A method for online identification of power transient disturbance

Abstract

The present invention discloses an online identification method for power transient disturbance, comprising the following steps: (1) collecting transient disturbance signals existing in the cable system and performing normalization processing; (2) deconstructing the transient disturbance signals into multiple components with different center frequencies, and extracting a waveform feature matrix therefrom; (3) performing Wigner distribution WVD time-frequency analysis on each component respectively, obtaining a time-frequency spectrum and extracting a time-frequency spectrum feature matrix of each component; (4) fusing the waveform feature matrix with the time-frequency spectrum feature matrix to obtain multiple time-frequency feature matrices, thereby realizing the classification and identification of transient disturbance signals. The transient disturbance signal collected by the high-frequency current sensor HFCT installed at the cable accessory proposed by the present invention contains more obvious feature details, is more conducive to realizing the classification and identification of transient disturbance signals, and fully extends the utilization value of the cable HFCT.

Description

Technical Field

[0001] The present invention belongs to the technical field of power disturbance, and in particular relates to an online identification method for power transient disturbance. Background Art

[0002] With the rapid development of smart grids, a large number of power electronic equipment, intelligent control devices and high-power nonlinear loads are continuously put into use, making various power quality disturbances in the power system normalized and complicated. Transient disturbances are a type of power quality disturbance with short duration, strong impact and high frequency. They can be caused by various switch operations, transient faults, lightning strikes, etc. in the power system. They pose a great threat to a large number of power electronic equipment in modern power systems. It is worth carrying out research on the identification and tracing of transient disturbances to provide a basis for their further quantitative control.

[0003] Existing research mainly focuses on disturbance identification using low and medium sampling rate monitoring equipment installed in system substations. However, the sampling rate is too low to reproduce the signal details, which makes feature extraction difficult and makes it impossible to accurately identify transient disturbance signals. Moreover, for most transient disturbances scattered throughout the system, the attenuation path of the disturbance signal propagating to the monitoring equipment at the head end of the substation is often too long, making it difficult to ensure the effectiveness of signal acquisition.

[0004] In addition, the commonly used disturbance signal feature extraction methods currently include short-time Fourier transform, wavelet transform, S transform and empirical mode decomposition. Since the window length of short-time Fourier transform is fixed, the time resolution and frequency resolution remain unchanged, which is not suitable for the analysis of transient disturbances; wavelet transform cannot meet the requirements of high time resolution and high frequency resolution at the same time and has poor noise resistance; S transform has a large amount of calculation and is not suitable for the analysis of high-frequency transient disturbance signals; empirical mode decomposition can realize adaptive decomposition of signals, but there are problems of modal aliasing and endpoint effects in practical applications; it can be seen that the above methods all have the problems of insufficient accuracy in extracting features of transient disturbance signals and low recognition rate. Summary of the invention

[0005] The purpose of the present invention is to provide an online identification method for power transient disturbance to solve the problems in the prior art of difficulty in extracting transient disturbance signal features, low recognition rate, and difficulty in ensuring the effectiveness of signal acquisition.

[0006] To achieve the above object, the technical solution of the present invention is to provide an online identification method for power transient disturbance, the innovation of which is that it includes the following steps:

[0007] (1) The transient disturbance signal existing in the cable system is collected by the high-frequency current sensor HFCT installed at the cable accessories, and the collected transient disturbance signal is normalized;

[0008] (2) The White Shark optimization algorithm is used to optimize the penalty factor α and the modal number K of the variational modal decomposition. Then, the variational modal decomposition algorithm after parameter optimization is used to deconstruct the transient disturbance signal into multiple components with different center frequencies, from which the waveform feature matrix is ​​extracted;

[0009] (3) Perform Wigner distribution WVD time-frequency analysis on each component to obtain the time-frequency spectrum and extract the time-frequency spectrum feature matrix of each component;

[0010] (4) The waveform feature matrix and the time-frequency spectrum feature matrix are integrated to construct multiple time-frequency feature matrices, which are used as the input of the long short-term memory network to realize the classification and identification of transient disturbance signals.

