain a measurement compression signal;

[0051] A time series embedding module, configured to embed a time series position code into the measurement compression signal to generate time series input data with a time evolution trend;

[0052] A signal analysis module is used to analyze the oscillation amplitude changes in different time windows in the time series input data using a multi-head attention mechanism to obtain an attention-weighted signal with a disturbance type label;

[0053] A disturbance analysis module is used to generate a reconstructed signal through a linear layer transformation according to the attention weighted signal, and determine the scope of the preliminary disturbance source area based on the reconstructed signal and the disturbance propagation path;

[0054] The hierarchical positioning module is used to analyze the reconstruction error between the reconstructed signal and the original measurement signal within the preliminary disturbance source area to obtain a hierarchical positioning result of the broadband oscillation disturbance source.

[0055] The present invention provides a method and system for hierarchical positioning of broadband oscillation disturbance sources in an electric power system. The method obtains an original measurement signal from an electric power system monitoring device, calculates the compression ratio of the measurement signal based on the sampling resolution of the original measurement signal and the real-time transmission interval, and obtains an initial compression coefficient; performs frequency domain decomposition on the original measurement signal to obtain spectrum feature data containing oscillation frequency distribution characteristics, and compresses the original measurement signal based on the spectrum feature data and the initial compression coefficient to obtain a measurement compression signal; embeds a time series position code in the measurement compression signal to generate time series input data with a time evolution trend; utilizes a multi-head attention mechanism to analyze the oscillation amplitude changes in different time windows in the time series input data to obtain an attention weighted signal with a disturbance type label; generates a reconstructed signal based on the attention weighted signal through a linear layer transformation, and determines the preliminary disturbance source area range based on the reconstructed signal and the disturbance propagation path; analyzes the reconstruction error between the reconstructed signal and the original measurement signal within the preliminary disturbance source area, and obtains a hierarchical positioning result of the broadband oscillation disturbance source. Compared with existing technologies, this method achieves rapid and accurate hierarchical positioning of broadband oscillation disturbance sources in complex power grid environments through technologies such as adaptive compression to retain broadband oscillation characteristics and topological propagation path analysis, significantly improving positioning accuracy and real-time performance, and providing technical support for power grid fault diagnosis and stable operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 This is a flow chart of a hierarchical location method for broadband oscillation disturbance sources in a power system provided by an embodiment of the present invention;

[0057] Figure 2 This is a block diagram of a hierarchical positioning system for broadband oscillation disturbance sources in a power system provided by an embodiment of the present invention.

[0058] Explanation of the accompanying reference numerals: 101, compression analysis module; 102, signal compression module; 103, timing embedding module; 104, signal analysis module; 105, disturbance analysis module; 106, hierarchical positioning module. DETAILED DESCRIPTION

[0059] The following describes the embodiments of the present invention in detail with reference to the accompanying drawings. The embodiments are provided for illustrative purposes only and are not to be construed as limiting the present invention. The accompanying drawings are provided for reference and illustration only and do not constitute a limitation on the scope of protection of the present invention. Many changes may be made to the present invention without departing from the spirit and scope of the present invention.

[0060] refer to Figure 1, the embodiment of the present invention provides a hierarchical positioning method for broadband oscillation disturbance sources in a power system, such as Figure 1 As shown, the method includes the following steps:

[0061] S1. Obtain the original measurement signal of the power system monitoring equipment, calculate the compression ratio of the measurement signal according to the sampling resolution of the original measurement signal and the real-time transmission interval, and obtain the initial compression coefficient.

[0062] In some embodiments, the step of calculating the compression ratio of the measurement signal according to the sampling resolution and the real-time transmission interval of the original measurement signal to obtain the initial compression coefficient includes:

[0063] Obtaining a sampling resolution and a sampling point density of the original measurement signal, and calculating a ratio of the sampling resolution to the sampling point density to obtain a density ratio parameter;

[0064] Obtaining a data upload frequency and a real-time transmission interval of the original measurement signal, and calculating a ratio of the data upload frequency to the real-time transmission interval to obtain a frequency control factor;

[0065] Calculating an initial compression ratio value based on the density ratio parameter and the frequency control factor, and comparing the initial compression ratio value with a preset compression upper limit threshold;

[0066] If the initial compression ratio value exceeds a preset compression upper limit threshold, the sampling resolution is reversely adjusted using the initial compression ratio value to obtain a sampling resolution update value;

[0067] The density ratio parameter is recalculated based on the sampling resolution update value to obtain a density ratio parameter update value, and an initial compression coefficient is obtained based on the density ratio parameter update value and a real-time transmission interval.

[0068] Specifically, this embodiment obtains the sampling resolution (data accuracy of a single sampling point per unit time) and the real-time transmission interval (the minimum time difference between two adjacent data uploads) of the original measurement signal uploaded by the power system monitoring equipment, and calculates the number of effective sampling points per unit time as the sampling point density. The sampling point density reflects the sampling granularity of the time dimension. The original measurement signal may include at least one of the voltage signal, current signal, power signal and frequency signal collected in real time by the power system monitoring equipment. For example, the voltage signal may include an instantaneous value sequence of the fundamental voltage, harmonic components and broadband oscillation components, and the current signal may include an instantaneous value sequence of amplitude, phase and frequency fluctuation characteristics. The power signal may include transient fluctuation data of active power and reactive power, and the frequency signal may include frequency deviation and oscillation modal components within the range of 10 Hz to 2 kHz. This embodiment performs a ratio operation on the sampling resolution and the sampling point density to obtain a density ratio parameter representing the sampling information density. A frequency control factor is then calculated based on the data upload frequency (the number of data packets allowed to be uploaded per unit time) of the original measurement signal and the real-time transmission interval. The frequency control factor is equal to the data upload frequency divided by the inverse of the real-time transmission interval. This embodiment uses the frequency control factor to quantify the communication bandwidth constraint into a calculable physical quantity, which is used to quantify the strength of the communication channel's restriction on the data volume.

