A broadband oscillation detection method, system and computer readable storage medium

Abstract

The present application belongs to the technical field of wideband oscillation detection, and particularly relates to a wideband oscillation detection method, system and computer readable storage medium. The method first acquires a monitoring signal, performs time-frequency conversion on the monitoring signal to obtain an original frequency spectrum signal; then performs frequency domain convolution on the original frequency spectrum signal and the frequency spectrum of a window function with a frequency spectrum leakage suppression function, obtains a frequency spectrum after noise reduction according to the frequency domain convolution result; and finally performs wideband oscillation detection using the frequency spectrum after noise reduction. The frequency domain convolution technology is used to suppress the frequency spectrum leakage caused by the truncation effect in the FFT operation, improve the signal-to-noise ratio of the high-frequency weak signal, improve the detection accuracy of the wideband oscillation, and reduce the possibility of oscillation misjudgment.

Description

Technical Field

[0001] This invention belongs to the field of broadband oscillation detection technology, specifically relating to a broadband oscillation detection method, system, and computer-readable storage medium. Background Technology

[0002] Currently, wind power and photovoltaic power account for a relatively high proportion of installed capacity in my country, and the penetration rate of power electronic converter equipment (such as IGBTs and SiC converters) is also high, leading to fundamental changes in the dynamic characteristics of the power system.

[0003] 1) Wideband oscillation frequency band extension: The traditional power frequency oscillation (0.1~50Hz) is extended to subsynchronous / supersynchronous oscillation (50~2000Hz);

[0004] 2) Decentralization of oscillation sources: New energy power plants, SVG, DC converter stations, etc. can all become oscillation sources. Traditional Fourier transform methods (such as Fast Fourier Transform, FFT) are difficult to achieve rapid localization of multi-source oscillations.

[0005] Chinese invention patent application CN120870667A, published on October 31, 2025, discloses a broadband oscillation identification method and system based on the multi-band Nyquist criterion. It divides the frequency of broadband oscillations into low-frequency, mid-frequency, and high-frequency bands with overlapping frequency bands, and acquires the oscillation signals of monitoring nodes within the power grid in different frequency bands. It then calculates the equivalent impedance of the non-overlapping frequency bands of each band and the equivalent impedance of the overlapping frequency bands between two adjacent bands to construct the low-frequency band with overlapping frequency bands. The equivalent impedances of the mid-frequency and high-frequency bands are determined; the Nyquist criterion for each frequency band is established based on the transfer function of the dynamic characteristics of each frequency band under the open-loop state of the system; a generalized Nyquist criterion for overlapping frequency bands is constructed using a multi-band coupling matrix; a comprehensive decision function is constructed based on the identification results of system stability, comprehensive characteristic quantities of broadband oscillations, and the rate of change of impedance magnitude with frequency in the high-frequency band using the generalized Nyquist criterion; the broadband oscillation identification results are determined according to the value of the comprehensive decision function and the threshold, thereby improving the identification capability of broadband oscillations.

[0006] Traditional FFT uses a rectangular window, which leads to energy dispersion of high-frequency oscillation components and easily causes misjudgment of oscillation frequency. The switching noise of power electronic equipment overlaps with the frequency band of the oscillation signal, and conventional filtering methods cause phase distortion. There is an urgent need for a broadband oscillation detection device with high frequency resolution and strong real-time performance to solve the above-mentioned industry pain points. Summary of the Invention

[0007] The purpose of this invention is to provide a broadband oscillation detection method, system, and computer-readable storage medium to solve the problem in the prior art where the use of a rectangular window in traditional FFT leads to the dispersion of high-frequency oscillation component energy, which easily causes oscillation misjudgment.

[0008] To solve the above-mentioned technical problems, the present invention provides a technical solution for a broadband oscillation detection method, as follows:

[0009] The present invention provides a broadband oscillation detection method, the method comprising:

[0010] 1) Acquire the monitoring signal, perform time-frequency conversion on the monitoring signal, and obtain the original spectrum signal;

[0011] 2) Perform frequency domain convolution between the original spectral signal and the spectrum of a window function with spectral leakage suppression function, and obtain the denoised spectrum based on the frequency domain convolution result;

[0012] 3) Use the noise-reduced spectrum for broadband oscillation detection.

[0013] The beneficial effects of the above technical solution are as follows: Since the traditional FFT using a rectangular window is prone to causing energy dispersion of high-frequency oscillation components, the present invention uses the spectrum of a window function with spectrum leakage suppression function to perform frequency domain convolution on the original frequency domain signal obtained after time-frequency conversion. This is to suppress spectrum leakage caused by truncation effect in FFT operation through frequency domain convolution technology, improve the signal-to-noise ratio of weak high-frequency signals, improve the detection accuracy of broadband oscillations, and reduce the possibility of oscillation misjudgment.

[0014] Furthermore, the monitoring signal is time-frequency converted using a hybrid-based FFT. The hybrid-based FFT is used to: ① preferentially decompose the number of sampling points into 4 a ×8 b ×3 c ×5 d The combination of a,b,c,d≥0; then the remaining factors are decomposed into 2 factors; ② For the remaining factors that cannot be processed after step ①, if there is an optimized DFT kernel, the pre-generated optimized DFT kernel is called for processing; if there is no optimized DFT kernel, zeros are padded to the nearest neighbor number of points that can be completely decomposed by the strategy in step ①.

[0015] Furthermore, the radix-2, radix-4, and radix-8 twitch factors in the hybrid radix FFT are fixed as constants in the hardware logic of the butterfly operation unit; the radix-3 and radix-5 twitch factors in the hybrid radix FFT are pre-stored in memory and read through a lookup table; the inter-level twitch factors in the hybrid radix FFT calculation are all pre-calculated and stored in memory and read through a lookup table; the non-standard twitch factors in the hybrid radix FFT are calculated and generated in real time using the CORDIC algorithm.

[0016] Furthermore, the radix-8 in the hybrid radix FFT is decomposed into two levels of radix-4 operations. The two levels of radix-4 operations share a radix-4 butterfly unit, and the following timing scheduling strategy is used to generate the 8-point FFT output: the original 8-point input sequence is split into two 4-point subsequences, namely group 0 and group 1; during clock cycles T1~T4, the radix-4 butterfly unit is enabled, and the data of group 0 is sent to the radix-4 butterfly unit for radix-4 operation to obtain the 4-point intermediate result Y0; during clock cycles T5~T8, the data of group 1 is sent to the radix-4 butterfly unit for radix-4 operation to obtain the 4-point intermediate result Y1; the intermediate result Y0 is read and multiplied with the pre-stored rotation factor in an independent multiplier array to obtain Y0'; during clock cycles T9~T12, the result Y0' and Y1 are sent as input to the radix-4 butterfly unit to add Y0' and Y1, thereby generating the 8-point FFT output; T1, T2, ..., T11, T12 are 12 clock cycles that advance according to time.

[0017] Furthermore, when resources are limited, the hybrid radix FFT switches radix 3 and radix 5 to radix 2 mode and schedules them to idle periods to perform radix 3 and radix 5 operations using radix 2 operation units.

[0018] Furthermore, the window function with spectral leakage suppression function is the Hanning window.

[0019] Furthermore, the frequency domain coefficients of the Hanning window are selected based on the signal-to-noise ratio (SNR) of the monitored signal; when the SNR of the monitored signal is less than the set SNR threshold, the Hanning window is enabled in wide window mode; otherwise, the Hanning window is enabled in narrow window mode. The frequency domain broadening coefficient in wide window mode is less than that in narrow window mode.

[0020] Furthermore, the method for obtaining the denoised spectrum based on the frequency domain convolution result includes: inversely compensating the phase of the frequency domain convolution result, i.e., the compensated phase is the original phase of the frequency domain convolution result minus the phase shift, where the phase shift is the phase of the Hanning window's spectrum; normalizing the amplitude of the frequency domain convolution result using a window energy correction coefficient, where the normalized amplitude is the ratio of the original amplitude of the frequency domain convolution result to the window energy correction coefficient; the window energy correction coefficient K is:

[0021]

[0022] In the formula, w(n) is the time-domain expression of the Hanning window, and L is the total number of points of the Hanning window w(n).

[0023] Furthermore, the method of using the denoised spectrum for broadband oscillation detection includes using the denoised spectrum to identify whether oscillation exists, and the method of identifying whether oscillation exists includes: determining the amplitude of each frequency point based on the denoised spectrum; if the amplitude of a certain frequency point continuously exceeds the set amplitude threshold, it is considered that an oscillation event has occurred.

[0024] Furthermore, the method of broadband oscillation detection using the denoised spectrum includes oscillation pattern recognition using the denoised spectrum, where the oscillation patterns include at least subsynchronous, supersynchronous, and harmonic resonance. The method of identifying oscillation patterns includes: extracting the amplitude of the denoised spectrum and then squaring the extracted amplitude to obtain the energy spectrum of each frequency point; dividing the entire frequency band into multiple sub-bands and calculating the energy integral value of each sub-band within a set time window; forming an energy integral feature vector from the energy integral values ​​of each sub-band and inputting it into the oscillation pattern recognition model to obtain the oscillation pattern recognition result; the oscillation pattern recognition model is obtained by training a classification model using a dataset containing the energy integral feature vector and the corresponding oscillation patterns.

[0025] Furthermore, the method of broadband oscillation detection using the denoised spectrum includes locating the oscillation source using the denoised spectrum. The method for locating the oscillation source includes: obtaining the phase difference of the denoised spectrum corresponding to the monitoring signals of each pair of source points in the power grid through a PMU, and calculating the signal arrival time difference based on the phase difference; preliminarily determining the oscillation source based on the time difference data, combined with the electrical geometric model of the power grid and the time difference positioning principle, and forming an initial candidate source set by preliminarily determining the oscillation sources; traversing the initial candidate source set to determine whether any source point meets the node priority matching condition, where the node priority matching condition refers to the source point falling into a certain network node in the power grid topology GIS database or within an allowable error range centered on the network node. Within the specified range, source points that meet the node priority matching condition are identified as oscillation sources. Then, for source points that do not meet the node priority matching condition, it is determined whether any source point meets the line projection matching condition, where the line projection matching condition refers to the source point falling within the buffer zone of a transmission line. For source points that meet the line projection matching condition, the precise location of the oscillation source on that line is calculated through vertical projection. Then, for source points that do not meet the line projection matching condition, the minimum electrical distance from the source point to all topological elements in the power grid topology is calculated, and the source point is assigned to the topological element with the smallest distance. These topological elements include nodes and lines. Finally, the confidence level of all the above location results is calculated, and the oscillation source is determined based on the confidence level.

[0026] Furthermore, the confidence factors used to calculate the confidence level include consistency based on time difference measurement, the degree of fit between the source point and the topology element, and the dispersion among multiple candidate points. Consistency based on time difference measurement measures the degree of consistency when the time difference data calculated by different PMU nodes is used to infer the location of the oscillation source. The degree of fit between the source point and the topology element quantifies the physical proximity of the candidate source point and the grid topology element it matches.

[0027] Furthermore, the method also includes determining the risk level of the identified oscillation, and the determination method includes: extracting four key features of the identified oscillation: amplitude, frequency band, duration, and rate of change; performing a weighted summation of the four key features to obtain a comprehensive risk value; determining the risk level of the oscillation event based on the interval in which the comprehensive risk value is located, and triggering different alarm strategies for oscillation events of different risk levels.

[0028] Furthermore, the monitoring signal is the signal obtained by sequentially passing the original acquired signal through an RVS protection circuit, a differential input buffer, an instrumentation amplifier, a programmable gain amplifier, an anti-aliasing filter, an ADC driver, and a synchronous ADC. The RVS protection circuit is used for overvoltage clamping, the differential input buffer is used for impedance transformation, the instrumentation amplifier is used for differential amplification and common-mode rejection, the programmable gain amplifier is used for dynamic gain adjustment, the anti-aliasing filter is used to filter out high-frequency noise, and the ADC driver is used to condition the input signal to meet the input requirements of the synchronous ADC. The synchronous ADC is used to convert the signal into a digital signal.

