Multi-band oscillation monitoring method, system, and medium for power systems

Abstract

The application discloses a kind of multi-band oscillation monitoring method, system and medium for power system, belong to power system oscillation monitoring field, method includes: the active power measurement is carried out fast Fourier transform, to extract the first characteristic vector including main harmonic amplitude, main harmonic frequency and main harmonic amplitude ratio in;First characteristic vector is input first stage Gaussian naive Bayes classifier and is classified, and output oscillation type;When oscillation type is low-frequency oscillation, the active power measurement is carried out synchronous compression wavelet transform and hilbert transform, to extract the second characteristic vector including harmonic component frequency, harmonic component attenuation factor;Second characteristic vector is input second stage Gaussian naive Bayes classifier and is classified, and output the oscillation state of low-frequency oscillation.Can accurately judge low-frequency oscillation type, and judge system stability, solved the problem that each type of oscillation sample is uneven and leads to low training efficiency and low accuracy of classifier.

Description

Technical Field

[0001] This invention belongs to the field of power system oscillation monitoring, and more specifically, relates to a method, system and medium for multi-band oscillation monitoring of power systems. Background Technology

[0002] With the introduction of high proportions of renewable energy and power electronic equipment, new energy power systems are experiencing increasingly complex multi-frequency oscillations. Real-time and accurate online monitoring of multi-frequency oscillations in power systems, and timely acquisition of information such as oscillation frequency and amplitude, can provide data support for subsequent oscillation localization and suppression, effectively improving the stability and safety of power system operation.

[0003] Existing multi-band oscillation monitoring methods are all based on spectrum analysis, comparing the obtained parameters with set thresholds to determine whether the system is oscillating and providing information on the frequency and amplitude of the oscillation. However, current online monitoring systems suffer from problems such as inability to identify attenuation factors and slow alarm times when identifying low-frequency oscillations. Furthermore, oscillation monitoring alarm thresholds based on data measured by a phasor measurement unit (PMU) rely on human experience, further exacerbating the problems of insufficient monitoring speed and accuracy. Compared with traditional multi-band oscillation analysis methods, artificial intelligence methods have many advantages; however, in PMU data, there are a large number of normal and non-oscillating samples, while the number of abnormal and oscillating samples is small, resulting in uneven sample distribution and insufficient detection accuracy. Summary of the Invention

[0004] To address the shortcomings and improvement needs of existing technologies, this invention provides a multi-band oscillation monitoring method, system, and medium for power systems. Its purpose is to solve the problems of slow alarm time, inability to identify specific parameters of low-frequency oscillations, and reliance on human experience for alarm thresholds in existing oscillation monitoring methods.

[0005] To achieve the above objectives, according to one aspect of the present invention, a method for monitoring multi-band oscillations in a power system is provided, comprising: S1, performing a fast Fourier transform on the active power measurement value of the power system to extract a first feature vector including the main harmonic amplitude, the main harmonic frequency, and the proportion of the main harmonic amplitude; S2, inputting the first feature vector into a first-stage Gaussian Naive Bayes classifier for classification and outputting the oscillation type; S3, when the oscillation type is low-frequency oscillation, performing a synchronous compressed wavelet transform and a Hilbert transform on the active power measurement value sequentially to extract a second feature vector including the harmonic component frequency and the harmonic component attenuation factor; S4, inputting the second feature vector into a second-stage Gaussian Naive Bayes classifier for classification and outputting the oscillation state of the low-frequency oscillation.

[0006] Furthermore, in S1, the frequency spectrum and amplitude spectrum of the active power measurement value are obtained after the fast Fourier transform. The main harmonic amplitude is the maximum amplitude value in the amplitude spectrum, the main harmonic frequency is the frequency corresponding to the maximum amplitude value in the frequency spectrum, and the proportion of the main harmonic amplitude is the ratio between the maximum amplitude value and the total amplitude value of all amplitudes in the amplitude spectrum.

