Photovoltaic system DC fault arc detection method based on VMD decomposition and convolutional neural network

Abstract

The invention provides a photovoltaic system DC fault arc detection method based on VMD decomposition and a convolutional neural network. The invention provides a high-efficiency detection method combining signal processing and deep learning in order to solve the problem of detection of direct-current fault arcs in the grid-connected process of a photovoltaic power generation system. Firstly, a direct-current series arc signal under a photovoltaic system is collected through a current transformer, and VMD decomposition is carried out on the signal; and an optimal modal component is selected by using a dispersion entropy value, so that interference caused by ambient noise and inverter power regulation (MPPT) is effectively reduced. Then, signal reconstruction is carried out according to the optimal modal component to enhance signal effectiveness, a time-frequency graph is generated through short-time Fourier transform (STFT), and input data is provided for a convolutional neural network; and finally, the convolutional neural network automatically extracts time-frequency characteristics of the signal, such as frequency abrupt change and transient fluctuation, and high-precision detection and classification of the direct-current fault arc are completed. According to the method, the detection accuracy of the direct-current fault arc in the photovoltaic system is remarkably improved, meanwhile, the interference of environmental noise and inverter power adjustment is effectively reduced, and the method has high robustness and high sensitivity. Experimental results show that the fault arc identification accuracy of the method can reach 99%, and the safety and stability of a photovoltaic system are greatly improved.

Description

Technical Field

[0001] The invention relates to a DC fault arc detection method based on a photovoltaic system, belongs to the field of electrical engineering and intelligent monitoring, and is particularly suitable for detecting DC fault arcs during the grid-connection process of a photovoltaic power generation system. Background Art

[0002] With the widespread application of photovoltaic power generation technology, DC fault arcs, as a potentially dangerous type of fault, have become a major problem affecting the safety and stability of photovoltaic systems. Existing DC fault arc detection methods have the disadvantage of being highly sensitive to interference from environmental noise and inverter power regulation. Traditional DC fault arc detection methods usually rely on time-frequency domain features, but due to the complex environment of photovoltaic systems (such as dynamic power fluctuations regulated by inverter MPPT) and the transient characteristics of non-stationary signals, these methods are more sensitive to external interference and have low detection accuracy. Especially in scenarios where photovoltaic systems are connected to complex loads or are connected to the grid, existing methods find it difficult to effectively separate useful signals from noise, thereby reducing the detection accuracy of fault arcs. Summary of the invention

[0003] The present invention proposes a detection method combining variational mode decomposition (VMD), spread entropy optimization and convolutional neural network (CNN) for DC fault arc detection under photovoltaic system grid connection. By decomposing the signal and optimizing the mode selection, the reconstructed signal can effectively reduce the influence of noise, and then combine the time-frequency feature map with the convolutional neural network model to achieve accurate identification of DC fault arc. The method achieves accurate detection through the following steps:

[0004] S1: Use current transformer to collect DC series arc signals in photovoltaic system.

[0005] S2: Perform variational mode decomposition (VMD) on the collected signal to decompose the signal into multiple intrinsic mode components (IMF). Then, calculate the spread entropy value of each IMF and select the two IMFs with the largest spread entropy values. This step aims to reduce the impact of environmental noise and inverter power regulation by selecting the IMF with the strongest dynamic changes.

[0006] S3: Reconstruct the signal using the selected two optimal IMF components to enhance the effectiveness of the signal and remove invalid noise.

[0007] S4: Perform short-time Fourier transform (STFT) on the reconstructed signal to generate a time-frequency diagram to provide input data for subsequent deep learning algorithms.

[0008] S5: The generated time-frequency diagram is input into the convolutional neural network to automatically identify and classify DC fault arcs, and finally achieve high-precision fault arc detection.

[0009] Furthermore, the step S1 is specifically as follows: the present invention builds a photovoltaic system connected to a resistive load, and the photovoltaic system is connected to an inverter and grid-connected to conduct a DC series fault arc detection study, and uses a high-precision current sensor to collect line current for subsequent signal analysis and processing.

[0010] Furthermore, the step S2 is specifically as follows: after the sensor collects the current, VMD decomposition is performed on it, and the optimal number of decomposition levels is determined according to the spread entropy of different decomposition levels.