[0011] Furthermore, in the step (1), the transient disturbance signal existing in the cable system is collected by the high-frequency current sensor HFCT installed at the cable accessory, and the collected transient disturbance signal is normalized in the specific process as follows: first, the transient disturbance signal waveform in the 110KV cable system is generated by simulation, and then the original signal of the disturbance current on the grounding wire of the cable accessory is generated by a signal generator and input into the high-frequency current sensor HFCT, and the output signal of the real high-frequency current sensor HFCT is collected by an oscilloscope to obtain the transient disturbance signal after filtering by the high-frequency current sensor HFCT, and then the collected transient disturbance signal is normalized by the maximum-minimum normalization method.

[0012] Furthermore, in step (2), the White Shark optimization algorithm is used to automatically optimize the two parameters of the penalty factor α and the modal number K of the variational modal decomposition to establish the optimal combination of the penalty factor α and the modal number K, and then the high-frequency current sensor HFCT filtering transient disturbance signal is adaptively and accurately decomposed into several components BIMF1~BIMF K , extract the peak coefficient C of each component respectively 1K , Kurtosis C 2K , Pulse Factor C 3K and margin factor C 4K There are four feature quantities in total, constructing the waveform feature matrix C i :

[0013]

[0014] In the formula, C i is the waveform feature matrix corresponding to the i-th sample, K is the number of components of variational mode decomposition;

[0015] For a transient disturbance signal x(t), its variational mode decomposition process includes two parts: the construction and solution of the variational problem. The model construction of the variational problem is:

[0016]

[0017] Where: s k (t) is the k modal components obtained by variational modal decomposition of x(t); {s k}={s1,s2,…,s k}; {ω k}={ω1,ω2,…,ω k}; δ(t) is the Dirac function; It means to find the partial derivative with respect to t;

[0018] The solution process of the variational problem is:

[0019]

[0020] Where: λ is the Lagrange multiplication operator; α is the quadratic penalty factor;

[0021] In addition, when the White Shark optimization algorithm is used to optimize the parameters of variational mode decomposition, the envelope entropy is used as the fitness function, and the expression is:

[0022]

[0023] Where: i = 1, 2, ..., N; E h is the envelope entropy; h i It is the normalized form of the envelope signal obtained after the original signal is Hilbert transformed.

[0024] Furthermore, the step (3) obtains the Wigner distribution of each component respectively, and extracts the following four time-frequency features for each component based on the Wigner distribution time-frequency spectrum: the mean F of the Wigner distribution 1K , Variance F 2K Energy F 3K The frequency F corresponding to the maximum amplitude 4K , thus obtaining the time-frequency spectrum feature matrix F of the original signal i :

[0025]

[0026] In the formula, F i is the time-frequency spectrum feature matrix corresponding to the i-th sample;

[0027] For a transient disturbance signal x(t), its Wigner distribution is defined as:

[0028]

[0029] Where: τ is the time difference variable; x*(t) is the conjugate complex number of x(t); is the instantaneous correlation function of x(t).

[0030] Furthermore, the multiple time-frequency feature matrix in step (4) is composed of the waveform feature matrix C in step (2) and step (3) i And the time-frequency spectrum feature matrix F i The merged structure is as follows:

[0031]

[0032] Where Y i is the multiple time-frequency feature matrix corresponding to the i-th sample;

[0033] A% of the multiple time-frequency feature matrices of all samples extracted by the above method are used as the training set, and (100-a)% is used as the test set. The training set is then input into the long short-term memory network for training. The features are screened through the three gating structures of the internal forget gate, input gate, and output gate and the gating characteristics of the activation function, and the correlation between various features is sought to obtain the optimal network model. The trained long short-term memory network is then used to learn the test set, and the Softmax classifier is used to classify the test data, and finally the identification result of the transient disturbance signal is output.