[0029] To address the aforementioned technical problems, the present invention also provides a technical solution for a broadband oscillation detection system, as detailed below:

[0030] A broadband oscillation detection system according to the present invention includes a processor, the processor being configured to execute computer program instructions stored in a memory to implement the steps of the following method:

[0031] 1) Acquire the monitoring signal, perform time-frequency conversion on the monitoring signal, and obtain the original spectrum signal;

[0032] 2) Perform frequency domain convolution between the original spectral signal and the spectrum of a window function with spectral leakage suppression function, and obtain the denoised spectrum based on the frequency domain convolution result;

[0033] 3) Use the noise-reduced spectrum for broadband oscillation detection.

[0034] The beneficial effects of the above technical solution are as follows: The broadband oscillation detection system of the present invention provides the basic hardware support for realizing the broadband oscillation detection method of the present invention. Given that the traditional FFT using a rectangular window easily leads to energy dispersion of high-frequency oscillation components, the core idea of ​​the method of the present invention is to perform frequency domain convolution with the original frequency domain signal obtained after time-frequency conversion using the spectrum of a window function with spectral leakage suppression function. This suppresses spectral leakage caused by truncation effect in FFT operation through frequency domain convolution technology, improves the signal-to-noise ratio of weak high-frequency signals, improves the detection accuracy of broadband oscillations, and reduces the possibility of oscillation misjudgment.

[0035] Furthermore, the monitoring signal is time-frequency converted using a hybrid-based FFT. The hybrid-based FFT is used to: ① preferentially decompose the number of sampling points into 4 a ×8 b ×3 c ×5 d The combination of a,b,c,d≥0; then the remaining factors are decomposed into 2 factors; ② For the remaining factors that cannot be processed after step ①, if there is an optimized DFT kernel, the pre-generated optimized DFT kernel is called for processing; if there is no optimized DFT kernel, zeros are padded to the nearest neighbor number of points that can be completely decomposed by the strategy in step ①.

[0036] Furthermore, the radix-2, radix-4, and radix-8 twitch factors in the hybrid radix FFT are fixed as constants in the hardware logic of the butterfly operation unit; the radix-3 and radix-5 twitch factors in the hybrid radix FFT are pre-stored in memory and read through a lookup table; the inter-level twitch factors in the hybrid radix FFT calculation are all pre-calculated and stored in memory and read through a lookup table; the non-standard twitch factors in the hybrid radix FFT are calculated and generated in real time using the CORDIC algorithm.

[0037] Furthermore, the radix-8 in the hybrid radix FFT is decomposed into two levels of radix-4 operations. The two levels of radix-4 operations share a radix-4 butterfly unit, and the following timing scheduling strategy is used to generate the 8-point FFT output: the original 8-point input sequence is split into two 4-point subsequences, namely group 0 and group 1; during clock cycles T1~T4, the radix-4 butterfly unit is enabled, and the data of group 0 is sent to the radix-4 butterfly unit for radix-4 operation to obtain the 4-point intermediate result Y0; during clock cycles T5~T8, the data of group 1 is sent to the radix-4 butterfly unit for radix-4 operation to obtain the 4-point intermediate result Y1; the intermediate result Y0 is read and multiplied with the pre-stored rotation factor in an independent multiplier array to obtain Y0'; during clock cycles T9~T12, the result Y0' and Y1 are sent as input to the radix-4 butterfly unit to add Y0' and Y1, thereby generating the 8-point FFT output; T1, T2, ..., T11, T12 are 12 clock cycles that advance according to time.

[0038] Furthermore, when resources are limited, the hybrid radix FFT switches radix 3 and radix 5 to radix 2 mode and schedules them to idle periods to perform radix 3 and radix 5 operations using radix 2 operation units.

[0039] Furthermore, the window function with spectral leakage suppression function is the Hanning window.

[0040] Furthermore, the frequency domain coefficients of the Hanning window are selected based on the signal-to-noise ratio (SNR) of the monitored signal; when the SNR of the monitored signal is less than the set SNR threshold, the Hanning window is enabled in wide window mode; otherwise, the Hanning window is enabled in narrow window mode. The frequency domain broadening coefficient in wide window mode is less than that in narrow window mode.

[0041] Furthermore, the method for obtaining the denoised spectrum based on the frequency domain convolution result includes: inversely compensating the phase of the frequency domain convolution result, i.e., the compensated phase is the original phase of the frequency domain convolution result minus the phase shift, where the phase shift is the phase of the Hanning window's spectrum; normalizing the amplitude of the frequency domain convolution result using a window energy correction coefficient, where the normalized amplitude is the ratio of the original amplitude of the frequency domain convolution result to the window energy correction coefficient; the window energy correction coefficient K is:

[0042]

[0043] In the formula, w(n) is the time-domain expression of the Hanning window, and L is the total number of points of the Hanning window w(n).

[0044] Furthermore, the method of using the denoised spectrum for broadband oscillation detection includes using the denoised spectrum to identify whether oscillation exists, and the method of identifying whether oscillation exists includes: determining the amplitude of each frequency point based on the denoised spectrum; if the amplitude of a certain frequency point continuously exceeds the set amplitude threshold, it is considered that an oscillation event has occurred.

[0045] Furthermore, the method of broadband oscillation detection using the denoised spectrum includes oscillation pattern recognition using the denoised spectrum, where the oscillation patterns include at least subsynchronous, supersynchronous, and harmonic resonance. The method of identifying oscillation patterns includes: extracting the amplitude of the denoised spectrum and then squaring the extracted amplitude to obtain the energy spectrum of each frequency point; dividing the entire frequency band into multiple sub-bands and calculating the energy integral value of each sub-band within a set time window; forming an energy integral feature vector from the energy integral values ​​of each sub-band and inputting it into the oscillation pattern recognition model to obtain the oscillation pattern recognition result; the oscillation pattern recognition model is obtained by training a classification model using a dataset containing the energy integral feature vector and the corresponding oscillation patterns.

[0046] Furthermore, the method of broadband oscillation detection using the denoised spectrum includes locating the oscillation source using the denoised spectrum. The method for locating the oscillation source includes: obtaining the phase difference of the denoised spectrum corresponding to the monitoring signals of each pair of source points in the power grid through a PMU, and calculating the signal arrival time difference based on the phase difference; preliminarily determining the oscillation source based on the time difference data, combined with the electrical geometry model of the power grid and the time difference positioning principle, and forming an initial candidate source set from the preliminarily determined oscillation sources; traversing the initial candidate source set to determine whether any source point meets the node priority matching condition, where the node priority matching condition refers to the source point falling into a network node in the power grid topology GIS database or the allowable error centered on the network node. Within the specified range, source points that meet the node priority matching condition are identified as oscillation sources. Then, for source points that do not meet the node priority matching condition, it is determined whether any source point meets the line projection matching condition, where the line projection matching condition refers to the source point falling within the buffer zone of a transmission line. For source points that meet the line projection matching condition, the precise location of the oscillation source on that line is calculated through vertical projection. Then, for source points that do not meet the line projection matching condition, the minimum electrical distance from the source point to all topological elements in the power grid topology is calculated, and the source point is assigned to the topological element with the smallest distance. These topological elements include nodes and lines. Finally, the confidence level of all the above location results is calculated, and the oscillation source is determined based on the confidence level.

[0047] Furthermore, the confidence factors used to calculate the confidence level include consistency based on time difference measurement, the degree of fit between the source point and the topology element, and the dispersion among multiple candidate points. Consistency based on time difference measurement measures the degree of consistency when the time difference data calculated by different PMU nodes is used to infer the location of the oscillation source. The degree of fit between the source point and the topology element quantifies the physical proximity of the candidate source point and the grid topology element it matches.

[0048] Furthermore, the method also includes determining the risk level of the identified oscillation, and the determination method includes: extracting four key features of the identified oscillation: amplitude, frequency band, duration, and rate of change; performing a weighted summation of the four key features to obtain a comprehensive risk value; determining the risk level of the oscillation event based on the interval in which the comprehensive risk value is located, and triggering different alarm strategies for oscillation events of different risk levels.

[0049] Furthermore, the monitoring signal is the signal obtained by sequentially passing the original acquired signal through an RVS protection circuit, a differential input buffer, an instrumentation amplifier, a programmable gain amplifier, an anti-aliasing filter, an ADC driver, and a synchronous ADC. The RVS protection circuit is used for overvoltage clamping, the differential input buffer is used for impedance transformation, the instrumentation amplifier is used for differential amplification and common-mode rejection, the programmable gain amplifier is used for dynamic gain adjustment, the anti-aliasing filter is used to filter out high-frequency noise, and the ADC driver is used to condition the input signal to meet the input requirements of the synchronous ADC. The synchronous ADC is used to convert the signal into a digital signal.

[0050] To address the aforementioned technical problems, the present invention also provides a technical solution for a computer-readable storage medium, as detailed below:

[0051] The present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the steps of the following method:

[0052] 1) Acquire the monitoring signal, perform time-frequency conversion on the monitoring signal, and obtain the original spectrum signal;

[0053] 2) Perform frequency domain convolution between the original spectral signal and the spectrum of a window function with spectral leakage suppression function, and obtain the denoised spectrum based on the frequency domain convolution result;

[0054] 3) Use the noise-reduced spectrum for broadband oscillation detection.

[0055] The beneficial effects of the above technical solution are as follows: The computer-readable storage medium of this invention stores the software code corresponding to the broadband oscillation detection method of this invention, laying the foundation for the reliable implementation of this method. Given that the traditional FFT using a rectangular window easily leads to energy dispersion of high-frequency oscillation components, the core idea of ​​this invention is to perform frequency domain convolution with the original frequency domain signal obtained after time-frequency conversion using the spectrum of a window function with spectral leakage suppression function. This frequency domain convolution technique suppresses spectral leakage caused by truncation effect in FFT operations, improves the signal-to-noise ratio of weak high-frequency signals, enhances the detection accuracy of broadband oscillations, and reduces the possibility of oscillation misjudgment.

[0056] Furthermore, the monitoring signal is time-frequency converted using a hybrid-based FFT. The hybrid-based FFT is used to: ① preferentially decompose the number of sampling points into 4 a ×8 b ×3 c ×5 dThe combination of a,b,c,d≥0; then the remaining factors are decomposed into 2 factors; ② For the remaining factors that cannot be processed after step ①, if there is an optimized DFT kernel, the pre-generated optimized DFT kernel is called for processing; if there is no optimized DFT kernel, zeros are padded to the nearest neighbor number of points that can be completely decomposed by the strategy in step ①.

[0057] Furthermore, the radix-2, radix-4, and radix-8 twitch factors in the hybrid radix FFT are fixed as constants in the hardware logic of the butterfly operation unit; the radix-3 and radix-5 twitch factors in the hybrid radix FFT are pre-stored in memory and read through a lookup table; the inter-level twitch factors in the hybrid radix FFT calculation are all pre-calculated and stored in memory and read through a lookup table; the non-standard twitch factors in the hybrid radix FFT are calculated and generated in real time using the CORDIC algorithm.

[0058] Furthermore, the radix-8 in the hybrid radix FFT is decomposed into two levels of radix-4 operations. The two levels of radix-4 operations share a radix-4 butterfly unit, and the following timing scheduling strategy is used to generate the 8-point FFT output: the original 8-point input sequence is split into two 4-point subsequences, namely group 0 and group 1; during clock cycles T1~T4, the radix-4 butterfly unit is enabled, and the data of group 0 is sent to the radix-4 butterfly unit for radix-4 operation to obtain the 4-point intermediate result Y0; during clock cycles T5~T8, the data of group 1 is sent to the radix-4 butterfly unit for radix-4 operation to obtain the 4-point intermediate result Y1; the intermediate result Y0 is read and multiplied with the pre-stored rotation factor in an independent multiplier array to obtain Y0'; during clock cycles T9~T12, the result Y0' and Y1 are sent as input to the radix-4 butterfly unit to add Y0' and Y1, thereby generating the 8-point FFT output; T1, T2, ..., T11, T12 are 12 clock cycles that advance according to time.