[0069] Next, this embodiment multiplies the density ratio parameter by the frequency control factor to obtain an initial compression ratio value, and compares the initial compression ratio value with a preset compression upper limit threshold. If the initial compression ratio value exceeds the compression upper limit threshold, the dynamic adjustment link is entered, and the sampling resolution is reversely adjusted according to the initial compression ratio value that exceeds the compression upper limit threshold. Specifically, the updated sampling resolution is generated by multiplying the sampling resolution by the quotient of the compression upper limit threshold and the initial compression ratio value. This embodiment can avoid the loss of high-frequency oscillation characteristics due to excessive compression through the threshold judgment and reverse adjustment mechanism. This embodiment recalculates the density ratio parameter based on the updated sampling resolution to obtain an updated density ratio parameter value, and normalizes the updated density ratio parameter value with the real-time transmission interval to finally generate an initial compression coefficient. The value of the initial compression coefficient is equal to the updated density ratio parameter value divided by the real-time transmission interval. The updated density ratio parameter value simultaneously meets the communication bandwidth limitation and the key feature retention requirements, and is used to guide the subsequent signal compression process.

[0070] S2. Perform frequency domain decomposition on the original measurement signal to obtain frequency spectrum feature data including oscillation frequency distribution characteristics, and compress the original measurement signal based on the frequency spectrum feature data and the initial compression coefficient to obtain a measurement compression signal.

[0071] In some embodiments, the step of performing frequency domain decomposition on the original measurement signal to obtain frequency spectrum feature data including oscillation frequency distribution characteristics includes:

[0072] The original measurement signal is converted into a measurement frequency domain signal by fast Fourier transform, and the measurement frequency domain signal is divided into frequency bands by using a bandpass filter based on the wide-band oscillation frequency distribution range to obtain a normalized frequency band spectrum;

[0073] A multi-head attention mechanism is used to extract the spatiotemporal distribution patterns of oscillation energy in different frequency bands from the normalized frequency band spectrum to obtain corresponding frequency band feature vectors;

[0074] Performing dimensionality reduction processing on the frequency band feature vector to obtain a frequency band reduced dimensionality feature vector, and inputting the frequency band reduced dimensionality feature vector into a pre-trained autoencoder to reconstruct the frequency band spectrum to obtain a reconstructed feature vector;

[0075] The reconstructed feature vectors of different frequency bands are arranged into row vectors in ascending order of frequency to construct a two-dimensional spectrum feature matrix, and the two-dimensional spectrum feature matrix is ​​normalized to obtain spectrum feature data.

[0076] Specifically, based on the requirements for monitoring broadband oscillations in power systems, this embodiment converts the original measurement signal into a measurement frequency domain signal through a fast Fourier transform. The measurement frequency domain signal is then amplitude-normalized to eliminate dimensional differences between different monitoring devices, generating a standardized measurement frequency domain signal with zero mean and unit variance. Based on the typical frequency distribution range of broadband oscillations, this embodiment uses a three-stage bandpass filter to divide the standardized measurement frequency domain signal into a low-frequency band (0.1-10 Hz), a mid-frequency band (10-100 Hz), and a high-frequency band (100-2000 Hz). The amplitude envelope and phase offset of each frequency band are extracted. The low-frequency band corresponds to the electromechanical oscillation mode, the mid-frequency band covers subsynchronous oscillations, and the high-frequency band includes switching harmonics of power electronic equipment. This embodiment then normalizes the amplitude of each frequency band, linearly mapping it to the range of 0 to 1, generating a normalized frequency band spectrum that can be compared across frequency bands.

[0077] For each normalized frequency band spectrum, this embodiment adopts a multi-head attention mechanism to parallelly calculate the energy gradient correlation of different frequency points, where each attention head focuses on the amplitude mutation characteristics of a specific frequency band, and captures the coupling characteristics across frequency bands through multi-head parallel calculation, such as the interaction phenomenon of low frequency modulating high frequency. This embodiment generates a 512-dimensional frequency band feature vector by splicing the output features of each attention head. The frequency band feature vector characterizes the spatiotemporal distribution pattern of the oscillation energy in the frequency band. At the same time, this embodiment performs principal component analysis on the 512-dimensional frequency band feature vector, retains the first 128 principal components in the direction of maximum variance, eliminates the linear correlation between features, and obtains the frequency band dimensionality reduction feature vector. The segment dimensionality reduction feature vector is input into the pre-trained autoencoder, and the frequency band spectrum is reconstructed through the encoder-decoder structure. The network parameters are optimized with the goal of minimizing the mean square error, and the reconstructed feature vector with noise suppression capability is output. Finally, this embodiment arranges the reconstructed feature vectors of the three frequency bands of low frequency, medium frequency, and high frequency into row vectors in ascending order of frequency, constructs a two-dimensional spectrum feature matrix, and standardizes the features of each column of the two-dimensional spectrum feature matrix to make it conform to the zero mean and unit standard deviation distribution, and finally generates standardized spectrum feature data containing the oscillation frequency distribution characteristics, wherein the row vector of the standardized spectrum feature data corresponds to the frequency band division, and the column vector represents the time-varying oscillation energy distribution.

[0078] In some embodiments, the step of compressing the original measurement signal based on the spectrum feature data and the initial compression coefficient to obtain a measurement compression signal includes:

[0079] According to the energy proportion of each frequency band in the spectrum feature data and the initial compression coefficient, a frequency band feature weight vector of each frequency band is obtained, and a feedforward neural network including three hidden layers is constructed;

[0080] The frequency band feature weight vector is input as the bias condition of the hidden layer, the original measurement signal is input into the trained feedforward neural network, and the signal dimension is compressed through the nonlinear transformation of the hidden layer to generate a measurement compression signal.