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] (1) The transient disturbance signal collected by the high-frequency current sensor HFCT installed at the cable accessories proposed by the present invention contains more obvious characteristic details, which is more conducive to the classification and identification of transient disturbance signals and fully extends the utilization value of cable HFCT;

[0036] (2) In the feature extraction of weak high-frequency transient disturbance signals, the White Shark optimization algorithm is used to optimize the parameters of variational mode decomposition, avoiding the shortcomings of parameter selection based on manual experience, achieving adaptive decomposition and improving decomposition accuracy;

[0037] (3) Performing Wigner distribution time-frequency analysis on each decomposed component separately not only eliminates the influence of cross-interference terms, but also retains the characteristic information of each frequency component to the maximum extent, so that the multiple time-frequency feature matrix constructed thereby has a higher feature discrimination degree;

[0038] (4) The multiple time-frequency feature matrix and identification method proposed in the present invention can fully explore the local details of relatively weak transient disturbance signals, and have the advantages of obvious feature distinction, high identification accuracy and good noise resistance. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 The present invention is a flow chart of a method for online identification of power transient disturbances.

[0040] Figure 2 The present invention is a flowchart of constructing multiple time-frequency characteristic matrices of the high-frequency current sensor HFCT filtering transient disturbance signal.

[0041] Figure 3 This is a flow chart of transient disturbance identification of the long short-term memory network of the present invention.

[0042] Figure 4 The diagram is a result diagram of the high-frequency current sensor HFCT in an embodiment of the present invention before and after sampling a common transient disturbance signal.

[0043] Figure 5 This is a process of extracting multiple time-frequency feature matrices of a capacitor switching disturbance signal in an embodiment of the present invention.

[0044] Figure 6 It is a multiple time-frequency feature matrix extracted from various transient disturbance signals in the embodiment of the present invention. DETAILED DESCRIPTION

[0045] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0046] The present invention provides a method for online identification of power transient disturbances, the flow chart of which is as follows: Figure 1 As shown, the following steps are included:

[0047] (1) The transient disturbance signal existing in the cable system is collected by the high-frequency current sensor HFCT installed at the cable accessories, and the collected transient disturbance signal is normalized; the specific process is: first, the transient disturbance signal waveform in the 110KV cable system is generated by simulation, and then the original signal of the disturbance current on the grounding wire of the cable accessories is generated by a signal generator and input into the high-frequency current sensor HFCT, and the output signal of the real high-frequency current sensor HFCT is collected by an oscilloscope to obtain the transient disturbance signal after filtering by the high-frequency current sensor HFCT, and then the collected transient disturbance signal is normalized by the maximum-minimum normalization method.

[0048] (2) The White Shark optimization algorithm is used to optimize the penalty factor α and the mode number K of the variational modal decomposition. Then, the variational modal decomposition algorithm after parameter optimization is used to deconstruct the transient disturbance signal into multiple components with different center frequencies, from which the waveform feature matrix is ​​extracted.

[0049] The specific process is as follows: the White Shark optimization algorithm is used to automatically optimize the two parameters of the penalty factor α and the modal number K of the variational modal decomposition to establish the optimal combination of the penalty factor α and the modal number K. Then, the high-frequency current sensor HFCT filtering transient disturbance signal is adaptively and accurately decomposed into several components of different frequency components BIMF1~BIMF through the variational modal decomposition after parameter optimization. K , extract the peak coefficient C of each component respectively 1K , Kurtosis C 2K , Pulse Factor C 3K and margin factor C 4K There are four feature quantities in total, constructing the waveform feature matrix C i :

[0050]

[0051] In the formula, C i is the waveform feature matrix corresponding to the i-th sample, K is the number of components of variational mode decomposition;

[0052] For a transient disturbance signal x(t), its variational mode decomposition process includes two parts: the construction and solution of the variational problem. The model construction of the variational problem is:

[0053]

[0054] Where: s k (t) is the k modal components obtained by variational modal decomposition of x(t); {s k}={s1,s2,…,s k}; {ω k}={ω1,ω2,…,ω k}; δ(t) is the Dirac function; It means to find the partial derivative with respect to t;

[0055] The solution process of the variational problem is:

[0056]

[0057] Where: λ is the Lagrange multiplication operator; α is the quadratic penalty factor;

[0058] In addition, when the White Shark optimization algorithm is used to optimize the parameters of variational mode decomposition, the envelope entropy is used as the fitness function, and the expression is:

[0059]

[0060] Where: i = 1, 2, ..., N; E h is the envelope entropy; h iIt is the normalized form of the envelope signal obtained after the original signal is Hilbert transformed.

