Intelligent detection method for DC arc fault of photovoltaic power generation system

Abstract

The invention relates to an intelligent detection method for a direct-current arc fault of a photovoltaic power generation system, and the method comprises the steps: carrying out the frequency mixing modulation of a broadband signal through employing a modulation broadband converter and employing a periodic pseudo-random sequence, compressing a frequency band to a baseband region, and completing the sampling of a signal compressed to a baseband through employing a low-rate analog-to-digital converter, the sampling mechanism effectively avoids the bottleneck of traditional Nyquist sampling, hardware cost and power consumption are reduced, then frequency domain features are extracted by adopting wavelet transform, an arc sample library is established, and finally rapid and accurate photovoltaic direct current fault arc detection is realized on the premise of reducing the hardware cost in combination with a deep learning algorithm.

Description

Technical Field

[0001] The invention relates to the technical field of direct current arc fault detection in a photovoltaic power generation system, and in particular to an intelligent detection method for direct current arc fault in a photovoltaic power generation system. Background Art

[0002] As an important part of renewable energy, photovoltaic power generation systems have been widely used and developed in recent years. However, with the increase in the complexity and scale of the system, DC arc faults occur frequently, which seriously affects the stable operation and safety performance of the system, and is easy to cause safety accidents such as fires, resulting in property losses and casualties. Therefore, it is crucial to detect DC fault arcs quickly and accurately. Due to the complex working conditions of photovoltaic power generation systems, they are easily affected by inverter switching noise, light changes, wire length and environmental changes, making it difficult to detect fault arcs and causing misjudgments.

[0003] At present, there are two main methods for detecting arc faults in DC power grids: one is to extract high-frequency noise features through Fourier analysis or wavelet transform, compare the difference between arc and non-arc spectra and set threshold judgment; the other is to use deep neural networks to train models to distinguish arc and non-arc data for detection. These methods mainly focus on low-frequency signals (10-250kHz) for identifying arc faults. However, compared with low-frequency signals, high-frequency signals show stronger anti-interference capabilities against factors such as inverter operation interference, power booster noise, and power line communication. Using traditional hardware, the acquisition and analysis of these high-frequency signals are limited by the Nyquist-Shannon theorem. The sampling frequency must be at least twice the highest frequency contained in the signal to accurately restore the signal, which leads to high hardware costs.

[0004] Patent CN117741370A calculates the autocorrelation coefficient, mean, variance and other parameters to reasonably add windows, and uses Fourier transform to extract time domain, frequency domain and high-frequency modal features, comprehensively considers the arc characteristics, and finally adopts weight fusion strategy and support vector machine to achieve accurate and reliable detection of series arc faults. Patent CN116150588A proposes a photovoltaic system DC arc detection method and device that combines time domain, wavelet domain and frequency domain multi-feature analysis to determine whether a fault arc phenomenon occurs in the photovoltaic system. However, the above detection methods all focus on the low frequency band of the arc signal, and the sampling frequency needs to comply with the Nyquist principle. There are problems such as high hardware cost, poor anti-interference and lack of practicality. Summary of the invention

[0005] The present invention aims to solve the above problems and provides an intelligent detection method for DC arc faults in a photovoltaic power generation system, which solves the problems of high hardware cost, poor anti-interference and lack of practicality of existing detection methods.

[0006] A method for intelligent detection of DC arc faults in a photovoltaic power generation system comprises the following steps: Step S1: Building a DC arc data acquisition platform for a photovoltaic power generation system to obtain current signals in the circuit; Step S2: filtering out the low frequency band of the current signal using a bandpass filter; Step S3: sampling the high frequency band signal after the bandpass filtering by modulating several parallel channels of the broadband converter; Step S4: reconstructing arc signals based on a greedy algorithm; Step S5: extracting the time-frequency domain features of the reconstructed arc signal (DC) by wavelet transform; Step S6: using the time-frequency domain features extracted by the wavelet transform as feature vectors for identifying arcs, and establishing an arc data sample library; Step S7: building a back propagation neural network model, using the arc data sample library to train the back propagation neural network model, and obtaining a trained arc detection model; Step S8: Detect the photovoltaic DC arc based on the arc detection model, and if the result is abnormal, it is determined to be a fault arc.