[0007] Furthermore, S3 includes: sequentially performing synchronous compressed wavelet transform and inverse transform on the active power measurement value to extract and reconstruct harmonic signal components; performing Hilbert transform on the harmonic signal components, and calculating the instantaneous amplitude and instantaneous frequency of the harmonic signal components based on the result of the Hilbert transform; calculating the harmonic component frequency based on the instantaneous frequency of the harmonic signal components, and calculating the harmonic component attenuation factor based on the instantaneous amplitude of the harmonic signal components.

[0008] Furthermore, the harmonic component frequencies and harmonic component attenuation factors are respectively:

[0009]

[0010]

[0011] Where ω is the frequency of the harmonic component, α is the harmonic component attenuation factor, and ω d Let A(t1) be the instantaneous frequency of the harmonic signal component, and A(t2) be the instantaneous amplitude of the harmonic signal component at time t1 and time t2, respectively.

[0012] Furthermore, in S2, when the main harmonic frequency is within the low-frequency oscillation frequency range, and the main harmonic amplitude is higher than the amplitude threshold and the proportion of the main harmonic amplitude is higher than the proportion threshold, the output oscillation type is low-frequency oscillation.

[0013] Furthermore, in step S4, the damping ratio is calculated based on the harmonic component frequency and the harmonic component attenuation factor; when the damping ratio is greater than a first threshold and the harmonic component attenuation factor is less than a second threshold, the low-frequency oscillation is a reduced-amplitude oscillation; when the damping ratio is between the first and third thresholds and the harmonic component attenuation factor is between the second and fourth thresholds, the low-frequency oscillation is a constant-amplitude oscillation; when the damping ratio is less than the third threshold and the harmonic component attenuation factor is greater than the fourth threshold, the low-frequency oscillation is an increased-amplitude oscillation; wherein, the first threshold is greater than the third threshold, and the second threshold is less than the fourth threshold.

[0014] Furthermore, before S1, the method further includes: optimizing the training of a classifier using a sample set, wherein the classifier is either the first-stage Gaussian Naive Bayes classifier or the second-stage Gaussian Naive Bayes classifier; during the optimization training process, the classifier is monitored using a ten-fold cross-validation method, and when the cross-validation accuracy is found to be lower than the accuracy threshold, the sample size is increased to optimize the training of the classifier again until the cross-validation is passed.

[0015] Furthermore, when the active power measurement value is not in the sample set, the active power measurement value and its corresponding oscillation type and oscillation state are added to the sample set to update the classifier.

[0016] According to another aspect of the present invention, a multi-band oscillation monitoring system for a power system is provided, comprising: a first extraction module for performing a fast Fourier transform on the active power measurement value of the power system to extract a first feature vector including the main harmonic amplitude, the main harmonic frequency, and the proportion of the main harmonic amplitude; a first classification module for inputting the first feature vector into a first-stage Gaussian Naive Bayes classifier for classification and outputting the oscillation type; a second extraction module for performing a synchronous compressed wavelet transform and a Hilbert transform on the active power measurement value sequentially when the oscillation type is low-frequency oscillation, to extract a second feature vector including the harmonic component frequency and the harmonic component attenuation factor; and a second classification module for inputting the second feature vector into a second-stage Gaussian Naive Bayes classifier for classification and outputting the oscillation state of the low-frequency oscillation.

[0017] According to another aspect of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, characterized in that, when the program is executed by a processor, it implements the multi-band oscillation monitoring method for power systems as described above.

[0018] In summary, the above-described technical solutions of this invention achieve the following beneficial effects: A multi-band oscillation monitoring method for power systems is provided. By introducing the Gaussian Naive Bayes algorithm from machine learning, the problem of slow alarm response and low accuracy caused by manually setting alarm thresholds in online monitoring systems is avoided. Simultaneously, by introducing parameter extraction methods using synchronous compressed wavelet transform and Hilbert transform, parameters such as the frequency and attenuation factor of low-frequency oscillations can be extracted with high accuracy, thereby accurately classifying low-frequency oscillation samples and accurately determining system stability. Furthermore, by designing a two-stage classifier, the low training efficiency and accuracy of the classifier caused by the uneven distribution of oscillation samples of different types in actual situations are improved, thus enhancing monitoring performance. Attached Figure Description