[0081] Specifically, this embodiment extracts the energy proportions of the low frequency band (0.1-10Hz), medium frequency band (10-100Hz) and high frequency band (100-2000Hz) in the spectrum feature data, which are respectively recorded as the low frequency energy proportion, the medium frequency energy proportion and the high frequency energy proportion, and the frequency band feature weights of the low frequency band, the medium frequency band and the high frequency band are calculated according to the energy proportion of each frequency band in the spectrum feature data, wherein the low frequency band feature weight is equal to the product of the low frequency energy proportion and the initial compression coefficient, ensuring the stable retention of low frequency modes such as electromechanical oscillations; the medium frequency band feature weight is equal to the product of the medium frequency energy proportion and the square root of the initial compression coefficient, and the medium frequency band adopts the square root relationship to average the energy proportion. The compression rate and subsynchronous oscillation feature retention requirements are balanced; the high-frequency band feature weight is dynamically adjusted by the high-frequency energy ratio and the inverse of the initial compression coefficient. The high-frequency band feature weight can avoid excessive discarding of power electronic harmonics through the dynamic adjustment mechanism, thereby generating a frequency band feature weight vector containing three types of frequency band feature weights to constrain the feature retention priority during the compression process. At the same time, this embodiment constructs a feedforward neural network including an input layer, three hidden layers and an output layer, wherein the input layer dimension is consistent with the original measurement signal length, that is, the number of input layer nodes is equal to the original measurement signal length; the hidden layer dimension is reduced in logarithmic proportion according to the initial compression coefficient, and the specific calculation formula is:

[0082]

[0083] Where, is the number of nodes in the nth hidden layer; is the number of input layer nodes; C is the initial compression coefficient; n is the hidden layer index. In this embodiment, n=1, 2, 3.

[0084] In this embodiment, the number of nodes in each hidden layer decreases exponentially according to the initial compression coefficient, which can avoid information loss caused by dimensionality mutation and ensure a smooth compression process; the number of nodes in the output layer is determined by the product of the initial compression coefficient and the original signal length. In this embodiment, the frequency band feature weight vector is input as the bias condition of the hidden layer to ensure that the network preferentially retains the time-frequency characteristics of the high-weight frequency band during the compression process.

[0085] For the training process of the feedforward neural network, this embodiment uses the original measurement signal as the input target and the output of the feedforward neural network as the reconstructed signal, and trains the network parameters by minimizing the joint loss function in the time-frequency domain. This embodiment can prevent the loss of single domain features by optimizing the joint loss function in the time-frequency domain, wherein the joint loss function includes the time domain loss function, the frequency domain loss function and the sparsity constraint. The time domain loss function calculates the mean square error between the reconstructed signal and the original measurement signal within the preset attenuation time window, focusing on retaining the attenuation slope characteristics of the oscillation amplitude; the frequency domain loss function calculates the cosine similarity of the energy distribution of the reconstructed signal and the original signal in the low-frequency band and the mid-frequency band to ensure the integrity of the dominant oscillation mode; the sparse constraint The beam applies L1 norm regularization to the high-frequency band reconstructed signal to balance the contradiction between retaining high-frequency components and data volume, and suppress the excessive retention of non-critical harmonic components. This embodiment can use the Adam optimizer for training until the total loss function converges to a preset threshold. This embodiment inputs the original measurement signal into the trained feedforward neural network, realizes signal dimension compression through the nonlinear transformation of the hidden layer, and dynamically trims the output layer dimension according to the initial compression coefficient to generate a measurement compression signal that simultaneously retains the time domain attenuation characteristics and frequency band energy distribution. Its data volume is reduced compared to the original signal and the key oscillation mode is retained. This embodiment realizes the adaptive retention of high-frequency transient characteristics through the coupled calculation of frequency band energy and compression coefficient.

[0086] S3. Embed time series position coding into the measured compressed signal to generate time series input data with a time evolution trend.

[0087] S4. Use a multi-head attention mechanism to analyze the changes in oscillation amplitudes in different time windows in the time series input data to obtain an attention weighted signal with a disturbance type label.

[0088] In some embodiments, the step of analyzing the oscillation amplitude changes in different time windows in the time series input data using a multi-head attention mechanism to obtain an attention-weighted signal with a disturbance type label includes:

[0089] According to the broadband oscillation characteristics of the power system, the dominant mode period of electromechanical oscillation is taken as the basic oscillation period;

[0090] Using a multi-head attention mechanism to process the time series input data in parallel, and using the basic oscillation period to divide the time series input data into multiple time window segments;

[0091] Calculate the amplitude change rate, energy accumulation and phase difference absolute value in each time window segment based on the time series input data, and generate an attention key value matrix for each time window segment;

[0092] Based on the pre-acquired load disturbance feature template, the attention key value matrix of each time window segment is matched with the disturbance feature template to obtain the attention weight vector of each time window segment;

[0093] The time series input data is weightedly fused according to the attention weight vector, and an attention weighted signal with a disturbance type label is output.

[0094] Specifically, in this embodiment, the sine-cosine position coding of the timestamp is superimposed on each sampling point of the measured compressed signal to generate time series input data containing absolute time information and relative timing relationship; wherein, the wavelength sequence of the position coding is scaled logarithmically by the length of the time window to ensure that oscillation modes of different time scales can be effectively characterized. For example, the wavelengths of the sine and cosine functions are scaled logarithmically by the length of the time window, so that electromechanical oscillations of seconds and power electronic disturbances of milliseconds can be effectively characterized. Then, in this embodiment, according to the typical time scale of broadband oscillations of the power system, the basic oscillation period is set to the period of the dominant mode of electromechanical oscillation. For example, in this embodiment, the basic oscillation period can be set to 5 seconds, and a multi-head attention mechanism is used to process the time series input data in parallel, dividing the time series input data into multiple time window segments, and the length of each time window segment is set to an integer multiple of the basic oscillation period, covering up to At least 3 complete oscillation cycles are used to capture the transient process. Then, this embodiment calculates the amplitude change rate, energy accumulation and phase difference absolute value of the time series input data in each time window segment, wherein the amplitude change rate is obtained by calculating the instantaneous slope through the amplitude difference of adjacent sampling points, which characterizes the oscillation amplitude attenuation characteristic; the energy accumulation is obtained by statistically calculating the integral value of the square of the time series input data in the time window segment, which reflects the spatiotemporal distribution of the oscillation energy; the phase difference absolute value is obtained by calculating the phase difference absolute value of the first and last sampling points of the time window, specifically, the Hilbert transform and other methods are used to obtain the phase information of the time series input data at both ends of the time window segment, and the phase difference absolute value of the first and last sampling points of the time window is calculated based on the phase information, and its quantization reflects the degree of phase offset, thereby generating an attention key value matrix containing multi-time window features, which contains three feature dimensions: amplitude change rate, energy accumulation and phase difference absolute value.