[0007] Furthermore, the DC arc data acquisition platform in step S1 includes a photovoltaic array, a combiner box, a fault arc generator, an arc measurement box and an inverter, the photovoltaic array is electrically connected to the combiner box, the combiner box is electrically connected to the fault arc generator, the fault arc generator is electrically connected to the arc measurement box, and the inverter is electrically connected to the arc measurement box and the combiner box, respectively.

[0008] Furthermore, the specific process of step S2 is: passing the current signal through a bandpass filter to retain only the high-frequency band of the signal, wherein the frequency of the high-frequency band is greater than 750 kHz.

[0009] Furthermore, the specific process of step S3 is as follows: Step S31: Input signal x(t) ; Step S32: Measured signal x(t) Divided by power divider m channels; Step S33: Each channel uses a different mixing signal p i ( t )right x(t) Modulation Step S34: Shift the signal spectrum to baseband, and each mixed signal is then f s / 2 through a low-pass filter and at a frequency fs =1 / T s The low-speed ADC is used for data acquisition. y i [ n ] completely contains the information of the original signal.

[0010] Furthermore, the input signal x(t) is a Fourier transform X ( f) A specific set of multi-band signals, in the frequency range F s =[−1 / 2 T , 1 / 2 T ] contains N disjoint intervals or subbands, where T = 1 / f NYQ is the inverse of the Nyquist rate, and the bandwidth of each subband does not exceed B Hz; The mixed signal p i ( t ) is a piecewise linear function with a period of T p ,Depend on M intervals, of which p i is a constant, p i is -1 or +1; In the frequency domain, the input signal x(t) and the sampled signal y i [ n ] is:

[0011] in, y ( f )for m A vector consisting of channel sampling sequences, z ( f )=[ z 1 ( f ), ..., z L ( f )] T is the vector formed by shifting the spectrum of the measured signal, and its length is L , A for m × L The CS measurement matrix of m < L , matrix A Each row element corresponds to each channel mixing signal p i ( t )’s Fourier coefficients; The cutoff frequency of the low-pass filter matches the sampling frequency, ensuring m The fixed periodic signal of the circuit remains independent, and the pulse expression of the filter is:

[0012] The cutoff frequency of the filter is f s / 2, and The sampling frequency for each channel.

[0013] Furthermore, the specific process of step S4 is as follows: Step S41: Use matrix transformation to transform the matrix y i [ n ] is reduced in dimension, namely:

[0014] in, J is the framework matrix; Step S42: Initialize and set the support set to S , the loop sign is k , the residual matrix e 0 Equal to the frame matrix J , enter k At the first iteration, the algorithm first calculates the matrix A With the previous residual matrix e k-1 The product of , thus obtaining a new correlation matrix; Step S43: Calculate the binary norm of each column of the correlation matrix, and find the column with the largest norm value. The corresponding position represents the support frequency band to be determined in this iteration, and include the frequency band symmetrical to this support set into the support set S for updating; Step S44: Calculate the orthogonal projection matrix, reconstruct the signal according to the selected atom set, and obtain the current reconstructed signal x k ; Step S45: Update residual e k+1 , that is, the original signal y and the current reconstructed signal x k The difference, number of iterations k+ 1; Step S46: When the number of iterationsk Reached the maximum number of iterations K When , the iteration stops, and the final reconstruction signal is the reconstruction signal when the iteration stops x k .

[0015] Furthermore, the specific process of step S5 is as follows: Step S51: collect the data samples x ( n ) conduct n Layer wavelet decomposition to obtain reconstructed signals in different frequency bands; Step S52: Calculate discrete wavelet coefficients D j ( n ) and the residual component coefficient A m ( n ) Wavelet energy of each frequency band; Step S53: construct the extracted wavelet energy of each frequency band into a feature vector Q, and use it as the input of the back propagation neural network for arc recognition.

[0016] Furthermore, the data sample x ( n ) conduct n Layer wavelet decomposition is equivalent to repeatedly applying a set of high-pass and low-pass filters to gradually decompose the signal. The high-pass filter is responsible for extracting the high-frequency detail components in the signal, and the low-pass filter is responsible for obtaining the low-frequency approximate components of the signal. x ( n ) can be expressed as the sum of these components, namely: ; The wavelet energy calculation formula for each frequency band of the reconstructed signal is:

[0017] is the wavelet decomposition i Tier j node, for The wavelet coefficients of For the i Tier j The energy of a node.