[0059] Furthermore, when resources are limited, the hybrid radix FFT switches radix 3 and radix 5 to radix 2 mode and schedules them to idle periods to perform radix 3 and radix 5 operations using radix 2 operation units.

[0060] Furthermore, the window function with spectral leakage suppression function is the Hanning window.

[0061] Furthermore, the frequency domain coefficients of the Hanning window are selected based on the signal-to-noise ratio (SNR) of the monitored signal; when the SNR of the monitored signal is less than the set SNR threshold, the Hanning window is enabled in wide window mode; otherwise, the Hanning window is enabled in narrow window mode. The frequency domain broadening coefficient in wide window mode is less than that in narrow window mode.

[0062] Furthermore, the method for obtaining the denoised spectrum based on the frequency domain convolution result includes: inversely compensating the phase of the frequency domain convolution result, i.e., the compensated phase is the original phase of the frequency domain convolution result minus the phase shift, where the phase shift is the phase of the Hanning window's spectrum; normalizing the amplitude of the frequency domain convolution result using a window energy correction coefficient, where the normalized amplitude is the ratio of the original amplitude of the frequency domain convolution result to the window energy correction coefficient; the window energy correction coefficient K is:

[0063]

[0064] In the formula, w(n) is the time-domain expression of the Hanning window, and L is the total number of points of the Hanning window w(n).

[0065] Furthermore, the method of using the denoised spectrum for broadband oscillation detection includes using the denoised spectrum to identify whether oscillation exists, and the method of identifying whether oscillation exists includes: determining the amplitude of each frequency point based on the denoised spectrum; if the amplitude of a certain frequency point continuously exceeds the set amplitude threshold, it is considered that an oscillation event has occurred.

[0066] Furthermore, the method of broadband oscillation detection using the denoised spectrum includes oscillation pattern recognition using the denoised spectrum, where the oscillation patterns include at least subsynchronous, supersynchronous, and harmonic resonance. The method of identifying oscillation patterns includes: extracting the amplitude of the denoised spectrum and then squaring the extracted amplitude to obtain the energy spectrum of each frequency point; dividing the entire frequency band into multiple sub-bands and calculating the energy integral value of each sub-band within a set time window; forming an energy integral feature vector from the energy integral values ​​of each sub-band and inputting it into the oscillation pattern recognition model to obtain the oscillation pattern recognition result; the oscillation pattern recognition model is obtained by training a classification model using a dataset containing the energy integral feature vector and the corresponding oscillation patterns.

[0067] Furthermore, the method of broadband oscillation detection using the denoised spectrum includes locating the oscillation source using the denoised spectrum. The method for locating the oscillation source includes: obtaining the phase difference of the denoised spectrum corresponding to the monitoring signals of each pair of source points in the power grid through a PMU, and calculating the signal arrival time difference based on the phase difference; preliminarily determining the oscillation source based on the time difference data, combined with the electrical geometric model of the power grid and the time difference positioning principle, and forming an initial candidate source set by preliminarily determining the oscillation sources; traversing the initial candidate source set to determine whether any source point meets the node priority matching condition, where the node priority matching condition refers to the source point falling into a certain network node in the power grid topology GIS database or within an allowable error range centered on the network node. Within the specified range, source points that meet the node priority matching condition are identified as oscillation sources. Then, for source points that do not meet the node priority matching condition, it is determined whether any source point meets the line projection matching condition, where the line projection matching condition refers to the source point falling within the buffer zone of a transmission line. For source points that meet the line projection matching condition, the precise location of the oscillation source on that line is calculated through vertical projection. Then, for source points that do not meet the line projection matching condition, the minimum electrical distance from the source point to all topological elements in the power grid topology is calculated, and the source point is assigned to the topological element with the smallest distance. These topological elements include nodes and lines. Finally, the confidence level of all the above location results is calculated, and the oscillation source is determined based on the confidence level.

[0068] Furthermore, the confidence factors used to calculate the confidence level include consistency based on time difference measurement, the degree of fit between the source point and the topology element, and the dispersion among multiple candidate points. Consistency based on time difference measurement measures the degree of consistency when the time difference data calculated by different PMU nodes is used to infer the location of the oscillation source. The degree of fit between the source point and the topology element quantifies the physical proximity of the candidate source point and the grid topology element it matches.

[0069] Furthermore, the method also includes determining the risk level of the identified oscillation, and the determination method includes: extracting four key features of the identified oscillation: amplitude, frequency band, duration, and rate of change; performing a weighted summation of the four key features to obtain a comprehensive risk value; determining the risk level of the oscillation event based on the interval in which the comprehensive risk value is located, and triggering different alarm strategies for oscillation events of different risk levels.

[0070] Furthermore, the monitoring signal is the signal obtained by sequentially passing the original acquired signal through an RVS protection circuit, a differential input buffer, an instrumentation amplifier, a programmable gain amplifier, an anti-aliasing filter, an ADC driver, and a synchronous ADC. The RVS protection circuit is used for overvoltage clamping, the differential input buffer is used for impedance transformation, the instrumentation amplifier is used for differential amplification and common-mode rejection, the programmable gain amplifier is used for dynamic gain adjustment, the anti-aliasing filter is used to filter out high-frequency noise, and the ADC driver is used to condition the input signal to meet the input requirements of the synchronous ADC. The synchronous ADC is used to convert the signal into a digital signal. Attached Figure Description

[0071] Figure 1 This is an overall framework diagram of the broadband oscillation detection method of the present invention;

[0072] Figure 2 This is a flowchart of the hybrid-based FFT processing of the present invention;

[0073] Figure 3 This is a flowchart of the Hanning window convolution module processing of the present invention;

[0074] Figure 4 This is a flowchart of the oscillation decision module of the present invention;

[0075] Figure 5 This is a flowchart of the timing scheduling strategy of the present invention;

[0076] Figure 6 This is a decision-making flowchart for locating the oscillation source of the present invention. Detailed Implementation

[0077] The broadband oscillation detection method of the present invention comprises the following steps: First, acquiring a monitoring signal and performing time-frequency conversion on the monitoring signal to obtain the original spectrum signal; then, performing frequency domain convolution between the original spectrum signal and the spectrum of a window function with spectral leakage suppression function, and obtaining the denoised spectrum based on the frequency domain convolution result; finally, using the denoised spectrum for broadband oscillation detection. This method suppresses spectral leakage caused by truncation effect in FFT operations through frequency domain convolution technology, improves the signal-to-noise ratio of weak high-frequency signals, enhances the detection accuracy of broadband oscillations, and reduces the possibility of oscillation misjudgment. To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings.

[0078] An implementation method for broadband oscillation detection:

[0079] A broadband oscillation detection method is proposed. This method employs a five-stage pipeline architecture: signal acquisition, preprocessing, mixed-basis FFT, Hanning window convolution, and oscillation decision. Through modular hardware design and algorithmic optimization, combined with an FPGA+ARM heterogeneous computing platform, it achieves closed-loop processing of broadband oscillation signals from acquisition to transformation, analysis, and decision. It also enables real-time detection and localization of broadband oscillations. Through dual innovation of hardware acceleration and algorithm adaptation, it overcomes the challenges of real-time performance, accuracy, and reliability in broadband oscillation detection in high-penetration renewable energy power grids. The overall system architecture is as follows: Figure 1 As shown, the modules in the system are described below:

[0080] 1) The signal acquisition module adopts a wide-bandwidth (0.1-2500Hz) and high common-mode rejection ratio (≥120dB) analog front-end circuit, which solves the problem of accurate extraction of small oscillation signals (as low as 10mV) under high voltage scenarios.

[0081] 2) The preprocessing module integrates a synchronous ADC to convert analog signals into 16-bit digital signal streams and achieves seamless data connection through a ping-pong buffering mechanism to avoid spectral aliasing caused by sampling intervals.

[0082] 3) The hybrid radix FFT module is based on the FPGA to dynamically reconstruct the butterfly operation unit and supports multi-level decomposition strategies such as radix 4, radix 3, and radix 5. It can compress the number of FFT operation levels of 8192 points from 13 levels in the traditional radix 2 to 6~8 levels, breaking through the timing convergence bottleneck.

[0083] 4) The Hanning window convolution module, based on FPGA, performs noise suppression and spectrum correction in the frequency domain by pre-stored frequency domain window function coefficients (stored in Block RAM) and a parallel multiply-accumulate array (16 DSP slices), avoiding the delay accumulation of time domain convolution.

[0084] 5) The oscillation decision module relies on the real-time threshold calculation and topology matching algorithm of the ARM processor to realize the rapid location (error ≤ ±5Hz) and risk level classification of multi-source oscillations.

[0085] The following section provides a detailed introduction to the broadband oscillation detection method implemented by the signal acquisition module, preprocessing module, mixed-basis FFT module, Hanning window convolution module, and oscillation decision module, based on the functional descriptions of each module.

[0086] Step 1: Use the signal acquisition module to acquire the high-voltage input signal (also known as the monitoring signal) and output it to the synchronous ADC of the preprocessing module.

[0087] The high-voltage input signal here specifically refers to the voltage or current signal acquired from the primary side (high-voltage side) of the power grid through a dedicated sensor, which reflects the wideband dynamic characteristics of the system. Specific examples are shown in Table 1 below.

[0088] Table 1

[0089]

[0090] Furthermore, the signal acquisition module employs a wide-bandwidth (0.1~2500Hz) and high common-mode rejection ratio (≥120dB) analog front-end circuit to ensure wide-bandwidth response and high common-mode rejection ratio, thus addressing the problem of accurately extracting minute oscillating signals (as low as 10mV) under high-voltage scenarios. The overall framework of the analog front-end circuit is a multi-stage processing chain, comprising "RVS protection circuit → differential input buffer → instrumentation amplifier → programmable gain amplifier → anti-aliasing filter → ADC driver." These stages are cascaded and connected in the order of signal flow, collectively achieving signal protection, isolation, amplification, filtering, and driving functions, thereby meeting the requirements of wide-bandwidth (0.1~2500Hz), high common-mode rejection ratio (≥120dB), and minute signal extraction (as low as 10mV). Finally, the extracted signal is transmitted to the hybrid-based FFT module via the ADC. Details of these key stages are as follows:

[0091] 1) RVS protection circuit: Located at the very front of the signal acquisition module, it is directly connected to the high-voltage input signal. Its function is to protect the subsequent amplification and filtering circuits from damage caused by transient overvoltage in high-voltage scenarios, while ensuring that the signal is transmitted without distortion over a wide frequency band.

[0092] 2) Differential input buffer: A unity-gain buffer (also known as a voltage follower) is constructed using an operational amplifier with high input impedance (≥1MΩ) to achieve signal isolation and impedance matching, and reduce the signal source load effect.

[0093] 3) Instrumentation Amplifier: The core amplification unit, employing a multi-stage operational amplifier differential structure, uses a precision resistor network to set the gain (typically 10~100 times), providing a high common-mode rejection ratio (≥120dB). This stage suppresses common-mode noise and supports wideband response through symmetrical layout and temperature compensation technology.

[0094] 4) Programmable gain amplifier (PGA): Based on digital potentiometer, the gain is dynamically adjusted (1~1000 times) to adapt to oscillation signals of different amplitudes (10mV~10V) and ensure that the signal is linearly amplified to the ADC input range.

[0095] 5) Anti-aliasing filter: A Butterworth low-pass filter (cutoff frequency ≥ 2.5kHz, order ≥ 4) is used. Through a hybrid design of passive LC and active RC, high-frequency noise and aliasing components are suppressed while maintaining phase linearity.

[0096] 6) ADC driver: The last stage uses a high-speed operational amplifier to provide low output impedance (≤50Ω) and high slew rate (≥100V / μs) to drive the sample-and-hold circuit of the synchronous ADC, ensuring signal integrity and conditioning the input signal to meet the input requirements of the ADC.