[0019] Figure 1A flowchart of a multi-band oscillation monitoring method for power systems provided in an embodiment of the present invention;

[0020] Figure 2 A schematic diagram of the two-stage Gaussian Naive Bayes classifier monitoring process provided in this embodiment of the invention;

[0021] Figure 3 This is a wiring diagram of a simulation model of a four-machine, two-region new energy system provided in an embodiment of the present invention;

[0022] Figure 4A , Figure 4B , Figure 4C This is a schematic diagram of the amplitude spectrum obtained in the first classification stage under different active power measurement waveforms provided in the embodiments of the present invention.

[0023] Figure 5A , Figure 5B These are time-frequency curves and waveform diagrams obtained by synchronous compressed wavelet transform according to embodiments of the present invention.

[0024] Figure 6A , Figure 6B These are schematic diagrams of instantaneous amplitude and instantaneous frequency obtained by Hilbert transform according to embodiments of the present invention;

[0025] Figure 7 A schematic diagram showing the comparison between the reconstructed signal and the original signal in the second classification stage provided in an embodiment of the present invention;

[0026] Figure 8 This is a block diagram of a multi-band oscillation monitoring system for power systems provided in an embodiment of the present invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0028] In this invention, the terms "first," "second," etc. (if present) in the invention and the accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0029] Example 1

[0030] Figure 1 A flowchart illustrating a multi-band oscillation monitoring method for power systems provided in an embodiment of the present invention. (See also...) Figure 1 , combined Figures 2-7The multi-band oscillation monitoring method for power systems in this embodiment is described in detail. The method includes operations S1-S4.

[0031] In this embodiment, a two-stage Gaussian Naive Bayes classifier is pre-constructed, comprising two Gaussian Naive Bayes classifiers. This two-stage Gaussian Naive Bayes classifier needs to be trained before executing operation S1.

[0032] Specifically, the first-stage Gaussian Naive Bayes classifier is optimized and trained using a sample set. During the optimization training process, a 10-fold cross-validation method is used to monitor the first-stage Gaussian Naive Bayes classifier. When the cross-validation accuracy is found to be lower than the accuracy threshold, the sample size is increased to further optimize the classifier training until cross-validation is passed. The sample set used to train the first-stage Gaussian Naive Bayes classifier is obtained as follows: based on the active power data from the existing Wide Area Measurement System (WAMS) and phasor measurement unit, or the active power data simulated by a simulation model, three types of oscillation samples—low-frequency oscillation, subsynchronous oscillation, and no oscillation—are selected and labeled accordingly to form the training sample set for the first-stage classifier. The synchronous measurement sampling frequency of the phasor measurement unit is set to, for example, 100Hz. The active power signal measured by the phasor measurement unit is analyzed every 5 seconds, meaning each sample data contains 500 sampling points. The analysis object is the active power signal of each generator, which balances a relatively small computational load with timely alarms.

[0033] The optimization training process for the second-stage Gaussian Naive Bayes classifier is the same as that for the first-stage Gaussian Naive Bayes classifier. The sample set used to train the second-stage Gaussian Naive Bayes classifier is obtained as follows: based on the active power data from existing wide-area measurement systems and phasor measurement units, or active power data simulated by simulation models, oscillation samples in three states—low-frequency oscillation with decreasing amplitude, constant amplitude oscillation, and increasing amplitude oscillation—are selected and labeled accordingly to form the training sample set for the second-stage classifier.

[0034] According to an embodiment of the present invention, during the monitoring phase after commissioning, when the acquired active power measurement value is not in the sample set, the active power measurement value and its corresponding oscillation type and oscillation state are added to the sample set to update the classifier. This enables the classifier to improve itself, enhancing its adaptability and monitoring accuracy. Furthermore, it allows multi-band oscillation monitoring systems in different regions to share the sample set, increasing the number of oscillation samples, improving system fault tolerance, and mitigating uneven sample data.