[0018] Furthermore, the specific process of step S6 is: using the time-frequency domain features extracted by the wavelet transform as the feature vector Q for identifying the arc, establishing an arc data sample library, and the feature vector , in is the wavelet coefficient Dj ( n ), is the residual component coefficient A m ( n )'s wavelet energy.

[0019] Furthermore, the specific process of step S7 is: building a back-propagation neural network model, using the arc data sample library to train the back-propagation neural network model, the back-propagation neural network model comprises an input layer, an output layer and a hidden layer, wherein the input layer is used to input the feature vector extracted after wavelet transform, the output layer is used to output the model discrimination result, and the hidden layer comprises one or more levels, each layer has a number of neurons.

[0020] The present invention has the following advantages: 1. The modulation broadband converter uses a periodic pseudo-random sequence to mix and modulate the broadband signal, compressing the frequency band to the baseband area, so that a low-rate analog-to-digital converter (ADC) can be used to complete the sampling of the signal compressed to the baseband. This sampling mechanism effectively avoids the bottleneck of traditional Nyquist sampling and reduces hardware cost and power consumption; 2. Signal reconstruction is realized based on the orthogonal matching pursuit algorithm. The number of iterations is reduced through orthogonalization processing, which improves the efficiency and accuracy of signal reconstruction. Before signal reconstruction, matrix transformation is used to reduce the dimension of the matrix, thereby reducing the computational complexity. 3. Pay attention to the high-frequency band of the arc signal, use wavelet transform to extract time-frequency domain features, establish an arc sample library, and combine it with deep learning algorithm to achieve fast and accurate photovoltaic DC fault arc detection while reducing hardware costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only one embodiment of the present invention, and for ordinary technicians in this field, other implementation drawings can be derived from the provided drawings without creative work.

[0022] Figure 1 : Flowchart of the present invention; Figure 2 : Schematic diagram of the photovoltaic DC fault arc data acquisition platform of the present invention; Figure 3 : A structural block diagram of the undersampling system of the present invention; Figure 4: System block diagram of the modulation broadband converter of the present invention; Figure 5 : Detail coefficient of normal signal after wavelet decomposition of the present invention; Figure 6 : Detail coefficient of arc signal after wavelet decomposition of the present invention; Figure 7 : The feature quantity extracted by wavelet transform of the present invention (normal signal); Figure 8 : The feature quantity (arc signal) extracted by wavelet transform of the present invention; Fig. 9 : A structural diagram of the back propagation neural network model of the present invention; Fig.10 : Block diagram of the algorithm steps of the intelligent detection method for DC arc faults in photovoltaic power generation systems based on Nyquist undersampling of the present invention. DETAILED DESCRIPTION

[0023] The present invention will be further described below in conjunction with the accompanying drawings and examples: Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0024] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "installed", "connected", and "connected" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be directly connected or indirectly connected through an intermediate medium. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0025] In the description of the present invention, it should be understood that the terms "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside" and "outside" etc., indicating orientations or positional relationships, are based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as a limitation on the present invention. like Figures 1 to 10 As shown, a method for intelligent detection of DC arc faults in a photovoltaic power generation system comprises the following steps: Step S1: Build a DC arc data acquisition platform for the photovoltaic power generation system to obtain the current signal in the circuit.

[0026] Furthermore, the DC arc data acquisition platform in step S1 includes a photovoltaic array, a combiner box, a fault arc generator, an arc measurement box and an inverter, the photovoltaic array is electrically connected to the combiner box, the combiner box is electrically connected to the fault arc generator, the fault arc generator is electrically connected to the arc measurement box, and the inverter is electrically connected to the arc measurement box and the combiner box, respectively.

[0027] Furthermore, if Figure 2 As shown, the photovoltaic array consists of 12 photovoltaic modules. The arc generator complies with the UL1699B standard and uses circular copper electrodes (the separation speed of the electrodes is adjusted at different levels of 0.1-20 mm / s, and the interval gap is between 0.1-10 mm) to simulate arcing. The arc measurement box consists of a differential amplifier and a high-throughput current transformer. In the experiment, the arc generator is controlled to simulate the normal state and the arc state, and then the current transformer is used to measure the ground current signal in the loop. Finally, a ZDS1000 oscilloscope is used to obtain the current signal in the circuit at a sampling frequency of 2Mhz and transmit it to the host computer.