[0011] Further, the step S2 is specifically as follows:

[0012]

[0013] The calculation formula of VMD decomposition is shown in formula (1), u k (t) represents the kth modal component, Indicates u k The frequency domain representation of (t) is: is a penalty factor used to balance the constraints in the time domain and the frequency domain, w k is the center frequency of the kth mode, and K is the number of decomposed modes. The VMD decomposition process is as follows: initialize the modal center frequency and signal; perform frequency domain analysis on the signal and assign the signal to each mode; update the frequency and signal of each mode through iterative optimization; until convergence or reaching the preset number of iterations, output the signal of each mode.

[0014] Dispersion entropy (DE) is a method proposed by ROS-TAGHI et al. to detect the dynamic changes or irregularity of time series. This method has fast calculation speed, strong robustness and takes into account the relationship between amplitudes. The larger the dispersion entropy value, the higher the degree of irregularity; the smaller the dispersion entropy value, the lower the degree of irregularity. It can be seen that the stronger the periodic shock of the time series and the more orderly the performance, the smaller the dispersion entropy value. Based on this, the present invention uses dispersion entropy to optimize the IMF selection after VMD decomposition, and selects the two IMF components with the largest dispersion entropy for signal reconstruction.

[0015] Furthermore, the step S3 is specifically as follows: reconstructing the optimal IMF component selected by the spread entropy to obtain a reconstructed signal, and performing time-frequency diagram conversion.

[0016] Further, the step S4 is specifically as follows:

[0017]

[0018] The formula of STFT is shown in formula (2). Among them, X(t,f) is the STFT value at time t and frequency f; x(t) is the original signal; w(τ-t) is the window function (usually Hamming window, Hanning window, rectangular window, etc.), which is used to limit the window size of the signal in the time domain so that the Fourier transform only acts on the signal within the window; e -j2πfτ It is the complex exponential kernel in Fourier transform. By performing STFT transformation on the reconstructed signal, rich information of the signal in the time and frequency domain can be obtained, thereby providing effective input features for the subsequent convolutional neural network. This process enables complex non-stationary signals (DC fault arc signals) to be effectively analyzed and classified, improving the accuracy and robustness of fault detection.

[0019] Further, the step S5 is specifically as follows: the generated time-frequency diagram is input into the CNN model to determine whether a DC fault arc is generated. The CNN extracts the time-frequency domain features (frequency mutation, amplitude fluctuation, transient change) through the convolution layer, and further integrates them through the pooling layer and the fully connected layer to achieve high-precision classification. By further compressing the spatial size of the time-frequency diagram through the pooling layer, CNN can learn more abstract and global features. The fully connected layer fuses these extracted features and finally outputs the classification result. During the training process, the network adjusts the parameters through the back-propagation algorithm, so that the network can accurately determine whether a DC fault arc occurs when facing unknown test data.

[0020] The beneficial effects of the present invention are:

[0021] The present invention combines VMD decomposition, spread entropy optimization and convolutional neural network to significantly improve the accuracy of DC fault arc detection. Compared with traditional methods, the present invention can effectively reduce the errors caused by environmental noise, inverter power regulation (MPPT) and external interference, and improve the sensitivity and reliability of fault arc detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 The overall flow chart of the technical solution of the present invention is

[0023] Figure 2 DC fault arc data acquisition experimental platform

[0024] Figure 3 The original data waveform

[0025] Figure 4 Different IMF waveforms after VMD decomposition

[0026] Figure 5 is the diffusion entropy value of different IMFs

[0027] Figure 6 The reconstructed signal after selecting the optimal IMF based on the spread entropy

[0028] Figure 7 The time-frequency diagram obtained after STFT of the reconstructed signal

[0029] Figure 8 The test set results DETAILED DESCRIPTION

[0030] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0031] Figure 1 The overall flow chart of the technical solution of the present invention is as follows. In the experimental scenario, the current signal is collected through the mutual inductor, and then the original signal is decomposed by VMD, and the optimal IMF component is selected according to the spread entropy to reconstruct the signal. The reconstructed signal is transformed by STFT to obtain a time-frequency diagram, and the generated time-frequency diagram is input into the convolutional neural network to determine whether a fault arc is generated.