[0097] The specific signal processing steps for each stage are as follows: the high-voltage input signal first undergoes overvoltage clamping via the RVS protection circuit, then enters the differential input buffer for impedance transformation; the buffered signal is sent to the instrumentation amplifier for differential amplification and common-mode rejection; the amplified signal undergoes dynamic gain adjustment via a programmable gain amplifier; subsequently, it passes through an anti-aliasing filter to remove high-frequency noise; finally, it is buffered by the ADC driver and output to the synchronous ADC in the preprocessing module. The entire circuit employs a symmetrical layout and shielding design to reduce electromagnetic interference and ensure wide bandwidth and high common-mode rejection ratio characteristics.

[0098] Step 2: Use the preprocessing module to convert the signal output from the signal acquisition module into a digital signal stream output.

[0099] The preprocessing module includes a synchronous ADC, which can convert analog signals into 16-bit digital signal streams and achieve seamless data connection through a ping-pong buffering mechanism to avoid spectral aliasing caused by sampling intervals.

[0100] Since the measurement here is not a single measurement, but a continuous sampling of the analog signal at a fixed rate, a digital sequence arranged in chronological order, or "stream," is generated—a continuous data sequence. Furthermore, the "16-bit digital signal stream" here can be understood as a series of high-precision readings output after a "high-precision digital ruler" measures the analog signal. "16-bit" refers to the resolution of the ADC; for a 16-bit synchronous ADC, it can divide the amplitude range of the analog signal into 65,536 different levels. Of course, the specific number of bits can be set according to the actual situation; 16 bits is a well-balanced solution for the application scenario of this implementation. The choice of 16 bits is based on a comprehensive consideration of accuracy requirements, cost, power consumption, and data processing burden. A comparison of various bit widths is shown in Table 2 below:

[0101] Table 2

[0102]

[0103] Step 3: The signal output from Step 2 is processed by a mixed-base FFT module to convert the time-domain signal into a frequency-domain signal to obtain the original spectrum signal.

[0104] The hybrid-radix FFT processing module is the core computing unit of this invention. Implemented on an FPGA, it solves key problems of low computational efficiency, high resource consumption, and insufficient real-time performance in broadband oscillation detection using traditional FFT algorithms through a dynamic multi-radix decomposition strategy, hardware-friendly twiddle factor optimization, reconfigurable butterfly operation units, and an efficient data scheduling mechanism. This significantly improves the detection accuracy and response speed of broadband oscillations (0.1~2000Hz) in high-penetration renewable energy power grids. The overall processing flow is as follows: Figure 2 As shown.

[0105] 1) Dynamic multi-cardinality decomposition rules and series compression.

[0106] ① First Priority (High Cardinality Decomposition Strategy): For the number of sampling points N, adopt the principle of "high cardinality first, mixed decomposition". Decompose N into 4 as much as possible. a ×8 b ×3 c ×5 d Combinations (a,b,c,d≥0) are used to maximize the use of efficient radix-4, radix-8, radix-3, and radix-5 butterfly arithmetic units, thereby compressing the number of arithmetic stages and reducing the number of multiplications.

[0107] ② Second priority (radix-2 decomposition): After performing high-radix decomposition, if the remaining factors are powers of 2 (i.e., 2^2 / 2^2), then... e If e≥1), then the radix-2 butterfly operation unit is invoked for processing. Radix-2, as the most basic and flexible operation unit, is the foundation for constructing the entire FFT algorithm.

[0108] ③ Universality Guarantee Mechanism: For the remaining prime factors that cannot be processed after the above two-step decomposition (i.e., small prime numbers that are neither 2, 3, nor 5, nor can they be covered by high cardinality combinations, such as 7, 11, 13, etc.), if there is an optimized DFT kernel for the small prime number, the pre-generated optimized small prime number DFT kernel is called for processing; if resources do not allow or there is no corresponding optimized kernel, intelligent zero-padding is used to the nearest neighbor number that can be completely decomposed by the above strategy.

[0109] 2) Rotation factor hierarchical optimization and storage compression.

[0110] ① Conventional Twist Factor Pre-stored and Fixed: For operations with strong regularity such as radix 2, radix 4, and radix 8, the twist factor is fixed as a constant in the hardware logic of the butterfly operation unit, achieving zero-latency access.

[0111] ② Small radix factor block storage: For fixed small radixes such as radix 3 and radix 5, their limited and fixed rotation factors are pre-calculated and stored in the FPGA's Block RAM, and can be quickly read through a lookup table.

[0112] ③ Standard process factor pre-stored: For a preset typical number of sampling points N, all necessary inter-stage rotation factors in the complete FFT calculation process are pre-calculated and stored in an orderly manner in a large Block RAM lookup table, supporting continuous access in the pipeline.

[0113] ④ Real-time generation of dynamic factors: Non-standard rotation factors generated during dynamic adaptation and irregular decomposition are calculated and generated in real time using the CORDIC algorithm. This strategy avoids allocating dedicated storage for low-frequency factors, achieving an optimal balance between storage resources and computational flexibility.

[0114] 3) Reconfigurable butterfly operation unit and parallel acceleration.

[0115] ① Radix-2 unit: using Radix-2 2 The structure requires only 3 real number multiplications per level and supports a throughput of 4 points per cycle.

[0116] ②Base 3 unit: utilizing symmetry ( The number of complex multiplications in a 3-point DFT is reduced from 4 to 2, and resource efficiency is improved by reusing adders.

[0117] ③ Radix-5 Unit: Based on the Winograd FFT algorithm, the number of multiplications in a 5-point DFT is reduced from 10 to 5, and critical path latency is reduced through a pre-adding / post-adding network. The "pre-adding / post-adding network" here is the concrete manifestation and core component of the "Winograd FFT algorithm" in hardware implementation. The Winograd algorithm replaces complex complex multiplication operations with efficient linear operations, thereby achieving a significant reduction in the number of multiplications. Winograd FFT is a DFT optimization algorithm specifically designed for small point numbers N, and its mathematical model is as follows:

[0118] Y=C×((B×x)⊙M)

[0119] In the formula, x is the input vector; Y is the output vector; B is the matrix responsible for linearly combining the input signals, i.e., the "pre-addition network"; M is a diagonal matrix composed of rotation factors; ⊙ represents dot product, which is the only multiplication operation required in the whole process. The essence of the Winograd algorithm is to reduce the number of multiplications to the theoretical lower limit; C is responsible for linearly combining the multiplication results again to obtain the final DFT output, i.e., the "post-addition network".

[0120] ④ Radix-8 Unit Reuse of Radix-4 Hardware: The radix-8 operation is decomposed into two levels of radix-4 operations, sharing a radix-4 butterfly unit. A timing scheduling strategy avoids resource conflicts to complete the 8-point FFT operation. Specifically, this timing scheduling strategy is a "time-division multiplexing and ping-pong buffering" strategy, achieving safe and efficient reuse of the radix-4 butterfly hardware unit for radix-8 operations. The core of this strategy lies in decoupling a single radix-8 operation in time into two independent radix-4 operation stages, and managing the data flow through precise control logic, thereby avoiding concurrent access conflicts to the same hardware resource. The process is as follows: Figure 5 As shown, it specifically includes the following steps:

[0121] (a) Data allocation and caching stage: The control unit splits the original 8-point input sequence into two 4-point subsequences, namely group 0 and group 1. This operation is equivalent to performing a 2×4 matrix transpose on the input data in hardware.

[0122] (b) First stage radix-4 operation: During clock cycles T1~T4, the scheduler enables the radix-4 butterfly unit, reads the data of group 0 from the buffer, completes the first radix-4 operation, and obtains 4 intermediate results Y0 (Y0=FFT4); these 4 intermediate results Y0 are written into an intermediate transpose buffer.

[0123] (c) Twist Factor Multiplication and Data Preparation: During clock cycles T5-T8, the radix-4 butterfly unit enters the next stage. The scheduler performs two operations: it sends the data from group 1 into the radix-4 butterfly unit, generating the result Y1 (Y1=FFT4); simultaneously, it reads Y0 from the intermediate buffer and performs complex multiplication with the pre-stored twist factor W in an independent multiplier array to obtain Y0'=Y0×W. That is, in this stage, the radix-4 unit performs operations on group 1 to generate Y1, while the multiplier array modulates the twist factor of Y0 in parallel to obtain Y0'.

[0124] (d) Second stage radix-4 operation: During clock cycles T9~T12, the scheduler combines the result Y0' after the twiddle factor multiplication with the previously calculated Y1, and uses Y0' and Y1 as inputs to feed them into the radix-4 butterfly unit for the second time. This operation is not a standard 4-point FFT, but a post-processing combination network composed of adders. It adds Y0' and Y1 according to the radix-8 algorithm rules, and finally generates an 8-point complete FFT output.

[0125] The T1, T2, ..., T11, T12 mentioned above are 12 clock cycles that progress according to time.

[0126] ⑤ Dynamic Time-Division Multiplexing: When resources are limited, radix-3 / 5 units are switched to radix-2 mode, utilizing idle periods to complete multi-level radix-2 operations, improving hardware utilization. Here, "idle period" refers to a predictable idle time window created by the system resource scheduler during high-performance computing units' task execution due to their longer pipeline delays. Only the radix-2 operations dynamically switched from radix-3 / 5 are scheduled for execution within the "idle period." The system's native, inherent radix-2 computing units continue to run in their dedicated, conventional time slots.

[0127] Step four: The original frequency domain signal obtained in step three is subjected to frequency domain convolution using the Hanning window convolution module, and the denoised spectrum is obtained based on the frequency domain convolution result.

[0128] The Hanning window convolution module is the core noise reduction and spectral correction unit in broadband oscillation detection. It suppresses spectral leakage caused by truncation effects in FFT operations through frequency domain convolution technology and improves the signal-to-noise ratio of weak high-frequency signals. Based on a parametric Hanning window design and combined with a hardware acceleration architecture, this module achieves high-fidelity extraction of broadband oscillation features while ensuring phase accuracy. The Hanning window convolution module employs a four-stage processing flow: window function generation, frequency domain coefficient pre-storage, parallel convolution, phase compensation, and amplitude calibration. It works in conjunction with the hybrid-based FFT module, and the processing flow is as follows: Figure 3 As shown, the details are as follows:

[0129] 1) Window function parameterization design and frequency domain convolution: Frequency domain convolution technology is used to suppress spectral leakage and improve the signal-to-noise ratio. A parameterized Hanning window design is adopted, and the optimal window function configuration is dynamically selected according to the signal-to-noise ratio of the real-time signal. Specifically, the original spectral signal is convolved with the spectrum of the window function with spectral leakage suppression function in the frequency domain.

[0130] ① Dual-mode window function library: Two sets of optimized Hanning window frequency domain coefficients H(k) are pre-stored in the Block RAM of the FPGA, and the appropriate Hanning window frequency domain coefficients H(k) are selected according to the signal-to-noise ratio of the monitored signal.

[0131] When SNR < 10dB, the wide window mode (high resolution mode) is enabled. This mode corresponds to a longer time domain window (window width W = FFT points N), with a narrow main lobe in the frequency domain (i.e., a small "frequency domain broadening factor"), resulting in high frequency resolution and effectively separating weak oscillation components from strong noise. When SNR ≥ 10dB, the narrow window mode (high positioning accuracy mode) is enabled. This mode corresponds to a shorter time domain window (window width W = N / 4), with a wider main lobe in the frequency domain (i.e., a large "frequency domain broadening factor"). Although it sacrifices some frequency resolution, it improves time positioning accuracy and is beneficial for capturing high-frequency transient oscillations.

[0132] Furthermore, adjusting the frequency domain broadening factor is equivalent to switching between window functions with different time domain lengths. According to the scaling property of the Fourier transform, compression of the time-domain signal leads to broadening of its frequency domain representation. Therefore: when a narrow window is selected, the main lobe of the corresponding frequency domain window function H(k) is naturally wider (i.e., the "frequency domain broadening factor" is large); when a wide window is selected, the main lobe of the corresponding frequency domain window function H(k) is naturally narrower (i.e., the "frequency domain broadening factor" is small).

[0133] The frequency domain coefficients H(k) are pre-converted into complex coefficients using FFT and then quantized into 16-bit fixed-point numbers. Utilizing the conjugate symmetry of the Hanning window, only the first N / 2+1 frequency coefficients are stored, reducing storage overhead by 50%.