[0028] Step S2: Filter out the low frequency band of the current signal using a bandpass filter. The current signal is passed through the bandpass filter, and only the high frequency band of the signal (>750 kHz) is retained, thereby enhancing the sparsity of the spectrum so that the signal can be sampled using a modulated broadband converter later.

[0029] Step S3: Sample the high frequency signal after the bandpass filtering by modulating a small number (the number of channels is less than 10) of parallel channels of the broadband converter.

[0030] Furthermore, combined with Figure 3 and Figure 4 The modulated broadband converter is based on the theory of compressed sensing (CS), which has proven that signals that satisfy certain sparsity properties can be perfectly reconstructed from information obtained at sub-Nyquist sampling rates. In high-frequency signal applications, the frequency components are in the megahertz range, and the Nyquist criterion indicates that an ADC capable of collecting millions of samples per second is required. Such ADCs do exist, but they are very expensive and perform poorly when the limits of their sampling capabilities are reached. Therefore, the focus of CS is to exploit the sparsity of the signal and sample at a lower rate than the Nyquist criterion. The CS theoretical framework consists of a sparse representation of the signal, a measurement matrix, and a reconstructed target signal.

[0031] Furthermore, the modulation broadband converter is a completely blind Nyquist undersampling method that uses multiple parallel modulation channels, each of which includes mixing, filtering and sampling. The system can be regarded as a collection of multiple random demodulation channels, which can efficiently process signals and improve performance. The specific steps of using the modulation broadband converter to achieve signal sampling are: first, input signal x(t) , then the measured signal x(t) Divided by power divider m channels, and then each channel uses a different mixing signal p i ( t )right x(t) Modulation is performed and the signal spectrum is shifted to baseband. Each mixed signal is then modulated at the cutoff frequency. f s / 2 through a low-pass filter and at a frequency f s =1 / T s The low-speed ADC is used for data acquisition. y i [ n ] completely contains the information of the original signal.

[0032] Furthermore, the input signal x(t) is a Fourier transform X ( f) A specific set of multi-band signals, in the frequency range F s =[−1 / 2 T , 1 / 2 T ] contains N disjoint intervals or subbands, where T = 1 / f NYQ is the inverse of the Nyquist rate, and the bandwidth of each subband does not exceed B Hz, such a target signal x(t) is considered a sparse bandwidth signal.

[0033] Furthermore, the mixing signal p i ( t ) is a piecewise linear function with a period of T p ,Depend on M intervals, of which p i is a constant, which can be -1 or +1. The purpose of this step is to extend the original sparse target signal in the spectrum domain to the entire spectrum domain.

[0034] Furthermore, in the frequency domain, the input signal x(t) and the sampled signaly i [ n ] is:

[0035] in, y ( f )for m A vector consisting of channel sampling sequences, z ( f )=[ z 1 ( f ), ..., z L ( f )] T is the vector formed by shifting the spectrum of the measured signal, and its length is L , A for m × L The CS measurement matrix of m < L , matrix A Each row element corresponds to each channel mixing signal p i ( t ). Since the input mixing signals of different channels are different, the matrix A The columns of are uncorrelated. Based on this feature, we can use the sampling sequence y i [ n ] and the matrix A To reconstruct the frequency band position of the measured signal, thereby restoring the original signal 。

[0036] Furthermore, the cutoff frequency and sampling frequency of the low-pass filter need to match, while ensuring m The fixed periodic signal of the circuit remains independent. The pulse expression of the filter is:

[0037] The cutoff frequency of the filter is f s / 2, and It is still the sampling frequency of each channel. It can be found that the sampling frequency of all channels is lower than the original Nyquist sampling rate. M times.

[0038] Step S4: Reconstruct the target signal based on the greedy algorithm. The greedy algorithm is an algorithm that takes the best or optimal choice in the current state in each step, hoping to lead to the best or optimal result in the world. It is a sparse approximation method. The most commonly used greedy algorithm is the orthogonal matching pursuit algorithm, which reduces the number of iterations through orthogonalization and improves the efficiency and accuracy of signal reconstruction. Before signal reconstruction, in order to reduce the computational complexity, the matrix can be transformed into y i [ n ] is reduced in dimension, namely:

[0039] in, J is the frame matrix.