[0035] Operation S1 performs a fast Fourier transform on the active power measurement of the power system to extract the first feature vector, which includes the amplitude of the main harmonic, the frequency of the main harmonic, and the proportion of the amplitude of the main harmonic.

[0036] The active power measurement value is subjected to a Fast Fourier Transform (FFT) to obtain its frequency spectrum and amplitude spectrum, both of which have a length of N. The amplitude of the dominant harmonic is A. m The maximum amplitude value in the amplitude spectrum, and the dominant harmonic frequency F. m P represents the frequency corresponding to the maximum amplitude value in the frequency spectrum, and the proportion of the main harmonic amplitude. m This is the ratio between the maximum amplitude value and the total amplitude value of all amplitudes in the amplitude spectrum. The resulting first eigenvector, Feature1, is:

[0037]

[0038] Where Ai is the amplitude of the i-th harmonic in the amplitude spectrum, and fm is the frequency corresponding to the maximum amplitude value in the frequency spectrum.

[0039] Operation S2 inputs the first feature vector into the first-stage Gaussian Naive Bayes classifier for classification and outputs the oscillation type.

[0040] The classification criteria for the first-stage Gaussian Naive Bayes classifier are as follows: when the main harmonic frequency is within the low-frequency oscillation frequency range, and the main harmonic amplitude is higher than the amplitude threshold and the proportion of the main harmonic amplitude is higher than the proportion threshold, the output oscillation type is low-frequency oscillation; when the main harmonic frequency is within the subsynchronous oscillation frequency range, and the main harmonic amplitude is higher than the amplitude threshold and the proportion of the main harmonic amplitude is higher than the proportion threshold, the output oscillation type is subsynchronous oscillation; when the main harmonic amplitude is lower than the amplitude threshold or the proportion of the main harmonic amplitude is lower than the proportion threshold, the output oscillation type is no oscillation.

[0041] Operation S3: When the oscillation type is low-frequency oscillation, the active power measurement value is subjected to synchronous compressed wavelet transform and Hilbert transform in sequence to extract the second feature vector containing the harmonic component frequency and harmonic component attenuation factor.

[0042] According to an embodiment of the present invention, operation S3 includes sub-operations S31-S33.

[0043] In sub-operation S31, the active power measurement value is subjected to synchronous compressed wavelet transform and inverse transform in sequence to extract and reconstruct harmonic signal components.

[0044] Suppose the expression for the active power measurement s(t) to be analyzed is:

[0045]

[0046] Where s(t) is the kth harmonic component sk The eigenfunctions of s(t). A continuous wavelet transform of s(t) is shown in the following equation:

[0047]

[0048] Where ψ is a properly chosen wavelet, a is the scaling factor, and b is the translation factor.

[0049] Then, the wavelet transform result W s (a, b) are compressed into the time-frequency surface to obtain the result T of the synchronous compressed wavelet transform. s (ω l b). Assume the data length of s(t) is n = 2. L+1 If the sampling time interval is defined as Δt, then the sampling frequency is... According to the Nyquist sampling theorem, the sampling frequency is At that time, the maximum signal frequency that can be monitored is Therefore, the frequency range of the harmonic signal components can be determined as follows: To divide this frequency range, the center frequency set is defined as... To determine the specific value of each center frequency, Based on the center frequency ω l It is 2 of the baseline value ω0 lΔω Times, where Δω is the frequency interval between adjacent center frequencies, thus through The specific value of the calculated frequency interval is

[0050] In summary, based on the center frequency set as The frequency range of the harmonic signal components is related to the frequency interval Δω. It can be divided into different frequency narrowbands Compress the harmonic signal to the center frequency ω. l Nearby, the results of the synchronous compressed wavelet transform are obtained, as shown below:

[0051]

[0052] By performing an inverse transform on the synchronous compressed wavelet transform, each harmonic signal component s is realized. k Extraction and reconstruction of (t), namely:

[0053]

[0054] Among them, L k (t m ) represents the k-th harmonic component s k The center frequency ω around the corresponding instantaneous frequency curve l The subscript set.

