A method for diagnosing arc faults in series connection of photovoltaic arrays

Abstract

The present invention proposes a photovoltaic array series arc fault diagnosis method, comprising the following steps: step S1, collecting the current signal at the photovoltaic array DC bus according to a preset sampling rate to form a fault timing signal, and calculating the current signal within a preset time window X The margin factor C ; Step S2, encode the one-dimensional fault timing signal into a two-dimensional feature image to amplify the intrinsic characteristics of the arc fault and form a data set of two-dimensional feature images; Step S3, train the deep convolutional generative adversarial network DCGAN with the data set of two-dimensional feature images; The deep convolutional generative adversarial network uses the data enhancement method to expand the arc fault samples; Step S4, train the convolutional neural network classification model CNN with the data set of two-dimensional feature images and the expanded arc fault samples, and use the convolutional neural network classification model to judge the acquired suspected arc fault samples to identify whether there is an arc fault; The present invention can complete the identification of series arc faults under limited fault sample conditions.

Description

Technical Field

[0001] The invention relates to the technical field of photovoltaic power generation, and in particular to a method for diagnosing a photovoltaic array series arc fault. Background Art

[0002] In recent years, the excessive exploitation and use of fossil energy by humans has led to the gradual deterioration of the global climate and rising temperatures. In response to the problem of global warming, countries have begun to pay attention to the comprehensive utilization of renewable energy. Solar energy is favored as a sustainable clean energy. According to statistics, as of 2021, the global photovoltaic power generation installed capacity has reached 773.2GW, and it is expected to exceed 2TW by 2025. Although the use of photovoltaic power generation to replace part of the traditional non-renewable energy generation has effectively alleviated the greenhouse effect, it also brings many dangerous hidden dangers. Due to the harsh working environment of the photovoltaic power generation system, it is prone to problems such as component aging, wire insulation damage, and loose cable joints, which leads to the frequent occurrence of arc faults, and arc faults are the main cause of fires in photovoltaic systems. In photovoltaic power generation systems, arc faults are generally divided into series arc faults and parallel arc faults. The occurrence of parallel arc faults is usually accompanied by drastic changes in electrical signals, which can be detected by conventional detection equipment. However, when a series arc fault occurs, the electrical signal changes weakly, which is not enough to start the protection device to cut off the fault circuit. With the continuous injection of solar energy, it will further make it difficult to extinguish the arc. Therefore, it is very necessary to design an efficient, fast and accurate series arc fault diagnosis method.

[0003] At present, the existing series arc fault diagnosis algorithms are quite diverse, which can be mainly summarized as time domain, frequency domain and time-frequency domain detection methods. They extract the characteristics of electrical signals in the time and frequency domains and use classifiers or threshold methods to complete the identification of series arc faults. In fact, the characteristics of arc signals are weak, coupled with the influence of inverter switching frequency and interference from environmental factors, resulting in low accuracy, poor robustness and weak generalization of existing recognition algorithms. In addition, the acquisition of series arc fault data is very difficult and time-consuming, while normal data samples are easy to collect. The problem of sample imbalance is also an important reason for the poor diagnostic effect of existing detection algorithms.

[0004] Therefore, the present invention proposes a series arc fault detection algorithm based on waveform image features. The advantage of this algorithm is that it encodes the one-dimensional fault time series signal into a two-dimensional feature image to amplify the intrinsic characteristics of the arc fault, thereby improving the recognition accuracy of the detection algorithm. By using the data enhancement algorithm to expand the arc fault samples, the identification of series arc faults can be completed under limited fault sample conditions. Summary of the invention

[0005] The present invention provides a photovoltaic array series arc fault diagnosis method, which can complete the identification of series arc faults under limited fault sample conditions.

[0006] The present invention adopts the following technical solutions.