[0040] Furthermore, the specific steps of data reconstruction based on the orthogonal matching pursuit algorithm are as follows: first, initialization is performed, and the support set is set to S , the loop sign is k , the residual matrix e 0 Equal to the frame matrix J , enter k At the first iteration, the algorithm first calculates the matrix A With the previous residual matrix e k-1 Then, the bi-norm of each column of the correlation matrix is ​​calculated, and the column with the largest norm value is found. The corresponding position represents the support band to be determined in this iteration, and the band symmetrical to this support set is included in the support set S for updating. Then the orthogonal projection matrix is ​​calculated, and the signal is reconstructed according to the selected atomic set to obtain the current reconstructed signal x k . Update the residual e k+1 , that is, the original signal y and the current reconstructed signal x k The difference, number of iterations k+ 1. When the number of iterations k Reached the maximum number of iterations K When , the iteration stops, and the final reconstruction signal is the reconstruction signal when the iteration stops x k ; Step S5: Using wavelet transform to extract the time-frequency domain features of the reconstructed arc signal (DC).

[0041] Furthermore, the wavelet decomposition step includes: firstly, the collected data samples are x ( n ) conduct nLayer wavelet decomposition is performed to obtain the reconstructed signals in different frequency bands, and then the discrete wavelet coefficients are calculated D j ( n ) and the residual component coefficient A m ( n ) The wavelet energy of each frequency band is extracted and finally the feature vector Q is constructed and used as the input of the back propagation neural network for arc recognition.

[0042] Furthermore, the data sample x ( n ) conduct n Layer wavelet decomposition is equivalent to repeatedly applying a set of high-pass and low-pass filters to gradually decompose the signal. The high-pass filter is responsible for extracting the high-frequency detail components in the signal, and the low-pass filter is responsible for obtaining the low-frequency approximate components of the signal. Therefore, the original signal x ( n ) can be expressed as the sum of these components, namely:

[0043] Furthermore, in order to ensure the effect of signal analysis and processing during wavelet decomposition, thereby reducing the information loss of fault arc signals and improving the stability and reliability of wavelet decomposition, the selection of wavelet basis is crucial. Common wavelet basis functions include Haar, Daubechines (db) and Symlets (sym). This patent has found through a large number of Matlab experiments that compared with other wavelet basis functions, using db4 wavelet basis to decompose signals, such as Figure 5 and Figure 6 , after 4-layer wavelet decomposition, the 4-layer detail coefficients of the normal signal and the fault arc signal are obtained. It can be found through observation that the detail coefficients of the normal operation signal and the fault signal are significantly different after wavelet packet decomposition. Therefore, by analyzing these corresponding detail coefficients, it is possible to effectively determine whether the signal is in a normal or faulty state. The detail coefficients of the reconstructed signal can clearly distinguish the normal signal from the fault arc signal, so the present invention selects the db4 wavelet basis function to process the signal, which is conducive to subsequent feature extraction.

[0044] Step S6: Using the time-frequency domain features extracted by the wavelet transform as feature vectors Q for identifying arcs, and establishing an arc data sample library.

[0045] Furthermore, the present invention obtains energy from the reconstructed detail coefficients as a feature quantity for subsequent identification of fault arcs, such as Figure 7 and Figure 8 As shown, the current signal is decomposed by 4 layers of wavelet to obtain the characteristic vector ,in is the wavelet coefficientD j ( n ), is the residual component coefficient A m ( n ), the wavelet energy calculation formula of each frequency band of the reconstructed signal is:

[0046] in, is the wavelet decomposition i Tier j node, for The wavelet coefficients of For the i Tier j The energy of a node.

[0047] Step S7: construct a back-propagation neural network model, and use the arc data sample library to train the back-propagation neural network model.