[0055] In suboperation S32, a Hilbert transform is performed on the harmonic signal components, and the instantaneous amplitude and instantaneous frequency of the harmonic signal components are calculated based on the result of the Hilbert transform.

[0056] The instantaneous amplitude and frequency of each harmonic are calculated using the Hilbert transform. (Signal s) k The Hilbert transform of (t) is:

[0057]

[0058] The result of the Hilbert transform, y(t), is compared with the function s. k By exchanging the positions of (t), we can obtain the inverse Hilbert transform, as shown in the following equation:

[0059] Z(t) = s k (t)+jy(t)=A k (t)e jθ(t)

[0060] Among them, signal s k The instantaneous amplitude A(t) is:

[0061]

[0062] signal s k The time-varying phase θ(t) of (t) is:

[0063]

[0064] Harmonic signal s k The instantaneous frequency ω of (t) d for:

[0065]

[0066] In sub-operation S33, the frequency of the harmonic component is calculated based on the instantaneous frequency of the harmonic signal component, and the attenuation factor of the harmonic component is calculated based on the instantaneous amplitude of the harmonic signal component.

[0067] According to an embodiment of the present invention, the harmonic component frequency and the harmonic component attenuation factor are respectively:

[0068]

[0069]

[0070] Where ω is the frequency of the harmonic component, α is the harmonic component attenuation factor, and ω d Let A(t1) be the instantaneous frequency of the harmonic signal component, and A(t2) be the instantaneous amplitude of the harmonic signal component at time t1 and time t2, respectively.

[0071] The resulting second feature vector, Feature2, is:

[0072]

[0073] Operation S4 inputs the second feature vector into the second-stage Gaussian Naive Bayes classifier for classification, and outputs the oscillation state of low-frequency oscillation.

[0074] According to an embodiment of the present invention, in operation S4, the damping ratio ε is calculated based on the harmonic component frequency and the harmonic component attenuation factor:

[0075]

[0076] The classification criteria for the second-stage Gaussian Naive Bayes classifier are as follows: when the damping ratio is greater than the first threshold and the harmonic component attenuation factor is less than the second threshold, the low-frequency oscillation is a decreasing amplitude oscillation; when the damping ratio is between the first and third thresholds and the harmonic component attenuation factor is between the second and fourth thresholds, the low-frequency oscillation is a constant amplitude oscillation; when the damping ratio is less than the third threshold and the harmonic component attenuation factor is greater than the fourth threshold, the low-frequency oscillation is an increasing amplitude oscillation; wherein, the first threshold is greater than the third threshold and the second threshold is less than the fourth threshold.

[0077] In this embodiment, the phasor measurement unit for monitoring generator active power is adjusted and improved to filter out most noise interference, ensuring that the measured oscillation data is within the measurement frequency band. Simultaneously, to ensure timely and accurate oscillation alarms and reduce the data analysis burden on the WAMS master station, the PMU substation is required to have the function of analyzing and identifying different types of multi-frequency oscillations before uploading this information to the WAMS master station.

[0078] Example 2

[0079] The following describes the multi-band oscillation monitoring method for power systems according to the present invention, with reference to specific embodiments. (See also...) Figure 2 The method includes the following steps.

[0080] Step 1: Based on the active power measured by the actual PMU or the artificially generated oscillation sample data of the simulated power system, establish training sets for the two Gaussian Naive Bayes classifiers respectively.

[0081] See Figure 3 The diagram shows the wiring diagram of a four-generator, two-region wind power coupling system built in MATLAB / Simulink according to actual conditions. Based on this simulation model, different operating conditions were set by changing parameters such as generator output, interference type, interference frequency, noise, and load level, and sample data were collected.

[0082] The system's oscillation-free condition under different operating conditions was simulated by adjusting the generator's active power, load level, and noise level (range 20–40 dB, in 5 dB increments). Different operating conditions were simulated by adding step disturbances or sinusoidal disturbances of different frequencies and amplitudes to the doubly-fed induction generator (DFIG) rotor control side, and by changing the DFIG's phase-locked loop (PLL) parameters. Similarly, considering different operating conditions, low-frequency oscillation sample data was collected by adding step disturbances or sinusoidal disturbances of different frequencies and amplitudes to the generator exciter input or prime mover torque, or by simulating short-circuit faults followed by reclosing on the power exchange transmission line between the two regions.