[0134] ② Frequency domain convolution execution: The frequency domain coefficients H(k) of the selected mode are directly read and multiplied pointwise by point with the original spectrum signal F(k) output by the mixed-basis FFT module (i.e., frequency domain convolution) to complete noise suppression and spectrum correction. That is, the FFT result F(k) is multiplied pointwise by the window coefficients H(k) in the frequency domain and the complex multiplication and accumulation are completed by a 16-channel parallel multiplier-accumulator array. 16 pairs of data are processed in a single cycle. The multiplier-accumulator supports saturation and rounding modes to ensure that the operation accuracy error is <0.05%.

[0135] The frequency domain characteristics of all window functions are pre-computed and optimized, avoiding the resource consumption of real-time computation. That is, in hardware implementation, within the FPGA, the frequency domain coefficients H_wide(k) and H_narrow(k) corresponding to the wide and narrow modes of the window function are pre-computed and stored in Block RAM. Switching modes only requires changing the coefficient lookup table used.

[0136] ③ Phase compensation and amplitude calibration: The phase of the frequency domain convolution result is reverse-compensated, that is, the compensated phase is the original phase of the frequency domain convolution result minus the phase shift, and the phase shift is the phase of the Hanning window spectrum; and the amplitude of the frequency domain convolution result is normalized by the window energy correction coefficient, and the normalized amplitude is the ratio of the original amplitude of the frequency domain convolution result to the window energy correction coefficient.

[0137] When performing convolution in the frequency domain (i.e., complex multiplication G(k) = F(k) ⊙ H(k)), the frequency domain response H(k) of the window function (i.e., the spectrum of the window function) is a complex number, and its phase arg(H(k)) is directly superimposed on the spectrum F(k) of the original signal. This causes a phase distortion in the output G(k), which affects subsequent phase-based accurate positioning. Therefore, the Hanning window introduces a phase shift. This value is the most direct and accurate, representing the inherent phase response of the window function at each frequency point k. Since the window function is known in advance, its complete frequency domain response H(k) should be accurately calculated and stored in advance through a single FFT. Therefore, directly using this pre-calculated arg(H(k)) during compensation is the optimal strategy. Thus, the phase shift introduced by the Hanning window... Inverse compensation is performed on the phase of the convolution result G(k). ), This eliminates the distortion of the phase information of the oscillating signal by the window function.

[0138] At the same time, through the window energy correction coefficient Amplitude normalization corrects spectral amplitude distortion caused by the attenuation of the window function amplitude. Here, w(n) is the time-domain expression of the Hanning window, and L is the total number of points of the Hanning window w(n).

[0139] Therefore, the above correction process first performs amplitude normalization, then phase compensation, resulting in the final corrected spectrum. The expression for (also known as the denoised spectrum) is:

[0140]

[0141] In the formula, G(k)=F(k)⊙H(k) is the original result after frequency domain convolution; |G(k)| is the amplitude of G(k); arg(G(k)) is the phase of G(k); and K is the window energy correction coefficient. It is the phase difference introduced by the window energy.

[0142] Step 5: Input the output results from Step 4 into the oscillation decision module for oscillation decision.

[0143] The oscillation decision module is the final decision-making unit of the broadband oscillation detection device. Executed by the ARM (Automatic Power Array), it is responsible for identifying oscillation characteristics, determining the risk level, and locating the oscillation source from the denoised spectrum data. Based on a dynamic threshold algorithm, multi-source phase difference localization technology, and power grid topology matching strategy, this module achieves fully automated processing from "spectrum analysis" to "fault alarm." The processing flow is as follows: Figure 4 As shown. Core functions include:

[0144] 1) Identify the presence of oscillation events: If the amplitude at a certain frequency point continuously exceeds a set amplitude threshold, an oscillation event is considered to have occurred. Specifically:

[0145] The amplitude threshold set here is based on the 3σ value of the background noise in each frequency band according to historical data statistics as the initial threshold. When the amplitude at a certain frequency exceeds the threshold for 5 consecutive cycles, the threshold for that frequency band is adaptively increased by 20% to prevent false alarms. This cycle belongs to the threshold adjustment event determination cycle.

[0146] The “background noise at each frequency band” here refers to the signal amplitude at each frequency point measured and recorded by this device over a long period of time when the power system is operating stably and without oscillation. These amplitudes constitute a “background noise sample set”, and σ is the standard deviation of this sample set. It quantifies the natural fluctuation range of the inherent noise at each frequency point of the power grid under normal conditions.

[0147] The 3σ threshold originates from the statistical "3σ criterion." For a dataset following a normal distribution, 99.7% of the data points will fall within the range of ±3σ around the mean μ, with only about 0.3% falling outside this range. For each frequency point k, the amplitude data of that frequency point is extracted from all the "background spectrum" to form a data sequence. The mean μ_k and standard deviation σ_k of this sequence are calculated, and the initial threshold T_k = μ_k + 3 × σ_k.

[0148] In the absence of oscillations, the noise amplitude at this frequency will not exceed this limit 99.7% of the time. Therefore, if the amplitude at a certain frequency continuously exceeds this threshold, it is considered with a very high confidence (99.7%) that this is not normal noise fluctuation, but an abnormal oscillation event that requires attention. It should be noted that the "continuous" in this paragraph is a different period from the "continuous" when adjusting the threshold (i.e., exceeding the threshold for 5 consecutive cycles mentioned earlier). The "continuous" in the oscillation event judgment in this paragraph is a short-term, sensitive trigger condition used to quickly capture the onset of oscillations, while the "continuous" in the threshold adjustment event is a long-term, conservative confirmation condition used to carefully confirm that the noise level has undergone a permanent change, thereby avoiding erroneous adjustments.

[0149] 2) Identify oscillation modes: Calculate the energy integral value of each sub-band across the entire frequency band from 0.1 to 2000 Hz. Combine the energy integral values ​​of each sub-band into an energy integral feature vector, which is then input into the oscillation mode identification model to identify the oscillation mode (subsynchronous / supersynchronous / harmonic resonance). Specifically:

[0150] ① Energy Calculation: The amplitude of the spectrum result G(k) output by the Hanning window convolution module is extracted and squared to obtain the energy spectrum |G(k)| at each frequency point. 2 .

[0151] ②Feature extraction: Divide the entire frequency band from 0.1 to 2000 Hz into multiple feature sub-bands, and calculate the energy integral value of each sub-band within each 10 ms time window.

[0152] ③ Pattern Recognition: The energy integral values ​​of these sub-frequency bands are combined into a feature vector, which is used as input to the oscillation pattern recognition model. The oscillation pattern recognition model automatically identifies and outputs the classification result of the oscillation pattern based on the energy distribution pattern of the input vector. The oscillation pattern recognition model here is obtained by training the classification model using a dataset containing energy integral feature vectors and corresponding oscillation patterns. The specific classification model can be a support vector machine (SVM), or other classification models obtained from existing technologies.

[0153] 3) Locating the oscillation source: Obtaining the phase difference of multiple nodes through a wide-area synchronous measurement unit (PMU). (Obtained from the output of the "Hamming window convolution module"), using the formula The signal arrival time difference Δt is calculated, and the location of the oscillation source is matched using the power grid GIS topology database. The core of this step lies in utilizing the physical connections of the power grid to uniquely determine one or a few most probable actual electrical nodes from the theoretically possible multiple oscillation source locations. The decision-making process is as follows: Figure 6 As shown, the details are as follows:

[0154] ① Time difference positioning and initial candidate set generation: using the formula Calculate the time difference of arrival (TWA) between all pairs of PMUs, where Δt_ij represents the TWA between nodes i and j. This represents the phase difference between nodes i and j. Based on the time difference data, in the electrical geometry model, each pair of PMUs can determine a hyperbola. The intersection of multiple hyperbolas is the theoretical location of the oscillation source, and these intersections constitute an "initial candidate source set". This step essentially uses the time difference positioning principle to initially determine the oscillation source.

[0155] ② Topological Constraints and Exact Matching: The above "initial candidate source set" is compared with the GIS topology database, which contains the exact coordinates and electrical connections of all substations, power plants, load nodes, and the transmission lines connecting them. Specifically:

[0156] Node priority matching principle (prioritize judging whether the source point meets the node priority matching condition): traverse the initial candidate source set and check whether there is a source point that happens to fall on a certain network node in the GIS database (such as a power plant bus or hub substation), or falls within its allowable error range (±500 meters). If such a node exists, the node is directly determined as the oscillation source.

[0157] Line projection matching principle (further determining whether a source point meets the line projection matching condition): For candidate source points that are not close to any node, check whether they fall within the buffer zone of a transmission line (a strip area consisting of 100 meters on both sides of the line). If they fall on a certain line, calculate the precise location of the oscillation source on that line (30% of the distance from the beginning of the line) through vertical projection. This corresponds to a fault or special phenomenon at a certain point in the middle of the line.

[0158] Nearest neighbor matching and conflict resolution principle: For candidate points that are neither on nodes nor on lines, a conflict resolution strategy is initiated. The minimum electrical distance from the candidate point to all topology elements (nodes and lines) is calculated, and the candidate point is assigned to the nearest topology element. Simultaneously, the system records a result with low confidence, which may require manual verification.

[0159] ③ Confidence Assessment and Output: The system assigns a confidence score to each location result. These results include those obtained from the node priority matching principle, line projection matching principle, and nearest neighbor matching and conflict resolution principle in step ②. The score is based on confidence factors such as the consistency of time difference measurements, the degree of fit between the source point and topological elements, and the dispersion among multiple candidate points. Based on the confidence scores, further rating and processing are performed. Finally, the most reliable oscillation source location is output, as shown in the following example format:

[0160] [Node ID: XXX Substation 110kV Busbar]

[0161] [Line ID: YYY-ZZZ, 45% distance from YYY terminal]

[0162] The specific calculation methods for the above three confidence factors are as follows:

[0163] (a) Consistency of time difference measurements measures the degree of consistency when calculating time difference data from different PMU nodes and then inferring the location of the oscillation source. The calculation method is as follows:

[0164] Assuming there are M PMUs, K valid node pairs will be formed. Calculate the Δt of the K time differences. k .

[0165] Using all K time difference data points, an optimal estimated source point coordinate is calculated through a time difference localization algorithm. .

[0166] Calculate the residual: for each time difference Δt k Based on its theoretical propagation model and estimated source point A theoretical time difference can be calculated. .

[0167] Consistency index CTDOA It is calculated using the normalized root mean square (NRMSE) of the residuals and mapped to the confidence score:

[0168]

[0169]

[0170] In the above formula, the smaller the NRMSE, the more consistent the time difference data. TDOA The closer it is to 1.

[0171] (b) The degree of fit between the source point and the topology element, quantifying the physical proximity of the candidate source point to the grid topology element it matches. The calculation method is as follows:

[0172] For node matching: calculate the Euclidean distance d from the candidate source point to the node coordinates, and then calculate the degree of fit according to the following formula. : R node It is the allowable error range for node priority matching.

[0173] For line projection matching: calculate the distance d from the candidate source point to its perpendicular projection point on the line, and then calculate the degree of fit according to the following formula. R line It is the radius of the line buffer.

[0174] (c) The dispersion among multiple candidate points: In the initial candidate source set, if the locations of multiple candidate points are highly concentrated, it indicates that the localization result is stable and reliable; if they are highly dispersed, it indicates that the result has high uncertainty. The calculation method is as follows:

[0175] Suppose the initial candidate source set contains N points. Calculate the mean center O of these points, and calculate the average distance from all candidate points to the mean center O. Then, the dispersion index is calculated based on this formula. : , among them It is a normalized constant, when the average distance much smaller hour When the average distance is close to 1 When it is very large, Close to 0.

[0176] 4) Risk Level Determination: A multi-feature weighted assessment model is used to classify the risk level of the identified oscillation events. Specifically:

[0177] First, extract four key features of the identified oscillation event: amplitude, frequency band, duration, and rate of change. Then, combine preset weighting coefficients to calculate the comprehensive risk value through weighted summation. Finally, based on the threshold range of the comprehensive risk value, the oscillation event is classified into three risk levels: "low," "medium," and "high," and differentiated alarm and control strategies are triggered for different levels.