[0048] Furthermore, if Fig. 9 and Fig.10 As shown, the back propagation neural network model includes an input layer, an output layer and a hidden layer, wherein the input layer is used to input the feature vector extracted after wavelet transform, the output layer is used to output the model discrimination result, and the hidden layer can include one or more layers, each layer has a number of neurons. The process of training the back propagation neural network model is mainly divided into forward propagation and error back propagation. The difference or error between the actual output and the expected output of the output layer is calculated through forward propagation, and then the weights and thresholds of the neurons in each layer are adjusted and optimized through the reverse correction stage to reduce the error between the actual output and the expected output, thereby improving the prediction or classification performance of the network.

[0049] Furthermore, the number of nodes of the back propagation neural network model is determined. The input of the model is the five feature quantities extracted after wavelet transform. Therefore, the number of nodes in the input layer is set to 5. The output layer only needs one node, and 1 and 2 are used to represent no arc and arc respectively. The number of nodes in the hidden layer is n It is determined to be 10 according to an empirical formula, and the empirical formula is:

[0050] in, n 1 is the number of nodes in the input layer, n 2 is the number of nodes in the output layer, and λ is a constant between 1 and 10.

[0051] Step S8: Detect the photovoltaic DC arc based on the arc detection model, and if the output result is 2, it means that a fault arc occurs.

[0052] The present invention is described above by way of examples, but the present invention is not limited to the above specific embodiments, and any changes or modifications made based on the present invention belong to the scope of protection claimed by the present invention.

Claims

1. An intelligent detection method for DC arc fault in photovoltaic power generation system, characterized in that: The following steps are involved: Step S1: Building a DC arc data acquisition platform for a photovoltaic power generation system to obtain current signals in the circuit; Step S2: filtering out the low frequency band of the current signal using a bandpass filter; Step S3: sampling the high frequency band signal after the bandpass filtering by modulating several parallel channels of the broadband converter; Step S4: reconstructing arc signals based on a greedy algorithm; Step S5: using wavelet transform to extract the time-frequency domain features of the reconstructed arc signal, where the reconstructed arc signal is a DC arc signal; Step S6: using the time-frequency domain features extracted by the wavelet transform as feature vectors for identifying arcs, and establishing an arc data sample library; Step S7: building a back propagation neural network model, using the arc data sample library to train the back propagation neural network model, and obtaining a trained arc detection model; Step S8: Detect the photovoltaic DC arc based on the arc detection model, and if the result is abnormal, it is determined to be a fault arc.

2. The method for intelligent detection of DC arc faults in a photovoltaic power generation system according to claim 1, characterized in that: The DC arc data acquisition platform in step S1 includes a photovoltaic array, a combiner box, a fault arc generator, an arc measurement box and an inverter. The photovoltaic array is electrically connected to the combiner box, the combiner box is electrically connected to the fault arc generator, the fault arc generator is electrically connected to the arc measurement box, and the inverter is electrically connected to the arc measurement box and the combiner box, respectively.

3. The intelligent detection method for DC arc fault in photovoltaic power generation system according to claim 1, characterized in that: The specific process of step S2 is: passing the current signal through a bandpass filter to retain only the high-frequency band of the signal, wherein the low-frequency band is less than or equal to 750 kHz, and the frequency of the high-frequency band is greater than 750 kHz.

4. The intelligent detection method for DC arc fault in photovoltaic power generation system according to claim 1, characterized in that: The specific process of step S3 is: Step S31: Input signal x(t) ; Step S32: Measured signal x(t) Divided by power divider m channels; Step S33: Each channel uses a different mixing signal p i ( t )right x(t) Modulation Step S34: Shift the signal spectrum to baseband, and each mixed signal is then f s / 2 through a low-pass filter and at a frequency f s =1 / T s The low-speed ADC is used for data acquisition. y i [ n ] completely contains the information of the original signal.

5. The method for intelligent detection of DC arc faults in a photovoltaic power generation system according to claim 4, characterized in that: The input signal x(t) is a Fourier transform X ( f) A specific set of multi-band signals, in the frequency range F s =[−1 / 2 T , 1 / 2 T ] contains N disjoint intervals or subbands, where T = 1 / f NYQ is the inverse of the Nyquist rate, and the bandwidth of each subband does not exceed B Hz; The mixed signal p i ( t ) is a piecewise linear function with a period of T p ,Depend on M intervals, of which p i is a constant, p i is -1 or +1; In the frequency domain, the input signal x(t) and the sampled signal y i [ n ] is: in, y ( f )for m A vector consisting of channel sampling sequences, z ( f )=[ z 1( f ), ..., z L ( f )] T is the vector formed by shifting the spectrum of the measured signal, and its length is L , A for m × L The CS measurement matrix of m < L , matrix A Each row element corresponds to each channel mixing signal p i ( t )’s Fourier coefficients; The cutoff frequency of the low-pass filter matches the sampling frequency, ensuring m The fixed periodic signal of the circuit remains independent, and the pulse expression of the filter is: The cutoff frequency of the filter is f s / 2, and The sampling frequency for each channel.