[0095] The load disturbance feature template pre-acquired in this embodiment includes an amplitude change rate feature pattern, an energy accumulation feature pattern, and a phase difference absolute value feature pattern when the load suddenly changes. For example, when the load suddenly changes, the amplitude change rate will show a feature of a sharp rise or fall, the energy accumulation will increase rapidly in a short period of time, and the absolute value of the phase difference will suddenly change. Then, a similarity matching algorithm such as the cosine similarity algorithm is used to match the attention key value matrix of each time window segment with the load disturbance feature template. Specifically, this embodiment calculates the cosine similarity of the amplitude change rate feature vector, the energy accumulation feature vector, and the phase difference absolute value feature vector of the time window segment with the corresponding feature vector in the load disturbance feature template, and performs a weighted summation on the cosine similarities of the three feature dimensions to obtain a comprehensive matching score for each time window segment with the load disturbance feature template. The weight coefficient can be set according to the actual application scenario, thereby obtaining an attention weight vector for each time window segment, which represents the degree of matching between the time window segment and the load disturbance feature template.

[0096] Next, this embodiment determines the disturbance type based on the attention weight vector. Specifically, when the comprehensive matching score in the attention weight vector exceeds a preset matching score threshold, it is determined that the time series input data corresponding to the time window segment has a load disturbance. It should be noted that those skilled in the art can set other types of disturbance feature templates at the same time according to the specific implementation. For example, the generator speed regulation system imbalance feature template can be matched and the comprehensive matching score can be calculated according to a similar method. It can be determined as a generator imbalance type label based on the comprehensive matching score. The disturbance type is then determined based on the size relationship of the comprehensive matching scores of different types of disturbances. Finally, the attention weight vector of each time window segment is used as a weight coefficient to perform a weighted average on the time series input data within the time window segment to obtain weighted time series input data. The weighted time series input data is then assigned a corresponding disturbance type label according to the disturbance type judgment rule. For example, if the load disturbance feature template has the highest comprehensive matching score, it is assigned a load disturbance type label, and the attention weighted signal with the disturbance type label is finally obtained.

[0097] S5. Generate a reconstructed signal through linear layer transformation according to the attention weighted signal, and determine the preliminary disturbance source area range based on the reconstructed signal and the disturbance propagation path.

[0098] In some embodiments, the step of determining the scope of the preliminary disturbance source region based on the reconstructed signal and the disturbance propagation path includes:

[0099] The electrical nodes of the power grid topology are analyzed through a graph attention network to extract node feature vectors that combine electrical coupling strength and power transmission direction.

[0100] Calculate the propagation path weight of each electrical node in the disturbance propagation path based on the node eigenvector and the oscillation propagation delay characteristics of the reconstructed signal;

[0101] Screening out a set of key nodes whose propagation path weight is greater than a preset area division threshold from the electrical nodes in the power grid topology structure, and removing isolated nodes and merging adjacent node groups based on the spatial distribution density of the key node set to form a candidate key node cluster;

[0102] Integrating the edge connection weights within the candidate key node cluster to calculate the disturbance propagation path density per unit area, and dividing the candidate key node cluster into different disturbance propagation levels according to the disturbance propagation path density; the different disturbance propagation levels include core propagation area, secondary propagation area and edge area;

[0103] According to the principle of shortest electrical distance, the connected domain analysis algorithm is used to merge the overlapping areas of different disturbance propagation levels to form a continuous preliminary disturbance source area range.

[0104] Specifically, this embodiment inputs the attention weighted signal into the linear layer network, maps and transforms the signal features by linear combination, and obtains a reconstructed signal after the linear layer transformation. The reconstructed signal retains the original signal features while highlighting the feature information related to the disturbance source for subsequent disturbance source location analysis, wherein the linear layer can adopt a fully connected layer network. At the same time, this embodiment uses the busbar or equipment access node in the power grid topology structure as the vertex, the transmission line as the edge, and the line impedance or power transmission capacity as the weight to construct a power grid topology graph structure model based on the graph attention network, and extracts the node feature vector of the power grid topology graph structure model through the multi-head attention mechanism. The node feature vector integrates the electrical node coupling strength and the power transmission direction, wherein the electrical node coupling strength is the product of the phase difference change rate of adjacent electrical nodes and the amplitude attenuation slope. Then, this embodiment extracts the node feature vector from the reconstructed signal. The oscillation propagation delay characteristics are extracted and the time characteristics of the disturbance propagation are quantified. The oscillation propagation delay characteristics are specifically obtained by comparing the time difference of the oscillation phase mutations of adjacent electrical nodes in the reconstructed signal, and the propagation path weight of each electrical node in the disturbance propagation path is calculated according to the node feature vector and the oscillation propagation delay characteristics extracted from the reconstructed signal; wherein the propagation path weight is equal to the weighted sum of the electrical node coupling strength and the inverse of the oscillation propagation delay characteristics. Then, this embodiment screens the electrical node set whose weight coefficient is greater than the preset propagation threshold based on the preset area division threshold, and according to the spatial distribution density of the screened electrical node set, eliminates isolated nodes and merges adjacent node groups to form a candidate key node cluster, wherein the isolated node is the adjacent node whose distance exceeds the preset electrical distance threshold; it should be noted that the electrical node in this embodiment is the bus or equipment access node in the power grid topology.

[0105] This embodiment performs a cumulative integral operation on the edge connection weights within the candidate key node cluster to calculate the disturbance propagation path density within a unit area. The candidate key node cluster is divided into three levels based on the disturbance propagation path density: a core propagation area, a secondary propagation area, and an edge area. For example, a disturbance propagation path density in the core propagation area is ≥80% percentile, indicating a strong disturbance propagation path; a disturbance propagation path density in the secondary propagation area is between 50% and 80% percentile, indicating a medium propagation path; and a disturbance propagation path density in the edge area is less than 50% percentile, indicating a weak propagation path. This embodiment maps the electrical nodes in the core propagation area and the secondary propagation area to a grid geographic wiring diagram of the grid topology structure, expands the area boundary according to the principle of shortest electrical distance, and ensures that the electrical distance between any two nodes in the area does not exceed a preset electrical distance threshold. Based on the expanded area boundary, this embodiment uses a connected domain analysis algorithm to merge the overlapping areas of the core propagation area and the secondary propagation area, and ultimately determines the scope of the preliminary disturbance source area containing the potential disturbance source.