[0178] The specific risk level can be adjusted according to the actual situation.

[0179] 5) Real-time alarm output: Triggers 4-20mA analog output, Modbus-TCP network messages and local LCD display, supports COMTRADE format fault recording (records 5 seconds of data before and after the event).

[0180] The 4-20mA analog output here is typically used to characterize the “severity” or “risk level” of the oscillation, and is a standard alarm signal widely recognized by industrial control systems.

[0181] The Modbus-TCP network message here contains structured and detailed oscillation diagnostic information, including oscillation flags, oscillation characteristic values, risk levels, oscillation modes, source location information, timestamps, etc.; the message receiver is the local / remote monitoring master station.

[0182] The fault recording here is a record of the raw sampled data (analog channel) and related switch quantities for 5 seconds before and after the event trigger point in standard COMTRADE format.

[0183] Thus, the broadband oscillation detection method of the present invention has been realized.

[0184] It should be noted that in this embodiment, the hybrid-based FFT module and the Hanning window convolution module are implemented using an FPGA, while the oscillation decision module is implemented using an ARM. This heterogeneous structure is a technical solution that comprehensively considers real-time performance, energy efficiency, development efficiency, system flexibility, and cost. Of course, the hybrid-based FFT module and the Hanning window convolution module can also be implemented using other hardware accelerators with parallel computing capabilities, and the oscillation decision module can also be implemented using other microprocessors, central processing units, or other general-purpose / dedicated processors.

[0185] An implementation method for a broadband oscillation detection system:

[0186] This invention discloses a wideband oscillation detection system, comprising a processor and a memory. The processor and memory communicate and exchange data via an internal bus, and the memory stores computer programs / instructions. The processor provides computational, signal processing, and control capabilities, and executes the computer program / instructions to implement the wideband oscillation detection method of this invention. Specifically, the processor can be a general-purpose processor, such as a digital signal processor (DSP). Alternatively, it can be a high-speed random access memory (RAM) processor.

[0187] This method employs a five-stage pipeline architecture: signal acquisition, preprocessing, mixed-basis FFT, Hanning window convolution, and oscillation decision. This achieves a closed-loop processing of broadband oscillating signals, encompassing acquisition, transformation, analysis, and decision. The basic implementation idea and process are as follows:

[0188] 1) The high-voltage input signal is acquired by the signal acquisition module and output to the synchronous ADC of the preprocessing module.

[0189] The specific signal processing procedure of the signal acquisition module is as follows: the high-voltage input signal first undergoes overvoltage clamping through the RVS protection circuit, and then enters the differential input buffer for impedance transformation; the buffered signal is sent to the instrumentation amplifier for differential amplification and common-mode rejection; the amplified signal undergoes dynamic gain adjustment through the programmable gain amplifier; subsequently, it passes through an anti-aliasing filter to remove high-frequency noise; finally, it is buffered by the ADC driver and output to the synchronous ADC in the preprocessing module. The entire circuit adopts a symmetrical layout and shielding design to reduce electromagnetic interference and ensure wide bandwidth and high common-mode rejection ratio characteristics.

[0190] 2) The preprocessing module converts the signal output from the signal acquisition module into a digital signal stream output.

[0191] The preprocessing module includes a synchronous ADC, which can convert analog signals into 16-bit digital signal streams and achieve seamless data connection through a ping-pong buffering mechanism to avoid spectral aliasing caused by sampling intervals.

[0192] 3) The signal output in step 2) is processed by a mixed-base FFT processing module to convert the time-domain signal into a frequency-domain signal to obtain the original spectrum signal.

[0193] The hybrid-radix FFT processing module features specific settings tailored to its application scenarios in areas such as dynamic multi-radix factorization rules and series compression, twitch factor hierarchical optimization and storage compression, reconfigurable butterfly operation units, and parallel acceleration. These include:

[0194] ① Mixed-basis FFT preferentially decomposes the number of sampling points into 4a ×8 b ×3 c ×5 d The combination of a, b, c, d ≥ 0; and for cases that cannot be factored into 2, 3, and 5 factors, the pre-generated optimized small prime number DFT kernel is called to process the remaining factors if they are small prime numbers; the remaining factors are processed by padding with zeros to the nearest neighbor number of decomposable points.

[0195] ② The radix-2, radix-4, and radix-8 twitch factors in the hybrid radix FFT are fixed as constants in the hardware logic of the butterfly operation unit; the radix-3 and radix-5 twitch factors in the hybrid radix FFT are pre-stored in memory and read through a lookup table; the inter-level twitch factors in the hybrid radix FFT calculation are all pre-calculated and stored in memory and read through a lookup table; the non-standard twitch factors in the hybrid radix FFT are calculated and generated in real time using the CORDIC algorithm.

[0196] ③ The radix-8 in the hybrid radix FFT is decomposed into two levels of radix-4 operations. The two levels of radix-4 operations share the radix-4 butterfly unit and use a timing scheduling strategy to generate an 8-point FFT output.

[0197] ④ When resources are limited, the hybrid radix FFT switches radix 3 and radix 5 to radix 2 mode and schedules them to idle periods to perform radix 3 and radix 5 operations using radix 2 operation units.

[0198] 4) The original frequency domain signal obtained in step 3) is convolved in the frequency domain using the Hanning window convolution module, and the denoised spectrum is obtained based on the frequency domain convolution result.

[0199] This step suppresses spectral leakage and improves the signal-to-noise ratio (SNR) through frequency domain convolution. A parametric Hanning window design is employed, dynamically selecting the optimal window function configuration based on the real-time SNR. When the SNR is less than a set SNR threshold, the Hanning window uses a wide-window mode; otherwise, it uses a narrow-window mode. The frequency domain broadening factor in the wide-window mode is less than that in the narrow-window mode.

[0200] Furthermore, the phase of the frequency domain convolution result is reverse-compensated, that is, the compensated phase is the original phase of the frequency domain convolution result minus the phase shift, and the phase shift is the phase of the Hanning window spectrum; the amplitude of the frequency domain convolution result is normalized by the window energy correction coefficient, and the normalized amplitude is the ratio of the original amplitude of the frequency domain convolution result to the window energy correction coefficient.

[0201] 5) Input the output result of step 4) into the oscillation decision module for oscillation decision.

[0202] This step involves: identifying the presence of oscillation events, identifying oscillation patterns, locating the oscillation source, determining the risk level, and outputting real-time alarms.

[0203] Among them, oscillation mode recognition is achieved based on artificial intelligence technology; the phase difference of multiple nodes is obtained through a wide-area synchronous measurement unit (PMU), and the arrival time difference is calculated. The location of the oscillation source is matched by combining the power grid GIS topology database.

[0204] For more detailed information, please refer to the detailed description of the method in "An Implementation Method of a Wideband Oscillation Detection Method".

[0205] One embodiment of a computer-readable storage medium:

[0206] The present invention provides a computer-readable storage medium storing computer-executable instructions / programs, which, when invoked and executed by a processor, cause the processor to implement a broadband oscillation detection method of the present invention.

[0207] This method employs a five-stage pipeline architecture: signal acquisition, preprocessing, mixed-basis FFT, Hanning window convolution, and oscillation decision. This achieves a closed-loop processing of broadband oscillating signals, encompassing acquisition, transformation, analysis, and decision. The basic implementation idea and process are as follows:

[0208] 1) The high-voltage input signal is acquired by the signal acquisition module and output to the synchronous ADC of the preprocessing module.

[0209] The specific signal processing procedure of the signal acquisition module is as follows: the high-voltage input signal first undergoes overvoltage clamping through the RVS protection circuit, and then enters the differential input buffer for impedance transformation; the buffered signal is sent to the instrumentation amplifier for differential amplification and common-mode rejection; the amplified signal undergoes dynamic gain adjustment through the programmable gain amplifier; subsequently, it passes through an anti-aliasing filter to remove high-frequency noise; finally, it is buffered by the ADC driver and output to the synchronous ADC in the preprocessing module. The entire circuit adopts a symmetrical layout and shielding design to reduce electromagnetic interference and ensure wide bandwidth and high common-mode rejection ratio characteristics.

[0210] 2) The preprocessing module converts the signal output from the signal acquisition module into a digital signal stream output.

[0211] The preprocessing module includes a synchronous ADC, which can convert analog signals into 16-bit digital signal streams and achieve seamless data connection through a ping-pong buffering mechanism to avoid spectral aliasing caused by sampling intervals.

[0212] 3) The signal output in step 2) is processed by a mixed-base FFT processing module to convert the time-domain signal into a frequency-domain signal to obtain the original spectrum signal.

[0213] The hybrid-radix FFT processing module features specific settings tailored to its application scenarios in areas such as dynamic multi-radix factorization rules and series compression, twitch factor hierarchical optimization and storage compression, reconfigurable butterfly operation units, and parallel acceleration. These include:

[0214] ① Mixed-basis FFT preferentially decomposes the number of sampling points into 4 a ×8 b ×3 c ×5 d The combination of a, b, c, d ≥ 0; and for cases that cannot be factored into 2, 3, and 5 factors, the pre-generated optimized small prime number DFT kernel is called to process the remaining factors if they are small prime numbers; the remaining factors are processed by padding with zeros to the nearest neighbor number of decomposable points.

[0215] ② The radix-2, radix-4, and radix-8 twitch factors in the hybrid radix FFT are fixed as constants in the hardware logic of the butterfly operation unit; the radix-3 and radix-5 twitch factors in the hybrid radix FFT are pre-stored in memory and read through a lookup table; the inter-level twitch factors in the hybrid radix FFT calculation are all pre-calculated and stored in memory and read through a lookup table; the non-standard twitch factors in the hybrid radix FFT are calculated and generated in real time using the CORDIC algorithm.

[0216] ③ The radix-8 in the hybrid radix FFT is decomposed into two levels of radix-4 operations. The two levels of radix-4 operations share the radix-4 butterfly unit and use a timing scheduling strategy to generate an 8-point FFT output.

[0217] ④ When resources are limited, the hybrid radix FFT switches radix 3 and radix 5 to radix 2 mode and schedules them to idle periods to perform radix 3 and radix 5 operations using radix 2 operation units.

[0218] 4) The original frequency domain signal obtained in step 3) is convolved in the frequency domain using the Hanning window convolution module, and the denoised spectrum is obtained based on the frequency domain convolution result.

[0219] This step suppresses spectral leakage and improves the signal-to-noise ratio (SNR) through frequency domain convolution. A parametric Hanning window design is employed, dynamically selecting the optimal window function configuration based on the real-time SNR. When the SNR is less than a set SNR threshold, the Hanning window uses a wide-window mode; otherwise, it uses a narrow-window mode. The frequency domain broadening factor in the wide-window mode is less than that in the narrow-window mode.

[0220] Furthermore, the phase of the frequency domain convolution result is reverse-compensated, that is, the compensated phase is the original phase of the frequency domain convolution result minus the phase shift, and the phase shift is the phase of the Hanning window spectrum; the amplitude of the frequency domain convolution result is normalized by the window energy correction coefficient, and the normalized amplitude is the ratio of the original amplitude of the frequency domain convolution result to the window energy correction coefficient.

[0221] 5) Input the output result of step 4) into the oscillation decision module for oscillation decision.

[0222] This step involves: identifying the presence of oscillation events, identifying oscillation patterns, locating the oscillation source, determining the risk level, and outputting real-time alarms.

[0223] Among them, oscillation mode recognition is achieved based on artificial intelligence technology; the phase difference of multiple nodes is obtained through a wide-area synchronous measurement unit (PMU), and the arrival time difference is calculated. The location of the oscillation source is matched by combining the power grid GIS topology database.

[0224] For more detailed information, please refer to the detailed description of the method in "An Implementation Method of a Wideband Oscillation Detection Method".