6. A method for intelligent detection of DC arc faults in a photovoltaic power generation system according to claim 1, characterized in that: The specific process of step S4 is as follows: Step S41: Use matrix transformation to transform the matrix y i [ n ] is reduced in dimension, namely: in, J is the framework matrix; Step S42: Initialize and set the support set to S , the loop sign is k , the residual matrix e 0 is equal to the frame matrix J , enter k At the first iteration, the algorithm first calculates the matrix A With the previous residual matrix e k-1 The product of , thus obtaining a new correlation matrix; Step S43: Calculate the binary norm of each column of the correlation matrix, and find the column with the largest norm value. The corresponding position represents the support frequency band to be determined in this iteration, and include the frequency band symmetrical to this support set into the support set S for updating; Step S44: Calculate the orthogonal projection matrix, reconstruct the signal according to the selected atom set, and obtain the current reconstructed signal x k ; Step S45: Update residual e k+1 , that is, the original signal y and the current reconstructed signal x k The difference, number of iterations k+ 1; Step S46: When the number of iterations k Reached the maximum number of iterations K When , the iteration stops, and the final reconstruction signal is the reconstruction signal when the iteration stops x k .

7. A method for intelligent detection of DC arc faults in a photovoltaic power generation system according to claim 1, characterized in that: The specific process of step S5 is as follows: Step S51: The collected data samples are x ( n ) conduct n Layer wavelet decomposition to obtain reconstructed signals in different frequency bands; Step S52: Calculate discrete wavelet coefficients D j ( n ) and the residual component coefficient A m ( n ) Wavelet energy of each frequency band; Step S53: construct the extracted wavelet energy of each frequency band into a feature vector Q, and use it as the input of the back propagation neural network for arc recognition.

8. The method for intelligent detection of DC arc faults in a photovoltaic power generation system according to claim 7, characterized in that: The data sample x ( n ) conduct n Layer wavelet decomposition is equivalent to repeatedly applying a set of high-pass and low-pass filters to gradually decompose the signal. The high-pass filter is responsible for extracting the high-frequency detail components in the signal, and the low-pass filter is responsible for obtaining the low-frequency approximate components of the signal. x ( n ) can be expressed as the sum of these components, namely: ； The wavelet energy calculation formula for each frequency band of the reconstructed signal is: is the wavelet decomposition i Tier j node, for The wavelet coefficients of For the i Tier j The energy of a node.

9. The method for intelligent detection of DC arc faults in a photovoltaic power generation system according to claim 1, characterized in that: The specific process of step S6 is: the current signal is decomposed by 4 layers of wavelet to obtain the characteristic vector ,in is the wavelet coefficient D j ( n ), is the residual component coefficient A m ( n ), the wavelet energy calculation formula for reconstructing each frequency band of the signal is: in, is the wavelet decomposition i Tier j node, for The wavelet coefficients of For the i Tier j The energy of a node.

10. The intelligent detection method for DC arc fault in photovoltaic power generation system according to claim 1, characterized in that: The specific process of step S7 is: building a back-propagation neural network model, using the arc data sample library to train the back-propagation neural network model, the back-propagation neural network model comprises an input layer, an output layer and a hidden layer, wherein the input layer is used to input the feature vector extracted after wavelet transform, the output layer is used to output the model discrimination result, and the hidden layer comprises one or more layers, each layer has a number of neurons; The process of training the back-propagation neural network model is mainly divided into forward propagation and error back propagation. The difference or error between the actual output and the expected output of the output layer is calculated through forward propagation, and then the weights and thresholds of the neurons in each layer are adjusted and optimized through the reverse correction stage to reduce the error between the actual output and the expected output.

Images