[0106] S6. Analyze the reconstruction error between the reconstructed signal and the original measurement signal within the preliminary disturbance source region to obtain a hierarchical positioning result of the broadband oscillation disturbance source.

[0107] In some embodiments, the step of analyzing the reconstruction error between the reconstructed signal and the original measurement signal within the preliminary disturbance source region to obtain a hierarchical positioning result of the broadband oscillation disturbance source includes:

[0108] Calculating a mean square error between the reconstructed signal and the original measurement signal in a preset frequency band within the preliminary disturbance source region, and screening out high-error electrical nodes from electrical nodes in the power grid topology structure based on the mean square error;

[0109] Based on the principle of shortest electrical distance, the boundary of the disturbance propagation level area is expanded to the minimum connected area covering all high-error electrical nodes to obtain the correction range of the disturbance source area;

[0110] Extracting attenuation characteristic parameters of high-error electrical nodes within the correction range of the disturbance source region, and matching the attenuation characteristic parameters with a pre-acquired device-level attenuation feature library to obtain an attenuation characteristic matching degree; the attenuation characteristic parameters include a signal attenuation slope and an oscillation amplitude attenuation time constant;

[0111] Screening out high-error electrical nodes whose attenuation characteristic matching degree is higher than a preset attenuation threshold as core candidate nodes, and performing spatial density clustering on the core candidate nodes to generate a core disturbance area;

[0112] Analyzing the spatial consistency of the phase propagation direction and the amplitude attenuation gradient direction of the reconstructed signal within the core perturbation region to generate a consistency score;

[0113] The core disturbance area is divided according to the consistency score to obtain hierarchical positioning areas, and the hierarchical positioning areas are associated with a device coordinate library of a power grid topology structure to obtain hierarchical positioning results of broadband oscillation disturbance sources.

[0114] Specifically, this embodiment calculates the mean square error between the reconstructed signal and the original measured signal in a preset frequency band within the scope of the preliminary disturbance source area. If the mean square error exceeds the preset dynamic adjustment threshold, the corresponding electrical node in the power grid topology is marked as a high-error electrical node, and the area boundary is expanded based on the principle of shortest electrical distance, and the disturbance propagation level area boundary is expanded to the minimum connected area (non-geographic rectangle) covering all high-error electrical nodes to generate a disturbance source area correction range. Then, the signal attenuation slope and the oscillation amplitude attenuation time constant of the high-error electrical node within the disturbance source area correction range are extracted to obtain the attenuation characteristic parameters of the high-error electrical node. This embodiment compares the attenuation characteristic parameters of the high-error electrical node with the device-level attenuation feature library. Matching is performed to obtain the attenuation characteristic matching degree. The device-level attenuation characteristic library may include device characteristics such as the generator rotor inertia time constant (2-10 seconds) and the converter switching frequency attenuation parameter (100-2000Hz). Then, according to the attenuation characteristic matching degree, high-error electrical nodes with attenuation characteristic matching degrees higher than the preset attenuation threshold are selected as core candidate nodes, and the core candidate nodes are spatially density clustered to eliminate isolated electrical nodes with spatial density lower than the preset density threshold to generate a core disturbance area. In this embodiment, the spatial consistency of the phase propagation direction and the amplitude attenuation gradient direction of the reconstructed signal is analyzed in the core disturbance area. Specifically, if the angle between the phase propagation direction and the attenuation gradient direction is ≤15°, it is determined to be a consistent propagation path. path; if the angle between the phase propagation direction and the attenuation gradient direction is greater than 15°, the propagation path weight coefficient is recalculated, and when it is determined to be a consistent propagation path, the angle ratio between the phase propagation direction and the attenuation gradient direction and 90° is calculated, and the angle ratio is normalized to generate a consistency score. In this embodiment, the core disturbance area is divided into three levels of positioning results according to the consistency score. For example, the consistency score of the first-level positioning area is ≥0.9, which can be directly mapped to the device-level coordinates. For example, the accuracy of the device-level coordinates is ≤500 meters, which corresponds to the generator export bus or the converter grid connection point; the consistency score range of the second-level positioning area is (0.7, 0.9), which can be marked as the substation interval level, the substation interval level The accuracy is ≤2 kilometers, which corresponds to the 220kV interval or the low-voltage side of the main transformer; the consistency score of the third-level positioning area is ≤0.7, which can be marked as the plant level. The accuracy of the plant level is ≤5 kilometers, which corresponds to the coverage of the 500kV substation. Finally, this embodiment associates the hierarchical positioning area with the equipment coordinate library of the power grid geographic information system, and outputs the hierarchical positioning result of the broadband oscillation disturbance source containing the following three levels of information. The hierarchical positioning result of the broadband oscillation disturbance source may include regional level, equipment type and coordinate range. For example, the regional level may include the geographical boundaries of the first / second / third level positioning area, the equipment type may include the generator speed control system, converter group or other power electronic equipment, and the coordinate range may be the longitude-latitude boundary box.

[0115] An embodiment of the present invention provides a method for hierarchical positioning of broadband oscillation disturbance sources in an electric power system. The method obtains an original measurement signal from an electric power system monitoring device, calculates a compression ratio of the measurement signal based on a sampling resolution of the original measurement signal and a real-time transmission interval, and obtains an initial compression coefficient; performs frequency domain decomposition on the original measurement signal to obtain spectrum feature data including oscillation frequency distribution characteristics, and compresses the original measurement signal based on the spectrum feature data and the initial compression coefficient to obtain a measurement compression signal; embeds a time series position code in the measurement compression signal to generate time series input data with a time evolution trend; utilizes a multi-head attention mechanism to analyze the oscillation amplitude changes in different time windows in the time series input data to obtain an attention weighted signal with a disturbance type label; generates a reconstructed signal based on the attention weighted signal through a linear layer transformation, and determines a preliminary disturbance source area range based on the reconstructed signal and the disturbance propagation path; analyzes a reconstruction error between the reconstructed signal and the original measurement signal within the preliminary disturbance source area, and obtains a hierarchical positioning result of the broadband oscillation disturbance source. Compared with existing technologies, this method achieves rapid and accurate hierarchical positioning of broadband oscillation disturbance sources in complex power grid environments through technologies such as adaptive compression to retain broadband oscillation characteristics and topological propagation path analysis, significantly improving positioning accuracy and real-time performance, and providing technical support for power grid fault diagnosis and stable operation.