[0061] (3) Perform Wigner distribution WVD time-frequency analysis on each component to obtain the time-frequency spectrum and extract the time-frequency spectrum feature matrix of each component; specifically: the mean F of the Wigner distribution 1K , variance F 2K Energy F 3K The frequency F corresponding to the maximum amplitude 4K , thus obtaining the time-frequency spectrum feature matrix F of the original signal i :

[0062]

[0063] In the formula, F i is the time-frequency spectrum feature matrix corresponding to the i-th sample;

[0064] For a transient disturbance signal x(t), its Wigner distribution is defined as:

[0065]

[0066] Where: τ is the time difference variable; x*(t) is the conjugate complex number of x(t); is the instantaneous correlation function of x(t).

[0067] (4) The waveform feature matrix and the time-frequency spectrum feature matrix are integrated to construct multiple time-frequency feature matrices, which are used as the input of the long short-term memory network to realize the classification and identification of transient disturbance signals.

[0068] The multiple time-frequency feature matrix is ​​composed of the waveform feature matrix C in step (2) and step (3) i And the time-frequency spectrum feature matrix F i The merged structure is as follows:

[0069]

[0070] Where Y i is the multiple time-frequency feature matrix corresponding to the i-th sample;

[0071] A% of the multiple time-frequency feature matrices of all samples extracted by the above method are used as the training set, and (100-a)% is used as the test set. The training set is then input into the long short-term memory network for training. The features are screened through the three gating structures of the internal forget gate, input gate, and output gate and the gating characteristics of the activation function, and the correlation between various features is sought to obtain the optimal network model. The trained long short-term memory network is then used to learn the test set, and the Softmax classifier is used to classify the test data, and finally the identification result of the transient disturbance signal is output.

[0072] The above content is a specific technical solution of the present invention. Based on the above technical solution, the specific implementation mode of the present invention further describes the above solution through the following content.

[0073] (1) First, MATLAB is used to simulate and generate six types of transient disturbance signal waveforms in the 110KV cable system: high-resistance arc grounding fault, low-resistance metallic grounding fault, and partial discharge; system operation types: capacitor switching, no-load line switching, and load switching. Then, the original signal of the disturbance current on the cable accessory grounding wire is generated by a signal generator and input into the high-frequency current sensor HFCT. The output signal of the real high-frequency current sensor HFCT is collected by an oscilloscope to obtain the transient disturbance signal after filtering by the high-frequency current sensor HFCT. The sampling frequency of the oscilloscope is 100MHz, as shown in Figure 1. Figure 4 The figure shows the results of the high-frequency current sensor HFCT before and after sampling the common transient disturbance signal. Figure 4 It can be seen that the disturbance signal after sampling by the high-frequency current sensor HFCT has filtered out the low-frequency redundant components, and the transient feature details contained are more obvious, which is more conducive to the recognition and analysis of transient disturbances. The parameters of the six disturbance types are changed respectively, and 100 samples are collected for each disturbance type under different working conditions, totaling 600. The parameter settings of various signals are shown in Table 1. The maximum-minimum normalization method is then used to normalize the collected transient disturbance signals.

[0074] Table 1

[0075]

[0076] (2) The White Shark optimization algorithm with strong global optimization capability is used to automatically optimize the two parameters of the variational modal decomposition, namely, the penalty factor α and the modal number K. The population size of the White Shark optimization algorithm is 50, the number of iterations is 50, and the optimization ranges of the parameters α and K are α∈[100,5000] and K∈[1,10], respectively. Thus, the optimal combination of the penalty factor α and the modal number K is established as [3134,5]. Then, the high-frequency current sensor HFCT filter transient disturbance signal is adaptively and accurately decomposed into several components BIMF1~BIMF with different frequency components through the variational modal decomposition after parameter optimization. K , extract the peak coefficient C of each component respectively 1K , Kurtosis C 2K , Pulse Factor C 3K and margin factor C 4K There are four feature quantities in total, constructing the waveform feature matrix C i , taking capacitor switching as an example, its waveform characteristic matrix is:

[0077]

[0078] (3) Perform Wigner distribution WVD time-frequency analysis on each component to obtain the time-frequency spectrum and extract the time-frequency spectrum feature matrix of each component; specifically: the mean F of the Wigner distribution 1K , Variance F 2K Energy F 3K The frequency F corresponding to the maximum amplitude 4K , thus obtaining the time-frequency spectrum feature matrix F of the original signal i , taking capacitor switching as an example, its time-frequency spectrum feature matrix is:

[0079]

[0080] (4) Fusion of the waveform feature matrix C in step (2) and step (3) i And the time-frequency spectrum feature matrix F i Construct a multiple time-frequency feature matrix, and its construction process is as follows Figure 2 As shown, Figure 5 The figure shows the construction process of multiple time-frequency characteristic matrices of capacitor switching disturbance signals. The constructed multiple time-frequency characteristic matrices are:

[0081]

[0082] According to the above steps, the multiple time-frequency feature matrices of six typical transient disturbance signals are extracted as follows: Figure 6 As shown in the figure, 70% of the extracted multiple time-frequency feature matrices of all samples are used as training sets and 30% as test sets. The training sets are then input into the long short-term memory network for training to obtain the best network model. The trained long short-term memory network is then used to learn the test set. The Softmax classifier is used to classify the test data, and finally the identification result of the transient disturbance signal is output. The identification process is as follows: Figure 3 As shown, at the same time, only the waveform feature matrix C is used i , time-frequency spectrum feature matrix F i As the input of the long short-term memory network, the proposed method is trained and tested. The experimental results are shown in Table 2. The overall recognition rate of the proposed method is as high as 99.62%, which is higher than that of the method that only extracts the waveform feature matrix C. i And only extract the time-frequency spectrum feature matrix F i In terms of the recognition rate, it has increased by 15.85% and 9.16% respectively, which can deeply mine the characteristic information of various transient disturbance signals, so as to accurately identify the transient disturbance signals existing in the system.

[0083] Table 2

[0084]

[0085] Taking into account the particularity that the high-frequency current sensor HFCT is generally interfered by noise in actual measurement sampling, five transient disturbance HFCT filter signals containing large noise, namely -10dB, -5dB, 5dB, 10dB and 20dB, are constructed respectively, and the present invention is further tested. The average recognition results after adding noise are shown in Table 3. In a high noise environment with an SNR of -10dB, the overall recognition rate can still reach 96.61%, and with the increase of SNR, the recognition accuracy also increases. When the SNR is 10dB and above, the recognition rate is higher than 99.27%. It can be seen that the present invention has good anti-noise performance and can be applied to the actual application environment of the high-frequency current sensor HFCT.

[0086] Table 3

[0087]

[0088] Through the above-mentioned method, the present invention discloses an online identification method for power transient disturbances, which collects transient disturbance signals through high-frequency current sensors HFCT distributed on the cable accessory grounding wires in the power system, and then performs parameter-optimized variational mode decomposition and Wigner distribution time-frequency analysis on the collected transient disturbance signals, and extracts multiple time-frequency feature matrices to classify and identify the transient disturbance signals. Experimental data show that the high-frequency current sensor HFCT on the cable accessory grounding wire can effectively collect transient disturbance signals in the system, and is an ideal signal source for monitoring system transient disturbances. The multiple time-frequency feature matrices extracted by collecting transient disturbance signals using the high-frequency current sensor HFCT have the advantages of obvious feature differentiation, high identification accuracy, and good noise resistance. The present invention provides a new reference scheme for system transient disturbance identification and has certain practical application potential.