[0083] The data in the first-stage sample set were processed using Fast Fourier Transform (FFT). The amplitude spectrum curves obtained after FFT for the three types of sample data are shown below. Figure 4A , Figure 4B and Figure 4C As shown. Figures 4A-4C This indicates that the three types of data differ significantly in three dimensions: main harmonic amplitude, main harmonic frequency, and main harmonic amplitude percentage. These three features are used as feature vectors to train the first-stage classifier.

[0084] The data in the second-stage sample set were processed using synchronous compressed wavelet transform and Hilbert transform to extract the harmonic component frequencies and harmonic component attenuation factors as feature vectors, which were then used to train the second-stage Gaussian Naive Bayes classifier.

[0085] Specifically, the accuracy of the parameter extraction method based on synchronous compressed wavelet transform and Hilbert transform is tested using a set of simulated low-frequency oscillation signals y(t). The low-frequency oscillation signal y(t) is:

[0086] y(t) = 15·e -0.3t cosπt+5 cos 4πt

[0087] To more closely approximate real-world conditions and make the simulation results more reliable, a sampling frequency of 100Hz was used, and the signal simulation time was set to 5 seconds. Simultaneously, 20dB of Gaussian white noise was added to the simulated signal to simulate the sampling noise of the PMU. Performing a synchronous compressed wavelet transform on the original signal yields a refined time-frequency curve, such as... Figure 5A and Figure 5B As shown. Figure 5A There are two curves, with frequencies around 0.5Hz and 2Hz, consistent with the theory. Figure 5B It can be qualitatively seen that IMT1 is a decaying waveform, while IMT2 is a constant amplitude oscillation waveform.

[0088] After separating the harmonic signal using synchronous compressed wavelet transform, the instantaneous frequency and instantaneous amplitude of each harmonic component are obtained using Hilbert transform. Figure 6A , Figure 6B The instantaneous amplitude and instantaneous frequency of IMT1 and IMT2 are respectively. The frequency and attenuation factor of IMT1 are 0.4912 and -0.3044, respectively, and the frequency and attenuation factor of IMT2 are 1.9978 and -0.0082, respectively. The frequency and attenuation factor of the two harmonic components are very close to the theoretical values.

[0089] To test the performance of the classifier, a 10-fold cross-validation method was used to test the sample data. The first-stage Gaussian Naive Bayes classifier was used to learn the first-stage sample set. Through 10-fold cross-validation, the accuracy of the first-stage Gaussian Naive Bayes classifier reached 98.67%. Similarly, the accuracy of the second-stage Gaussian Naive Bayes classifier reached 98.46%, indicating that both stages of the classifier have good classification performance.

[0090] The two-stage Gaussian Naive Bayes classifier requires 73.7483 seconds to train and has a classification accuracy of 97.54%, which is significantly better than other machine learning methods, such as the K-Nearest Neighbors algorithm, which has a training time of 332.6302 seconds and a classification accuracy of 78.75%.

[0091] Step 2: After processing the data monitored by the PMU substation through Fast Fourier Transform, feature vectors are extracted. The first-stage Gaussian Naive Bayes classifier is used for preliminary identification and corresponding alarm signals are issued. If the sample is judged to be a low-frequency oscillation signal, synchronous compressed wavelet transform and Hilbert transform are used to extract feature vectors, which are then input into the second-stage Gaussian Naive Bayes classifier to determine the specific low-frequency oscillation type and system stability.