[0116] It should be noted that the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiment of this application.

[0117] In one embodiment, Figure 2 As shown, an embodiment of the present invention provides a hierarchical positioning system for broadband oscillation disturbance sources in a power system, the system comprising:

[0118] The compression analysis module 101 is used to obtain the original measurement signal of the power system monitoring equipment, calculate the compression ratio of the measurement signal according to the sampling resolution and real-time transmission interval of the original measurement signal, and obtain the initial compression coefficient;

[0119] a signal compression module 102 configured to perform frequency domain decomposition on the original measurement signal to obtain frequency spectrum characteristic data including oscillation frequency distribution characteristics, and to compress the original measurement signal based on the frequency spectrum characteristic data and the initial compression coefficient to obtain a measurement compression signal;

[0120] A time series embedding module 103 is configured to embed a time series position code into the measured compressed signal to generate time series input data with a time evolution trend;

[0121] A signal analysis module 104 is configured to analyze the oscillation amplitude changes in different time windows of the time series input data using a multi-head attention mechanism to obtain an attention-weighted signal with a disturbance type label;

[0122] The disturbance analysis module 105 is configured to generate a reconstructed signal through a linear layer transformation according to the attention weighted signal, and determine the scope of a preliminary disturbance source region based on the reconstructed signal and the disturbance propagation path;

[0123] The hierarchical positioning module 106 is configured to analyze the reconstruction error between the reconstructed signal and the original measurement signal within the preliminary disturbance source region to obtain a hierarchical positioning result of the broadband oscillation disturbance source.

[0124] For the specific definition of a hierarchical positioning system for broadband oscillation disturbance sources in a power system, please refer to the above-mentioned definition of a hierarchical positioning method for broadband oscillation disturbance sources in a power system, which will not be repeated here. A person of ordinary skill in the art will appreciate that the various modules and steps described in conjunction with the embodiments disclosed in this application can be implemented in hardware, software, or a combination of both. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0125] An embodiment of the present invention provides a hierarchical positioning system for broadband oscillation disturbance sources in a power system. The system obtains original measurement signals from power system monitoring equipment through a compression analysis module, calculates the compression ratio of the measurement signals based on the sampling resolution and real-time transmission interval of the original measurement signals, and obtains an initial compression coefficient. The signal compression module performs frequency domain decomposition on the original measurement signals to obtain spectrum feature data containing oscillation frequency distribution characteristics, and compresses the original measurement signals based on the spectrum feature data and the initial compression coefficient to obtain a measurement compression signal. The time series embedding module embeds time series position codes in the measurement compression signal to generate time series input data with a time evolution trend. The signal analysis module uses a multi-head attention mechanism to analyze the oscillation amplitude changes in different time windows in the time series input data to obtain an attention-weighted signal with a disturbance type label. The disturbance analysis module generates a reconstructed signal through a linear layer transformation based on the attention-weighted signal, and determines the scope of a preliminary disturbance source area based on the reconstructed signal and the disturbance propagation path. The hierarchical positioning module analyzes the reconstruction error between the reconstructed signal and the original measurement signal within the preliminary disturbance source area to obtain a hierarchical positioning result for the broadband oscillation disturbance source. Compared with existing technologies, this system achieves rapid and accurate hierarchical positioning of broadband oscillation disturbance sources in complex power grid environments through technologies such as adaptive compression to retain broadband oscillation characteristics and topological propagation path analysis, significantly improving positioning accuracy and real-time performance, and providing technical support for power grid fault diagnosis and stable operation.

[0126] The above-described embodiments merely represent several preferred implementations of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent. It should be noted that a person skilled in the art could make several improvements and substitutions without departing from the technical principles of the present invention, and these improvements and substitutions should also be considered within the scope of protection of the present application. Therefore, the scope of protection of the present patent application shall be based on the scope of protection of the claims.