[0032] Figure 2 The invention is a DC fault arc data acquisition experimental platform built according to the UL1699B standard. A photovoltaic panel or a DC power supply is used as the DC source input of the experimental platform. The arc generator is connected in series in the main circuit, and the DC switch is used to control the on and off of the branch circuit.

[0033] The arc generator was built with reference to the UL1699B standard. The fault arc generating device mainly includes a fixed electrode, a movable electrode, a spiral rod, a regulator and a fixed base. In order to improve the accuracy and reliability of the data collected in the experiment, the electrode was moved as smoothly as possible, and the copper rod surface was polished with sandpaper to eliminate the residual traces of the previous arcing in each experiment. When the electrode was severely damaged, it was replaced with a new electrode, which can greatly reduce the impact of electrode damage on subsequent signal acquisition and analysis.

[0034] In the experiment, the current signal in the loop is measured by a current transformer, collected by a PicoScope 5000 Series oscilloscope and transmitted to the PC to obtain the normal and fault signals required for signal processing.

[0035] Figure 3 The waveforms of the normal operation and fault current when the arc generator generates an arc are collected under the photovoltaic inverter grid connection of the present invention. Each signal contains 20,000 points. It can be seen from the waveform that the current fluctuates violently when the fault arc occurs. In the grid-connected inverter scenario, due to the influence of noise and the influence of MPPT regulation in the inverter, the normal signal contains signal fluctuations, which increases the difficulty of fault arc diagnosis when the photovoltaic system is grid-connected.

[0036] It can be seen from the original signal waveform that the fault arc signal has the characteristics of non-stationary and transient changes. In view of the typical non-stationary and random vibration signals of the fault arc, VMD decomposition is used to process the collected signals.

[0037] Figure 4 It is the modal components obtained after VMD decomposition of the acquired signal.

[0038]

[0039] The calculation formula of VMD decomposition is shown in formula (1), u k (t) represents the kth modal component, Indicates u k The frequency domain representation of (t) is: is a penalty factor used to balance the constraints in the time domain and frequency domain, w k is the center frequency of the kth mode, and K is the number of decomposed modes. The core idea of ​​VMD is to regard each IMF as a narrowband signal centered on a certain center frequency, and estimate the center frequency and shape of each IMF by minimizing constrained optimization. VMD can ensure that each IMF is smooth and can completely reconstruct the original signal. Since VMD can adaptively decompose different signals, it is more suitable for analyzing non-stationary signals, such as arc signals. It determines the center frequency and bandwidth of each component by searching for the optimal solution of the variational model, so that each component of the arc signal can be analyzed more clearly.

[0040] Figure 5 is the spread entropy of different IMFs after VMD decomposition. The larger the spread entropy value, the higher the degree of irregularity; the smaller the spread entropy value, the lower the degree of irregularity. It can be seen that the stronger the periodic shock of the time series and the more orderly the performance, the smaller the spread entropy value. Based on this, the present invention uses the spread entropy to optimize the IMF selection after VMD decomposition, and selects the two IMF components with the largest spread entropy for signal reconstruction.

[0041] Figure 6 The signal is reconstructed by selecting the optimal modal number through the spread entropy. It can be seen from the reconstructed waveform that the fault reconstructed signal is more regular than the original signal, and the normal reconstructed signal is more regular than the original signal, showing periodic changes.

[0042] Figure 7 The reconstructed signal is subjected to STFT transformation to obtain the time-frequency diagram. It can be seen from the time-frequency diagram that there is an obvious difference between the time-frequency diagrams of the normal reconstructed signal and the faulty reconstructed signal. Using the time-frequency diagram for feature selection is easier to identify faults.

[0043]

[0044] The formula of STFT is shown in formula (2). Among them, X(t,f) is the STFT value at time t and frequency f; x(t) is the original signal; w(τ-t) is the window function (usually Hamming window, Hanning window, rectangular window, etc.), which is used to limit the window size of the signal in the time domain so that the Fourier transform only acts on the signal within the window; e -j2πfτ It is the complex exponential kernel in Fourier transform. By performing STFT transformation on the reconstructed signal, rich information of the signal in the time and frequency domain can be obtained, thereby providing effective input features for the subsequent convolutional neural network. This process enables complex non-stationary signals (DC fault arc signals) to be effectively analyzed and classified, improving the accuracy and robustness of fault detection.