[0225] In summary, this invention addresses the problems of spectrum leakage, noise interference, and insufficient real-time performance in the detection of wide-frequency oscillations (0.1-2000Hz) in high-penetration power grids of new energy sources. It proposes a hybrid-based FFT dynamic decomposition algorithm and a parameterized Hanning window frequency domain convolution collaborative processing method. By optimizing the number of operation stages through multi-level decomposition and suppressing noise through parallel frequency domain convolution, combined with dynamic threshold decision and phase difference positioning technology, high-precision real-time detection of wide-frequency oscillations can be achieved.

Claims

1. A broadband oscillation detection method, characterized in that, The method includes: 1) Acquire the monitoring signal, and perform time-frequency conversion on the monitoring signal using a hybrid-based FFT to obtain the original spectrum signal; the hybrid-based FFT is used for: ① Decompose the number of sampling points into 4 a ×8 b ×3 c ×5 d The combination of a, b, c, d ≥ 0; then the remaining factors are decomposed into 2 factors; ② For the remaining factors that cannot be processed after step ①, if there is an optimized DFT kernel, the pre-generated optimized DFT kernel is called for processing; if there is no optimized DFT kernel, zeros are padded to the nearest neighbor number of points that can be completely decomposed by the strategy in step ①. 2) The original spectral signal is convolved in the frequency domain with the spectrum of the Hanning window function, which has spectral leakage suppression capabilities. The denoised spectrum is obtained from the frequency domain convolution result as follows: The phase of the frequency domain convolution result is inversely compensated, i.e., the compensated phase is the original phase of the frequency domain convolution result minus the phase shift, where the phase shift is the phase of the Hanning window spectrum; the amplitude of the frequency domain convolution result is normalized using a window energy correction coefficient, where the normalized amplitude is the ratio of the original amplitude of the frequency domain convolution result to the window energy correction coefficient; the window energy correction coefficient K is: w(n) is the time-domain expression of the Hanning window, and L is the total number of points of the Hanning window w(n); 3) Utilize the denoised spectrum for broadband oscillation detection: identify the presence of oscillation events, identify oscillation modes, locate oscillation sources, determine the risk level of oscillation events, and output real-time alarms.

2. The broadband oscillation detection method according to claim 1, characterized in that, In the hybrid radix FFT, the radix-2, radix-4, and radix-8 twitch factors are fixed as constants in the hardware logic of the butterfly operation unit; the radix-3 and radix-5 twitch factors are pre-stored in memory and retrieved via lookup tables; the inter-level twitch factors in the hybrid radix FFT calculation are all pre-calculated and stored in memory and retrieved via lookup tables; the non-standard twitch factors in the hybrid radix FFT are generated in real time using the CORDIC algorithm.

3. The broadband oscillation detection method according to claim 1, characterized in that, In the hybrid radix FFT, the radix-8 is decomposed into two levels of radix-4 operations. The two levels of radix-4 operations share a radix-4 butterfly unit, and the following timing scheduling strategy is used to generate an 8-point FFT output: The original 8-point input sequence is split into two 4-point subsequences, group 0 and group 1 respectively; During clock cycles T1~T4, enable the radix 4 butterfly cell, send group 0 data into the radix 4 butterfly cell for radix 4 operation, and obtain the 4-point intermediate result Y0; During clock cycles T5~T8, the data of group 1 is sent to the radix-4 butterfly unit for radix-4 operation to obtain 4 intermediate results Y1; the intermediate result Y0 is read and multiplied by the pre-stored rotation factor in an independent multiplier array to obtain Y0'; During clock cycles T9~T12, the results Y0' and Y1 are fed into the radix-4 butterfly unit as inputs to add Y0' and Y1, thereby generating an 8-point FFT output; The T1, T2, ..., T11, T12 mentioned above are 12 clock cycles that progress according to time.

4. The broadband oscillation detection method according to claim 1, characterized in that, When resources are limited, the hybrid radix FFT switches radix 3 and radix 5 to radix 2 mode and schedules them to idle periods to perform radix 3 and radix 5 operations using radix 2 operation units.

5. The broadband oscillation detection method according to claim 1, characterized in that, The frequency domain coefficients of the Hanning window are selected based on the signal-to-noise ratio (SNR) of the monitored signal. When the SNR of the monitored signal is less than the set SNR threshold, the Hanning window is set to wide window mode; otherwise, the Hanning window is set to narrow window mode. The frequency domain broadening coefficient in the wide window mode is less than that in the narrow window mode.

6. The broadband oscillation detection method according to claim 1, characterized in that, Methods for identifying the presence of oscillations include: determining the amplitude of each frequency point based on the denoised spectrum; if the amplitude of a certain frequency point continuously exceeds a set amplitude threshold, an oscillation event is considered to have occurred.

7. The broadband oscillation detection method according to claim 1, characterized in that, Oscillation modes include at least subsynchronous, supersynchronous, and harmonic resonance; and the methods for identifying oscillation modes include: The amplitude of the denoised spectrum is extracted, and then the extracted amplitude is squared to obtain the energy spectrum of each frequency point. The entire frequency band is divided into multiple sub-bands, and the energy integral value of each sub-band is calculated within a set time window. The energy integral values ​​of each sub-frequency band are combined into an energy integral feature vector, which is then input into the oscillation pattern recognition model to obtain the oscillation pattern recognition result. The oscillation pattern recognition model is obtained by training a classification model using a dataset containing the energy integral feature vector and the corresponding oscillation pattern.

8. The broadband oscillation detection method according to claim 1, characterized in that, Methods for locating oscillation sources include: The phase difference of the noise-reduced spectrum corresponding to the monitoring signals of each pair of source points in the power grid is obtained by the PMU, and the arrival time difference of the signal is calculated based on the phase difference. Based on the time difference data, the oscillation source is initially determined by combining the electrical geometric model of the power grid and the time difference positioning principle, and the initially determined oscillation source is formed into an initial candidate source set. Traverse the initial candidate source set and determine whether there are any source points that meet the node priority matching condition. The node priority matching condition refers to the source point falling into a certain network node in the power grid topology GIS database or within the allowable error range centered on the network node. Source points that meet the node priority matching condition are determined to be oscillation sources. Then, for source points that do not meet the node priority matching condition, it is determined whether there are source points that meet the line projection matching condition. The line projection matching condition refers to the source point falling in the buffer zone of a transmission line. For source points that meet the line projection matching condition, the precise location of the oscillation source on the line is calculated by vertical projection. Furthermore, for source points that do not meet the line projection matching conditions, the minimum electrical distance from the source point to all topological elements in the power grid topology is calculated, and the source point is assigned to the topological element with the smallest distance; the topological elements include nodes and lines. Then, the confidence level of all the above positioning results is calculated, and the oscillation source is determined based on the confidence level.

9. The broadband oscillation detection method according to claim 8, characterized in that, The confidence factors used to calculate the confidence level include consistency based on time difference measurement, the degree of fit between the source point and the topology elements, and the dispersion among multiple candidate points. Consistency based on time difference measurement measures the degree of consistency when calculating the time difference data through different PMU nodes and then inferring the location of the oscillation source. The degree of fit between the source point and the topology element is a quantitative measure of how close the candidate source point is to the grid topology element it matches in physical space.

10. The broadband oscillation detection method according to claim 1, characterized in that, The methods for determining the risk level of identified oscillation events include: extracting four key features of the identified oscillations: amplitude, frequency band, duration, and rate of change; weighting and summing the four key features to obtain a comprehensive risk value; judging the risk level of the oscillation event based on the range in which the comprehensive risk value is located, and triggering different alarm strategies for oscillation events of different risk levels.

11. The broadband oscillation detection method according to any one of claims 1 to 10, characterized in that, The monitoring signal is the signal obtained by sequentially passing the original acquired signal through the RVS protection circuit, differential input buffer, instrumentation amplifier, programmable gain amplifier, anti-aliasing filter, ADC driver and synchronous ADC; The RVS protection circuit is used for overvoltage clamping; the differential input buffer is used for impedance transformation; the instrumentation amplifier is used for differential amplification and common-mode rejection; the programmable gain amplifier is used for dynamic gain adjustment; the anti-aliasing filter is used to filter out high-frequency noise; the ADC driver is used to condition the input signal to meet the input requirements of the synchronous ADC; the synchronous ADC is used to convert the signal into a digital signal.

12. The broadband oscillation detection method according to claim 1, characterized in that, The monitoring signal is a voltage or current signal collected from the primary side of the power grid that reflects the wideband dynamic characteristics of the system.

13. A broadband oscillation detection system, comprising a processor, characterized in that, The processor is used to execute computer program instructions stored in memory to implement the steps of the following method: 1) Acquire the monitoring signal, and perform time-frequency conversion on the monitoring signal using a hybrid-based FFT to obtain the original spectrum signal; the hybrid-based FFT is used for: ① Decompose the number of sampling points into 4 a ×8 b ×3 c ×5 d The combination of a, b, c, d ≥ 0; then the remaining factors are decomposed into 2 factors; ② For the remaining factors that cannot be processed after step ①, if there is an optimized DFT kernel, the pre-generated optimized DFT kernel is called for processing; if there is no optimized DFT kernel, zeros are padded to the nearest neighbor number of points that can be completely decomposed by the strategy in step ①. 2) Convolve the original spectral signal with the spectrum of the Hanning window function with spectral leakage suppression in the frequency domain, and obtain the denoised spectrum from the frequency domain convolution result in the following way: perform inverse compensation on the phase of the frequency domain convolution result, that is, the compensated phase is the original phase of the frequency domain convolution result minus the phase shift, and the phase shift is the phase of the Hanning window spectrum. The amplitude of the frequency domain convolution result is normalized by the window energy correction coefficient. The normalized amplitude is the ratio of the original amplitude of the frequency domain convolution result to the window energy correction coefficient. The window energy correction factor K is: w(n) is the time-domain expression of the Hanning window, and L is the total number of points of the Hanning window w(n); 3) Utilize the denoised spectrum for broadband oscillation detection: identify the presence of oscillation events, identify oscillation modes, locate oscillation sources, determine the risk level of oscillation events, and output real-time alarms.

14. The broadband oscillation detection system according to claim 13, characterized in that, In the hybrid radix FFT, the radix-2, radix-4, and radix-8 twitch factors are fixed as constants in the hardware logic of the butterfly operation unit; the radix-3 and radix-5 twitch factors are pre-stored in memory and retrieved via lookup tables; the inter-level twitch factors in the hybrid radix FFT calculation are all pre-calculated and stored in memory and retrieved via lookup tables; the non-standard twitch factors in the hybrid radix FFT are generated in real time using the CORDIC algorithm.

15. The broadband oscillation detection system according to claim 13, characterized in that, In the hybrid radix FFT, the radix-8 is decomposed into two levels of radix-4 operations. The two levels of radix-4 operations share a radix-4 butterfly unit, and the following timing scheduling strategy is used to generate an 8-point FFT output: The original 8-point input sequence is split into two 4-point subsequences, group 0 and group 1 respectively; During clock cycles T1~T4, enable the radix 4 butterfly cell, send group 0 data into the radix 4 butterfly cell for radix 4 operation, and obtain the 4-point intermediate result Y0; During clock cycles T5~T8, the data of group 1 is sent to the radix-4 butterfly unit for radix-4 operation to obtain 4 intermediate results Y1; the intermediate result Y0 is read and multiplied by the pre-stored rotation factor in an independent multiplier array to obtain Y0'; During clock cycles T9~T12, the results Y0' and Y1 are fed into the radix-4 butterfly unit as inputs to add Y0' and Y1, thereby generating an 8-point FFT output; The T1, T2, ..., T11, T12 mentioned above are 12 clock cycles that progress according to time.

16. The broadband oscillation detection system according to claim 13, characterized in that, When resources are limited, the hybrid radix FFT switches radix 3 and radix 5 to radix 2 mode and schedules them to idle periods to perform radix 3 and radix 5 operations using radix 2 operation units.

17. The broadband oscillation detection system according to claim 13, characterized in that, The frequency domain coefficients of the Hanning window are selected based on the signal-to-noise ratio (SNR) of the monitored signal. When the SNR of the monitored signal is less than the set SNR threshold, the Hanning window is set to wide window mode; otherwise, the Hanning window is set to narrow window mode. The frequency domain broadening coefficient in the wide window mode is less than that in the narrow window mode.