Claims

1. A hierarchical location method for broadband oscillation disturbance sources in a power system, characterized in that: The following steps are involved: Obtaining the original measurement signal of the power system monitoring equipment, calculating the compression ratio of the measurement signal according to the sampling resolution of the original measurement signal and the real-time transmission interval, and obtaining the initial compression coefficient; Obtaining a sampling resolution and a sampling point density of the original measurement signal, and calculating a ratio of the sampling resolution to the sampling point density to obtain a density ratio parameter; Obtaining a data upload frequency and a real-time transmission interval of the original measurement signal, and calculating a ratio of the data upload frequency to the real-time transmission interval to obtain a frequency control factor; According to the density ratio parameter and the frequency control factor, the density ratio parameter is multiplied by the frequency control factor to calculate an initial compression ratio value, and the initial compression ratio value is compared with a preset compression upper limit threshold; If the initial compression ratio value exceeds a preset compression upper threshold, the sampling resolution is reversely adjusted using the initial compression ratio value, and the sampling resolution is multiplied by the quotient of the compression upper threshold and the initial compression ratio value to obtain a sampling resolution update value; Recalculating the density ratio parameter based on the sampling resolution update value to obtain an updated density ratio parameter value, and obtaining an initial compression coefficient based on the updated density ratio parameter value and a real-time transmission interval, wherein the value of the initial compression coefficient is equal to the updated density ratio parameter value divided by the real-time transmission interval; Performing frequency domain decomposition on the original measurement signal to obtain frequency spectrum characteristic data including oscillation frequency distribution characteristics, and compressing the original measurement signal based on the frequency spectrum characteristic data and the initial compression coefficient to obtain a measurement compression signal; Obtaining a frequency band feature weight vector for each frequency band according to the energy proportion of each frequency band in the frequency spectrum feature data and the initial compression coefficient; Extract the energy proportions of the low-frequency, mid-frequency, and high-frequency bands from the spectral feature data. The low-frequency feature weight is equal to the product of the low-frequency energy proportion and the initial compression coefficient. The mid-frequency feature weight is equal to the product of the mid-frequency energy proportion and the square root of the initial compression coefficient. The high-frequency feature weight is dynamically adjusted by the high-frequency energy proportion and the inverse of the initial compression coefficient. A feedforward neural network with three hidden layers is constructed. Input the frequency band feature weight vector as a bias condition of the hidden layer, input the original measurement signal into the trained feedforward neural network, perform signal dimension compression through the nonlinear transformation of the hidden layer, and generate a measurement compression signal; Embedding a time series position code in the measurement compression signal to generate time series input data with a time evolution trend; Utilizing a multi-head attention mechanism to analyze the oscillation amplitude changes in different time windows in the time series input data, and obtaining an attention-weighted signal with a disturbance type label; Generate a reconstructed signal through a linear layer transformation according to the attention weighted signal, and determine the scope of the preliminary disturbance source area based on the reconstructed signal and the disturbance propagation path; Analyzing the reconstruction error between the reconstructed signal and the original measured signal within the preliminary disturbance source region to obtain a hierarchical positioning result of the broadband oscillation disturbance source; and taking the dominant mode period of the electromechanical oscillation as the basic oscillation period according to the broadband oscillation characteristics of the power system; Using a multi-head attention mechanism to process the time series input data in parallel, and using the basic oscillation period to divide the time series input data into multiple time window segments; Calculate the amplitude change rate, energy accumulation and phase difference absolute value in each time window segment based on the time series input data, and generate an attention key value matrix for each time window segment; Based on the pre-acquired load disturbance feature template, the attention key value matrix of each time window segment is matched with the disturbance feature template. The cosine similarity between the amplitude change rate feature vector, energy accumulation feature vector, and phase difference absolute value feature vector of the time window segment and the corresponding feature vector in the load disturbance feature template is calculated respectively. The cosine similarity of the three feature dimensions is weighted and summed to obtain the comprehensive matching score of each time window segment with the load disturbance feature template, and the attention weight vector of each time window segment is obtained. Performing weighted fusion on the time series input data according to the attention weight vector, and outputting an attention weighted signal with a disturbance type label; The electrical nodes of the power grid topology are analyzed through a graph attention network to extract node feature vectors that combine electrical coupling strength and power transmission direction. The oscillation propagation delay characteristics are specifically obtained by comparing the timing differences of the oscillation phase mutations of adjacent electrical nodes in the reconstructed signal; Based on the node eigenvector and the oscillation propagation delay characteristics of the reconstructed signal, the propagation path weight of each electrical node in the disturbance propagation path is calculated. The propagation path weight is equal to the weighted sum of the electrical node coupling strength and the inverse of the oscillation propagation delay characteristics. Screening out a set of key nodes whose propagation path weight is greater than a preset area division threshold from the electrical nodes in the power grid topology structure, and removing isolated nodes and merging adjacent node groups based on the spatial distribution density of the key node set to form a candidate key node cluster; Integrating the edge connection weights within the candidate key node cluster to calculate the disturbance propagation path density per unit area, and dividing the candidate key node cluster into different disturbance propagation levels according to the disturbance propagation path density; According to the principle of shortest electrical distance, the connected domain analysis algorithm is used to merge the overlapping areas of different disturbance propagation levels to form a continuous preliminary disturbance source area range; Calculating a mean square error between the reconstructed signal and the original measurement signal in a preset frequency band within the preliminary disturbance source region, and screening out high-error electrical nodes from electrical nodes in the power grid topology structure based on the mean square error; Based on the principle of shortest electrical distance, the boundary of the disturbance propagation level area is expanded to the minimum connected area covering all high-error electrical nodes to obtain the correction range of the disturbance source area; Extracting attenuation characteristic parameters of high-error electrical nodes within the correction range of the disturbance source region, and matching the attenuation characteristic parameters with a pre-acquired device-level attenuation feature library to obtain an attenuation characteristic matching degree; Screening out high-error electrical nodes whose attenuation characteristic matching degree is higher than a preset attenuation threshold as core candidate nodes, and performing spatial density clustering on the core candidate nodes to generate a core disturbance area; Analyzing the spatial consistency of the phase propagation direction and the amplitude attenuation gradient direction of the reconstructed signal within the core perturbation region to generate a consistency score; The core disturbance area is divided according to the consistency score to obtain hierarchical positioning areas, and the hierarchical positioning areas are associated with a device coordinate library of a power grid topology structure to obtain hierarchical positioning results of broadband oscillation disturbance sources.

2. A hierarchical location method for broadband oscillation disturbance sources in a power system according to claim 1, characterized in that: The step of performing frequency domain decomposition on the original measurement signal to obtain frequency spectrum feature data including oscillation frequency distribution characteristics includes: The original measurement signal is converted into a measurement frequency domain signal by fast Fourier transform, and the measurement frequency domain signal is divided into frequency bands by using a bandpass filter based on the wide-band oscillation frequency distribution range to obtain a normalized frequency band spectrum; A multi-head attention mechanism is used to extract the spatiotemporal distribution patterns of oscillation energy in different frequency bands from the normalized frequency band spectrum to obtain corresponding frequency band feature vectors; Performing dimensionality reduction processing on the frequency band feature vector to obtain a frequency band reduced dimensionality feature vector, and inputting the frequency band reduced dimensionality feature vector into a pre-trained autoencoder to reconstruct the frequency band spectrum to obtain a reconstructed feature vector; The reconstructed feature vectors of different frequency bands are arranged into row vectors in ascending order of frequency to construct a two-dimensional spectrum feature matrix, and the two-dimensional spectrum feature matrix is ​​normalized to obtain spectrum feature data.

3. The method for hierarchically locating a broadband oscillation disturbance source in a power system according to claim 1, wherein: The number of nodes in each hidden layer of the feedforward neural network decreases exponentially according to the initial compression coefficient.

4. The method for hierarchically locating a broadband oscillation disturbance source in a power system according to claim 1, wherein: The length of each time window segment is set to an integer multiple of the basic oscillation period.