[0092] Specifically, five typical sample types were used to simulate and test the classifier: no oscillation, subsynchronous oscillation, reduced-amplitude low-frequency oscillation, increased-amplitude low-frequency oscillation, and constant-amplitude low-frequency oscillation. For no-oscillation samples, 25 dB of Gaussian white noise was added to the PMU sampling data to simulate the actual situation. For synchronous oscillation samples, the parameters of the doubly-fed induction generator (DFIG) phase-locked loop were adjusted, and a small step interference signal was added to the rotor control side; similarly, 20 dB of Gaussian white noise was added. For low-frequency oscillation samples, after adjusting the system power flow, a small step interference signal was added to the generator excitation input side, adding 20 dB to 30 dB of Gaussian white noise to the data.

[0093] These five sets of sample signals were analyzed and identified by a two-stage Gaussian Naive Bayes classifier, which achieved a classification accuracy of 100%.

[0094] To further examine the parameter recognition performance of the two-stage Gaussian Naive Bayes classifier, the low-frequency oscillation signal extracted and reconstructed through synchronous compressed wavelet transform was compared with the original signal. Taking the amplified low-frequency oscillation signal as an example... Figure 7 As shown, the reconstructed amplified low-frequency oscillation signal fits the original signal well, proving the accuracy of the synchronous compressed wavelet transform.

[0095] Step 3 involves incorporating the five newly generated sample data sets into the Gaussian Naive Bayes classifier sample set. After cross-validation, the two-stage Gaussian Naive Bayes classifier is updated, achieving self-updating of the classifier. Once the monitoring system is in use, this process is continuously iterated to achieve self-improvement of the classifier, thereby enhancing its adaptability and detection accuracy.

[0096] Example 3

[0097] Figure 8 This is a block diagram of a multi-band oscillation monitoring system for power systems provided in an embodiment of the present invention. (See also...) Figure 8 The multi-band oscillation monitoring system 800 for power systems includes a first extraction module 810, a first classification module 820, a second extraction module 830, and a second classification module 840.

[0098] The first extraction module 810, for example, performs operation S1 to perform a fast Fourier transform on the active power measurement value of the power system to extract a first feature vector including the main harmonic amplitude, the main harmonic frequency and the proportion of the main harmonic amplitude.

[0099] The first classification module 820, for example, performs operation S2, which inputs the first feature vector into the first-stage Gaussian Naive Bayes classifier for classification and outputs the oscillation type.

[0100] The second extraction module 830, for example, performs operation S3 to sequentially perform synchronous compressed wavelet transform and Hilbert transform on the active power measurement value when the oscillation type is low-frequency oscillation, in order to extract a second feature vector containing harmonic component frequency and harmonic component attenuation factor.

[0101] The second classification module 840, for example, performs operation S4 to input the second feature vector into the second-stage Gaussian Naive Bayes classifier for classification and output the oscillation state of low-frequency oscillation.

[0102] A multi-band oscillation monitoring system 800 for power systems is used to perform the above... Figures 1-7 The illustrated embodiment presents a multi-band oscillation monitoring method for power systems. For details not covered in this embodiment, please refer to the foregoing. Figures 1-7 The multi-band oscillation monitoring method for power systems shown in the embodiments will not be described in detail here.

[0103] This invention also provides a computer-readable storage medium having a computer program stored thereon. When executed by a processor, the program implements as follows: Figures 1-7 The embodiment shown illustrates a multi-band oscillation monitoring method for power systems.

[0104] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for monitoring multi-band oscillations in power systems, characterized in that, include: S1, perform a fast Fourier transform on the active power measurement of the power system to extract the first feature vector containing the main harmonic amplitude, main harmonic frequency and the proportion of the main harmonic amplitude; S2, input the first feature vector into the first-stage Gaussian Naive Bayes classifier for classification, and output the oscillation type; S3, when the oscillation type is low-frequency oscillation, the active power measurement value is subjected to synchronous compressed wavelet transform and Hilbert transform in sequence to extract the second feature vector including harmonic component frequency and harmonic component attenuation factor. S4, input the second feature vector into the second-stage Gaussian Naive Bayes classifier for classification, and output the oscillation state of the low-frequency oscillation.