18. The broadband oscillation detection system according to claim 13, characterized in that, Methods for identifying the presence of oscillations include: determining the amplitude of each frequency point based on the denoised spectrum; if the amplitude of a certain frequency point continuously exceeds a set amplitude threshold, an oscillation event is considered to have occurred.

19. The broadband oscillation detection system according to claim 13, characterized in that, Oscillation modes include at least subsynchronous, supersynchronous, and harmonic resonance; and the methods for identifying oscillation modes include: The amplitude of the denoised spectrum is extracted, and then the extracted amplitude is squared to obtain the energy spectrum of each frequency point. The entire frequency band is divided into multiple sub-bands, and the energy integral value of each sub-band is calculated within a set time window. The energy integral values ​​of each sub-frequency band are combined into an energy integral feature vector, which is then input into the oscillation pattern recognition model to obtain the oscillation pattern recognition result. The oscillation pattern recognition model is obtained by training a classification model using a dataset containing the energy integral feature vector and the corresponding oscillation pattern.

20. The broadband oscillation detection system according to claim 13, characterized in that, Methods for locating oscillation sources include: The phase difference of the noise-reduced spectrum corresponding to the monitoring signals of each pair of source points in the power grid is obtained by the PMU, and the arrival time difference of the signal is calculated based on the phase difference. Based on the time difference data, the oscillation source is initially determined by combining the electrical geometric model of the power grid and the time difference positioning principle, and the initially determined oscillation source is formed into an initial candidate source set. Traverse the initial candidate source set and determine whether there are any source points that meet the node priority matching condition. The node priority matching condition refers to the source point falling into a certain network node in the power grid topology GIS database or within the allowable error range centered on the network node. Source points that meet the node priority matching condition are determined to be oscillation sources. Then, for source points that do not meet the node priority matching condition, it is determined whether there are source points that meet the line projection matching condition. The line projection matching condition refers to the source point falling in the buffer zone of a transmission line. For source points that meet the line projection matching condition, the precise location of the oscillation source on the line is calculated by vertical projection. Furthermore, for source points that do not meet the line projection matching conditions, the minimum electrical distance from the source point to all topological elements in the power grid topology is calculated, and the source point is assigned to the topological element with the smallest distance; the topological elements include nodes and lines. Then, the confidence level of all the above positioning results is calculated, and the oscillation source is determined based on the confidence level.

21. The broadband oscillation detection system according to claim 20, characterized in that, The confidence factors used to calculate the confidence level include consistency based on time difference measurement, the degree of fit between the source point and the topology elements, and the dispersion among multiple candidate points. Consistency based on time difference measurement measures the degree of consistency when calculating the time difference data through different PMU nodes and then inferring the location of the oscillation source. The degree of fit between the source point and the topology element is a quantitative measure of how close the candidate source point is to the grid topology element it matches in physical space.

22. The broadband oscillation detection system according to claim 13, characterized in that, The methods for determining the risk level of identified oscillation events include: extracting four key features of the identified oscillations: amplitude, frequency band, duration, and rate of change; weighting and summing the four key features to obtain a comprehensive risk value; judging the risk level of the oscillation event based on the range in which the comprehensive risk value is located, and triggering different alarm strategies for oscillation events of different risk levels.

23. The broadband oscillation detection system according to any one of claims 13 to 22, characterized in that, The monitoring signal is the signal obtained by sequentially passing the original acquired signal through the RVS protection circuit, differential input buffer, instrumentation amplifier, programmable gain amplifier, anti-aliasing filter, ADC driver and synchronous ADC; The RVS protection circuit is used for overvoltage clamping; the differential input buffer is used for impedance transformation; the instrumentation amplifier is used for differential amplification and common-mode rejection; the programmable gain amplifier is used for dynamic gain adjustment; the anti-aliasing filter is used to filter out high-frequency noise; the ADC driver is used to condition the input signal to meet the input requirements of the synchronous ADC; the synchronous ADC is used to convert the signal into a digital signal.

24. The broadband oscillation detection system according to claim 13, characterized in that, The monitoring signal is a voltage or current signal collected from the primary side of the power grid that reflects the wideband dynamic characteristics of the system.

25. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, this computer program implements the following steps for a wideband oscillation detection method: 1) Acquire the monitoring signal, and perform time-frequency conversion on the monitoring signal using a hybrid-based FFT to obtain the original spectrum signal; the hybrid-based FFT is used for: ① Decompose the number of sampling points into 4 a ×8 b ×3 c ×5 d The combination of a, b, c, d ≥ 0; then the remaining factors are decomposed into 2 factors; ② For the remaining factors that cannot be processed after step ①, if there is an optimized DFT kernel, the pre-generated optimized DFT kernel is called for processing; if there is no optimized DFT kernel, zeros are padded to the nearest neighbor number of points that can be completely decomposed by the strategy in step ①. 2) Convolve the original spectral signal with the spectrum of the Hanning window function with spectral leakage suppression in the frequency domain, and obtain the denoised spectrum from the frequency domain convolution result in the following way: perform inverse compensation on the phase of the frequency domain convolution result, that is, the compensated phase is the original phase of the frequency domain convolution result minus the phase shift, and the phase shift is the phase of the Hanning window spectrum. The amplitude of the frequency domain convolution result is normalized by the window energy correction coefficient. The normalized amplitude is the ratio of the original amplitude of the frequency domain convolution result to the window energy correction coefficient. The window energy correction factor K is: w(n) is the time-domain expression of the Hanning window, and L is the total number of points of the Hanning window w(n); 3) Utilize the denoised spectrum for broadband oscillation detection: identify the presence of oscillation events, identify oscillation modes, locate oscillation sources, determine the risk level of oscillation events, and output real-time alarms.

26. The computer-readable storage medium according to claim 25, characterized in that, In the hybrid radix FFT, the radix-2, radix-4, and radix-8 twitch factors are fixed as constants in the hardware logic of the butterfly operation unit; the radix-3 and radix-5 twitch factors are pre-stored in memory and retrieved via lookup tables; the inter-level twitch factors in the hybrid radix FFT calculation are all pre-calculated and stored in memory and retrieved via lookup tables; the non-standard twitch factors in the hybrid radix FFT are generated in real time using the CORDIC algorithm.

27. The computer-readable storage medium according to claim 25, characterized in that, In the hybrid radix FFT, the radix-8 is decomposed into two levels of radix-4 operations. The two levels of radix-4 operations share a radix-4 butterfly unit, and the following timing scheduling strategy is used to generate an 8-point FFT output: The original 8-point input sequence is split into two 4-point subsequences, group 0 and group 1 respectively; During clock cycles T1~T4, enable the radix 4 butterfly cell, send group 0 data into the radix 4 butterfly cell for radix 4 operation, and obtain the 4-point intermediate result Y0; During clock cycles T5~T8, the data of group 1 is sent to the radix-4 butterfly unit for radix-4 operation to obtain 4 intermediate results Y1; the intermediate result Y0 is read and multiplied by the pre-stored rotation factor in an independent multiplier array to obtain Y0'; During clock cycles T9~T12, the results Y0' and Y1 are fed into the radix-4 butterfly unit as inputs to add Y0' and Y1, thereby generating an 8-point FFT output; The T1, T2, ..., T11, T12 mentioned above are 12 clock cycles that progress according to time.

28. The computer-readable storage medium according to claim 25, characterized in that, When resources are limited, the hybrid radix FFT switches radix 3 and radix 5 to radix 2 mode and schedules them to idle periods to perform radix 3 and radix 5 operations using radix 2 operation units.

29. The computer-readable storage medium according to claim 25, characterized in that, The frequency domain coefficients of the Hanning window are selected based on the signal-to-noise ratio (SNR) of the monitored signal. When the SNR of the monitored signal is less than the set SNR threshold, the Hanning window is set to wide window mode; otherwise, the Hanning window is set to narrow window mode. The frequency domain broadening coefficient in the wide window mode is less than that in the narrow window mode.

30. The computer-readable storage medium according to claim 25, characterized in that, Methods for identifying the presence of oscillations include: determining the amplitude of each frequency point based on the denoised spectrum; if the amplitude of a certain frequency point continuously exceeds a set amplitude threshold, an oscillation event is considered to have occurred.

31. The computer-readable storage medium according to claim 25, characterized in that, Oscillation modes include at least subsynchronous, supersynchronous, and harmonic resonance; and the methods for identifying oscillation modes include: The amplitude of the denoised spectrum is extracted, and then the extracted amplitude is squared to obtain the energy spectrum of each frequency point. The entire frequency band is divided into multiple sub-bands, and the energy integral value of each sub-band is calculated within a set time window. The energy integral values ​​of each sub-frequency band are combined into an energy integral feature vector, which is then input into the oscillation pattern recognition model to obtain the oscillation pattern recognition result. The oscillation pattern recognition model is obtained by training a classification model using a dataset containing the energy integral feature vector and the corresponding oscillation pattern.

32. The computer-readable storage medium according to claim 25, characterized in that, Methods for locating oscillation sources include: The phase difference of the noise-reduced spectrum corresponding to the monitoring signals of each pair of source points in the power grid is obtained by the PMU, and the arrival time difference of the signal is calculated based on the phase difference. Based on the time difference data, the oscillation source is initially determined by combining the electrical geometric model of the power grid and the time difference positioning principle, and the initially determined oscillation source is formed into an initial candidate source set. Traverse the initial candidate source set and determine whether there are any source points that meet the node priority matching condition. The node priority matching condition refers to the source point falling into a certain network node in the power grid topology GIS database or within the allowable error range centered on the network node. Source points that meet the node priority matching condition are determined to be oscillation sources. Then, for source points that do not meet the node priority matching condition, it is determined whether there are source points that meet the line projection matching condition. The line projection matching condition refers to the source point falling in the buffer zone of a transmission line. For source points that meet the line projection matching condition, the precise location of the oscillation source on the line is calculated by vertical projection. Furthermore, for source points that do not meet the line projection matching conditions, the minimum electrical distance from the source point to all topological elements in the power grid topology is calculated, and the source point is assigned to the topological element with the smallest distance; the topological elements include nodes and lines. Then, the confidence level of all the above positioning results is calculated, and the oscillation source is determined based on the confidence level.

33. The computer-readable storage medium according to claim 32, characterized in that, The confidence factors used to calculate the confidence level include consistency based on time difference measurement, the degree of fit between the source point and the topology elements, and the dispersion among multiple candidate points. Consistency based on time difference measurement measures the degree of consistency when calculating the time difference data through different PMU nodes and then inferring the location of the oscillation source. The degree of fit between the source point and the topology element is a quantitative measure of how close the candidate source point is to the grid topology element it matches in physical space.

34. The computer-readable storage medium according to claim 25, characterized in that, The methods for determining the risk level of identified oscillation events include: extracting four key features of the identified oscillations: amplitude, frequency band, duration, and rate of change; weighting and summing the four key features to obtain a comprehensive risk value; judging the risk level of the oscillation event based on the range in which the comprehensive risk value is located, and triggering different alarm strategies for oscillation events of different risk levels.

35. The computer-readable storage medium according to any one of claims 25 to 34, characterized in that, The monitoring signal is the signal obtained by sequentially passing the original acquired signal through the RVS protection circuit, differential input buffer, instrumentation amplifier, programmable gain amplifier, anti-aliasing filter, ADC driver and synchronous ADC; The RVS protection circuit is used for overvoltage clamping; the differential input buffer is used for impedance transformation; the instrumentation amplifier is used for differential amplification and common-mode rejection; the programmable gain amplifier is used for dynamic gain adjustment; the anti-aliasing filter is used to filter out high-frequency noise; the ADC driver is used to condition the input signal to meet the input requirements of the synchronous ADC; the synchronous ADC is used to convert the signal into a digital signal.

36. The computer-readable storage medium according to claim 25, characterized in that, The monitoring signal is a voltage or current signal collected from the primary side of the power grid that reflects the wideband dynamic characteristics of the system.

Images