5. A hierarchical positioning system for broadband oscillation disturbance sources in power systems, characterized in that: The system comprises: The compression analysis module is used to obtain the original measurement signal of the power system monitoring equipment, calculate the compression ratio of the measurement signal according to the sampling resolution of the original measurement signal and the real-time transmission interval, and obtain the initial compression coefficient; Obtaining a sampling resolution and a sampling point density of the original measurement signal, and calculating a ratio of the sampling resolution to the sampling point density to obtain a density ratio parameter; Obtaining a data upload frequency and a real-time transmission interval of the original measurement signal, and calculating a ratio of the data upload frequency to the real-time transmission interval to obtain a frequency control factor; According to the density ratio parameter and the frequency control factor, the density ratio parameter is multiplied by the frequency control factor to calculate an initial compression ratio value, and the initial compression ratio value is compared with a preset compression upper limit threshold; If the initial compression ratio value exceeds a preset compression upper threshold, the sampling resolution is reversely adjusted using the initial compression ratio value, and the sampling resolution is multiplied by the quotient of the compression upper threshold and the initial compression ratio value to obtain a sampling resolution update value; Recalculating the density ratio parameter based on the sampling resolution update value to obtain an updated density ratio parameter value, and obtaining an initial compression coefficient based on the updated density ratio parameter value and a real-time transmission interval, wherein the value of the initial compression coefficient is equal to the updated density ratio parameter value divided by the real-time transmission interval; a signal compression module, configured to perform frequency domain decomposition on the original measurement signal to obtain frequency spectrum characteristic data including oscillation frequency distribution characteristics, and compress the original measurement signal based on the frequency spectrum characteristic data and the initial compression coefficient to obtain a measurement compression signal; Obtaining a frequency band feature weight vector for each frequency band according to the energy proportion of each frequency band in the frequency spectrum feature data and the initial compression coefficient; Extract the energy proportions of the low-frequency, mid-frequency, and high-frequency bands from the spectral feature data. The low-frequency feature weight is equal to the product of the low-frequency energy proportion and the initial compression coefficient. The mid-frequency feature weight is equal to the product of the mid-frequency energy proportion and the square root of the initial compression coefficient. The high-frequency feature weight is dynamically adjusted by the high-frequency energy proportion and the inverse of the initial compression coefficient. A feedforward neural network with three hidden layers is constructed. Input the frequency band feature weight vector as a bias condition of the hidden layer, input the original measurement signal into the trained feedforward neural network, perform signal dimension compression through the nonlinear transformation of the hidden layer, and generate a measurement compression signal; A time series embedding module, configured to embed a time series position code into the measurement compression signal to generate time series input data with a time evolution trend; A signal analysis module is used to analyze the oscillation amplitude changes in different time windows in the time series input data using a multi-head attention mechanism to obtain an attention-weighted signal with a disturbance type label; A disturbance analysis module is used to generate a reconstructed signal through a linear layer transformation according to the attention weighted signal, and determine the scope of the preliminary disturbance source area based on the reconstructed signal and the disturbance propagation path; a hierarchical positioning module, configured to analyze a reconstruction error between the reconstructed signal and the original measurement signal within the preliminary disturbance source region to obtain a hierarchical positioning result of the broadband oscillation disturbance source; According to the broadband oscillation characteristics of the power system, the dominant mode period of electromechanical oscillation is taken as the basic oscillation period; Using a multi-head attention mechanism to process the time series input data in parallel, and using the basic oscillation period to divide the time series input data into multiple time window segments; Calculate the amplitude change rate, energy accumulation and phase difference absolute value in each time window segment based on the time series input data, and generate an attention key value matrix for each time window segment; Based on the pre-acquired load disturbance feature template, the attention key value matrix of each time window segment is matched with the disturbance feature template. The cosine similarity between the amplitude change rate feature vector, energy accumulation feature vector, and phase difference absolute value feature vector of the time window segment and the corresponding feature vector in the load disturbance feature template is calculated respectively. The cosine similarity of the three feature dimensions is weighted and summed to obtain the comprehensive matching score of each time window segment with the load disturbance feature template, and the attention weight vector of each time window segment is obtained. Performing weighted fusion on the time series input data according to the attention weight vector, and outputting an attention weighted signal with a disturbance type label; The electrical nodes of the power grid topology are analyzed through a graph attention network to extract node feature vectors that combine electrical coupling strength and power transmission direction. The oscillation propagation delay characteristics are specifically obtained by comparing the timing differences of the oscillation phase mutations of adjacent electrical nodes in the reconstructed signal; Based on the node eigenvector and the oscillation propagation delay characteristics of the reconstructed signal, the propagation path weight of each electrical node in the disturbance propagation path is calculated. The propagation path weight is equal to the weighted sum of the electrical node coupling strength and the inverse of the oscillation propagation delay characteristics. Screening out a set of key nodes whose propagation path weight is greater than a preset area division threshold from the electrical nodes in the power grid topology structure, and removing isolated nodes and merging adjacent node groups based on the spatial distribution density of the key node set to form a candidate key node cluster; Integrating the edge connection weights within the candidate key node cluster to calculate the disturbance propagation path density per unit area, and dividing the candidate key node cluster into different disturbance propagation levels according to the disturbance propagation path density; According to the principle of shortest electrical distance, the connected domain analysis algorithm is used to merge the overlapping areas of different disturbance propagation levels to form a continuous preliminary disturbance source area range; Calculating a mean square error between the reconstructed signal and the original measurement signal in a preset frequency band within the preliminary disturbance source region, and screening out high-error electrical nodes from electrical nodes in the power grid topology structure based on the mean square error; Based on the principle of shortest electrical distance, the boundary of the disturbance propagation level area is expanded to the minimum connected area covering all high-error electrical nodes to obtain the correction range of the disturbance source area; Extracting attenuation characteristic parameters of high-error electrical nodes within the correction range of the disturbance source region, and matching the attenuation characteristic parameters with a pre-acquired device-level attenuation feature library to obtain an attenuation characteristic matching degree; Screening out high-error electrical nodes whose attenuation characteristic matching degree is higher than a preset attenuation threshold as core candidate nodes, and performing spatial density clustering on the core candidate nodes to generate a core disturbance area; Analyzing the spatial consistency of the phase propagation direction and the amplitude attenuation gradient direction of the reconstructed signal within the core perturbation region to generate a consistency score; The core disturbance area is divided according to the consistency score to obtain hierarchical positioning areas, and the hierarchical positioning areas are associated with a device coordinate library of a power grid topology structure to obtain hierarchical positioning results of broadband oscillation disturbance sources.

Images