2. The multi-band oscillation monitoring method for power systems as described in claim 1, characterized in that, In S1, the frequency spectrum and amplitude spectrum of the active power measurement value are obtained after the fast Fourier transform. The main harmonic amplitude is the maximum amplitude value in the amplitude spectrum, the main harmonic frequency is the frequency corresponding to the maximum amplitude value in the frequency spectrum, and the proportion of the main harmonic amplitude is the ratio between the maximum amplitude value and the total amplitude value of all amplitudes in the amplitude spectrum.

3. The multi-band oscillation monitoring method for power systems as described in claim 1, characterized in that, S3 includes: The active power measurement values ​​are sequentially subjected to synchronous compressed wavelet transform and inverse transform to extract and reconstruct harmonic signal components; Perform a Hilbert transform on the harmonic signal components, and calculate the instantaneous amplitude and instantaneous frequency of the harmonic signal components based on the result of the Hilbert transform; The harmonic component frequency is calculated based on the instantaneous frequency of the harmonic signal component, and the harmonic component attenuation factor is calculated based on the instantaneous amplitude of the harmonic signal component.

4. The multi-band oscillation monitoring method for power systems as described in claim 3, characterized in that, The harmonic component frequencies and harmonic component attenuation factors are respectively: Where ω is the frequency of the harmonic component, α is the harmonic component attenuation factor, and ω d Let A(t1) be the instantaneous frequency of the harmonic signal component, and A(t2) be the instantaneous amplitude of the harmonic signal component at time t1 and time t2, respectively.

5. The multi-band oscillation monitoring method for power systems as described in claim 1, characterized in that, In S2, when the main harmonic frequency is within the low-frequency oscillation frequency range, and the main harmonic amplitude is higher than the amplitude threshold and the proportion of the main harmonic amplitude is higher than the proportion threshold, the output oscillation type is low-frequency oscillation.

6. The multi-band oscillation monitoring method for power systems as described in claim 1, characterized in that, In S4, the damping ratio is calculated based on the harmonic component frequency and the harmonic component attenuation factor. When the damping ratio is greater than the first threshold and the harmonic component attenuation factor is less than the second threshold, the low-frequency oscillation is a reduced-amplitude oscillation state. When the damping ratio is between the first threshold and the third threshold, and the harmonic component attenuation factor is between the second threshold and the fourth threshold, the low-frequency oscillation is a constant-amplitude oscillation. When the damping ratio is less than the third threshold and the harmonic component attenuation factor is greater than the fourth threshold, the low-frequency oscillation is an amplified oscillation state; wherein, the first threshold is greater than the third threshold and the second threshold is less than the fourth threshold.

7. The multi-band oscillation monitoring method for power systems as described in any one of claims 1-6, characterized in that, Before S1, the following also applies: The classifier is optimized and trained using a sample set, wherein the classifier is either the first-stage Gaussian Naive Bayes classifier or the second-stage Gaussian Naive Bayes classifier. During the optimization training process, the classifier is monitored using the 10-fold cross-validation method. When the cross-validation accuracy is found to be lower than the accuracy threshold, the sample size is increased to further optimize the training of the classifier until it passes cross-validation.

8. The multi-band oscillation monitoring method for power systems as described in claim 7, characterized in that, When the active power measurement value is not in the sample set, the active power measurement value and its corresponding oscillation type and oscillation state are added to the sample set to update the classifier.

9. A multi-band oscillation monitoring system for power systems, characterized in that, include: The first extraction module is used to perform a fast Fourier transform on the active power measurement value of the power system to extract a first feature vector containing the main harmonic amplitude, the main harmonic frequency and the proportion of the main harmonic amplitude. The first classification module is used to input the first feature vector into the first-stage Gaussian Naive Bayes classifier for classification and output the oscillation type. The second extraction module is used to perform synchronous compressed wavelet transform and Hilbert transform on the active power measurement value in sequence when the oscillation type is low frequency oscillation, so as to extract a second feature vector including harmonic component frequency and harmonic component attenuation factor. The second classification module is used to input the second feature vector into the second-stage Gaussian Naive Bayes classifier for classification and output the oscillation state of the low-frequency oscillation.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the multi-band oscillation monitoring method for power systems as described in any one of claims 1-8.

Images