Claims

1. A method for online identification of power transient disturbances, characterized in that: The steps include: (1) The transient disturbance signal in the cable system is collected by the high-frequency current sensor HFCT installed at the cable accessories, and the collected transient disturbance signal is normalized; the specific process is to first use simulation to generate the transient disturbance signal waveform in the 110KV cable system, and then use the signal generator to generate the original signal of the disturbance current on the grounding wire of the cable accessories, input it into the high-frequency current sensor HFCT, and use the oscilloscope to collect the output signal of the real high-frequency current sensor HFCT to obtain the transient disturbance signal after filtering by the high-frequency current sensor HFCT, and then use the maximum-minimum normalization method to normalize the collected transient disturbance signal; (2) Penalty factor for variational mode decomposition using the White Shark optimization algorithm a and the modal number K The two parameters are automatically optimized to establish the penalty factor a and the modal number K The optimal combination of the high-frequency current sensor HFCT filter transient disturbance signal is adaptively and accurately decomposed into several components of different frequency components BIMF1~BIMF through the variational mode decomposition after parameter optimization. K , extract the peak coefficient C of each component respectively 1K , Kurtosis C 2K , Pulse Factor C 3K and margin factor C 4K There are four feature quantities in total, constructing the waveform feature matrix C i ; (3) Perform Wigner distribution WVD time-frequency analysis on each component to obtain the time-frequency spectrum and extract the time-frequency spectrum feature matrix of each component; The Wigner distribution of each component is obtained respectively, and the following four time-frequency features are extracted for each component based on the Wigner distribution time-frequency spectrum: the mean of the Wigner distribution F 1K ,variance F 2K ,energy F 3K The frequency corresponding to the maximum amplitude F 4K , thus obtaining the time-frequency spectrum feature matrix of the original signal F i ; (4) The waveform feature matrix and the time-frequency spectrum feature matrix are integrated to construct a multiple time-frequency feature matrix, which is used as the input of the long short-term memory network to realize the classification and identification of transient disturbance signals.

2. The method for online identification of power transient disturbance according to claim 1, characterized in that: The waveform feature matrix constructed in step (2) C i as follows: In the formula, C i For the i The waveform feature matrix corresponding to the samples is K is the number of components of variational mode decomposition; For a transient disturbance signal x ( t ), its variational mode decomposition process includes two parts: the construction and solution of the variational problem, where the model construction of the variational problem is: Where: s k ( t )for x ( t ) obtained by variational mode decomposition k modal components; s k }={ s 1, s 2,…, s k };{ w k }={ w 1, w 2,…, w k }; d ( t ) is the Dirac function;¶ t Express t Find partial derivatives; The solution process of the variational problem is: Where: l is the Lagrange multiplication operator; α is the quadratic penalty factor; In addition, when the White Shark optimization algorithm is used to optimize the parameters of variational mode decomposition, the envelope entropy is used as the fitness function, and the expression is: Where: i =1,2,…, N ; E h is the envelope entropy; h i It is the normalized form of the envelope signal obtained after the original signal is Hilbert transformed.

3. The method for online identification of power transient disturbance according to claim 1, characterized in that: The time-frequency spectrum feature matrix of the original signal obtained in step (3) F i as follows: In the formula, F i For the i The time-frequency spectrum feature matrix corresponding to the samples; For a transient disturbance signal x ( t ), whose Wigner distribution is defined as: Where: t is the time difference variable; x *( t )for x ( t )'s conjugate complex number; for x ( t )'s instantaneous correlation function.

4. The method for online identification of power transient disturbance according to claim 1, characterized in that: The multiple time-frequency feature matrix in step (4) is composed of the waveform feature matrix in step (2) and step (3) C i And the time-frequency spectrum feature matrix F i The merged structure is as follows: In the formula, Y i is the multiple time-frequency feature matrix corresponding to the i-th sample; The multiple time-frequency feature matrices of all samples extracted by the above method are a % as training set, (100- a )% as the test set, and then input the training set into the long short-term memory network for training. The features are screened through the three gating structures of the internal forget gate, input gate, and output gate and the gating characteristics of the activation function, and the correlation between various features is sought to obtain the best network model. The trained long short-term memory network is then used to learn the test set, and the Softmax classifier is used to classify the test data, and finally the recognition result of the transient disturbance signal is output.

Images