[0045] Figure 8 In order to input the generated time-frequency graph into the recognition result of the convolutional neural network, the method proposed in this patent has a fault recognition accuracy of 99% on the test data set.

Claims

1. A photovoltaic system DC fault arc detection method based on VMD decomposition and convolutional neural network, characterized in that: The following steps are involved: S1: Use current transformer to collect DC series arc signals in photovoltaic system. S2: Perform variational mode decomposition (VMD) on the collected signal to decompose the signal into multiple intrinsic mode components (IMF). Then, calculate the spread entropy value of each IMF and select the two IMFs with the largest spread entropy values. This step aims to reduce the impact of environmental noise and inverter power regulation by selecting the IMF with the strongest dynamic changes. S3: Reconstruct the signal using the selected two optimal IMF components to enhance the effectiveness of the signal and remove invalid noise. S4: Perform short-time Fourier transform (STFT) on the reconstructed signal to generate a time-frequency diagram to provide input data for subsequent deep learning algorithms. S5: The generated time-frequency diagram is input into the convolutional neural network to automatically identify and classify DC fault arcs, and finally achieve high-precision fault arc detection.

2. The method according to claim 1, characterized in that: In step S1, the present invention sets up a scenario where a photovoltaic system is connected to a resistive load and an inverter is connected to the grid to conduct a DC series fault arc detection study, and uses a high-precision current sensor to collect line current for subsequent signal analysis and processing.

3. The method according to claim 1, characterized in that: In step S2, after the sensor collects the current, it performs VMD decomposition on it, and determines the optimal decomposition level according to the spread entropy of different decomposition levels. The calculation formula of VMD decomposition is shown in formula (1), u k (t) represents the kth modal component, Indicates u k The frequency domain representation of (t) is: is a penalty factor used to balance the constraints in the time domain and the flat frequency domain, w k is the center frequency of the kth mode, and K is the number of decomposed modes. The decomposition process of VMD is to initialize the modal center frequency and signal; perform frequency domain analysis on the signal and assign the signal to each mode; Through iterative optimization, the frequency and signal of each mode are updated until convergence or the preset number of iterations is reached, and the signal of each mode is output.

4. The method according to claim 1, characterized in that: In S3, the two optimal IMF components selected by the spread entropy are reconstructed to obtain a reconstructed signal.

5. The method according to claim 1, characterized in that: In step S4, a short-time Fourier transform (STFT) is performed on the reconstructed signal to generate a time-frequency diagram to provide input data for a subsequent neural network model. The formula of STFT is shown in formula (2). Among them, X(t,f) is the STFT value at time t and frequency f; x(t) is the original signal; w(τ-t) is the window function (usually Hamming window, Hanning window, rectangular window, etc.), which is used to limit the window size of the signal in the time domain so that the Fourier transform only acts on the signal within the window; e -j2πfτ It is the complex exponential kernel in Fourier transform. By performing STFT transformation on the reconstructed signal, rich information of the signal in the time and frequency domain can be obtained, thereby providing effective input features for subsequent deep learning models (such as convolutional neural networks). This process enables complex non-stationary signals (such as DC fault arc signals) to be effectively analyzed and classified, improving the accuracy and robustness of fault detection.

6. The method according to claim 1, characterized in that: In step S5, the generated time-frequency graph is input into the CNN model to determine whether a DC fault arc is generated. CNN extracts time-frequency domain features (frequency mutation, amplitude fluctuation, transient change) through the convolution layer, and further integrates them through the pooling layer and the fully connected layer to achieve high-precision classification. By further compressing the spatial size of the time-frequency graph through the pooling layer, CNN can learn more abstract and global features. The fully connected layer fuses these extracted features and finally outputs the classification result. During the training process, the network adjusts the parameters through the back propagation algorithm so that the network can accurately determine whether a DC fault arc occurs when facing unknown test data.

Images