[0007] A method for diagnosing a photovoltaic array series arc fault, the method comprising the following steps:

[0008] Step S1: collect the current signal at the DC bus of the photovoltaic array at a preset sampling rate, and calculate the margin factor C of the current signal X in a preset time window, and further obtain the absolute difference value C of the margin factor between two adjacent time windows. d , so as to obtain the fault timing signal;

[0009] Step S2, encoding the one-dimensional fault time series signal into a GASF feature image to amplify the intrinsic characteristics of the arc fault and form a GASF image data set;

[0010] Step S3, training an improved deep convolutional generative adversarial network with the GASF image dataset; the improved deep convolutional generative adversarial network is used to augment arc fault samples;

[0011] Step S4: training a convolutional neural network classification model CNN with the augmented faulty GASF feature image samples and normal GASF image samples, and using the convolutional neural network classification model to judge the acquired suspected arc fault samples to identify whether there is an arc fault.

[0012] The step S1 is a data collection step, which includes the following steps:

[0013] Step S11: Collect the current signal at the DC bus of the photovoltaic array at a preset sampling rate, and calculate the margin factor C of the current signal X in each time window of preset duration, which is specifically defined by the formula:

[0014] Where max and min represent the maximum and minimum values ​​of the current signal X, respectively, N is the length of X, and X i is a single data value in the current signal X;

[0015] Step S12: Using the time window of the preset time length in step S11 as the step length, sliding acquisition and calculating the margin factor C of the next window i+1 , further, calculate the absolute difference value C of the margin factor of two adjacent time windows d , expressed as;

[0016] C d =|C i -C i+1 | Formula 2;

[0017] Among them C i is the margin factor value of the current window, C i+1 is the margin factor value of the next time window;

[0018] Step S13: C calculated each time d Compared with the fault judgment threshold Z, if C d ≥Z, it is determined that a suspected arc fault has occurred. At this time, the acquisition equipment is started to capture the first M data points and the last N data points including the time window as suspected fault samples that need to be further identified; if C d If it is less than Z, then continue to collect and calculate C with the preset time length of step S11 as the time window sliding d .

[0019] In step S11, the sampling rate is 200kHz and the preset duration is 0.25ms; in step S13, M and N are both 500; and Z is 0.22.

[0020] The step S2 is an image conversion step, comprising the following steps:

[0021] Step S21: using the segmented aggregation approximation method to reduce the dimension of the time series current signal X with a length of n, as shown in the following formula 3;

[0022]

[0023] Step S22: The signal after dimension reduction Normalize as shown in the following formula 4;

[0024]

[0025] in is the jth sequence value after normalization;

[0026] Step S23: encoding the normalized data into a polar coordinate system according to the following formula 5;

[0027]

[0028] Among them, φ j express The angle in polar coordinates, r j is the radius, t j is the timestamp, N is the total duration;

[0029] Step S24: Calculate φ at two different times j and, based on the cosine function, the Gram angle sum field (GASF) matrix is ​​constructed, as shown in the following formula 6;

[0030]

[0031] Step S25: Generate a GASF feature image according to the constructed GASF matrix, and use the GASF image to form a two-dimensional feature image data set.

[0032] The step S3 is a data augmentation step, comprising the following steps:

[0033] Step S31: after the arc fault dataset obtained in step S1 is converted into a GASF image through step S2, an improved deep convolutional generative adversarial network DCGAN is trained using the GASF image;

[0034] Step S31: Generate a sufficient number of arc fault GASF feature images using the trained improved DCGAN.

[0035] The improved feature of the deep convolutional generative adversarial network is that a channel attention module SE Model is introduced into the generator to optimize the deep convolutional generative adversarial network's focus performance on useful information of the current task, and the channel attention module includes a compression module, an excitation module and an attention module;

[0036] The compression operation of the compression module is to compress the global spatial information into the channel by using global average pooling, thereby obtaining a set of feature channel weight vectors with global information, and its mathematical expression is:

[0037]

[0038] The excitation operation of the excitation module uses the aggregate information of each channel to obtain the correlation between channels. The specific steps are: the channel feature vector obtained by the compression operation is reduced to T / r channels through a fully connected layer to reduce the amount of calculation, and then the number of channels is restored to T through a second fully connected layer, and finally the sigmoid function is used to obtain the normalized weight of each channel. The mathematical expression of this process is:

[0039] s=F ex (z,W)=σ[g(z,W)]=σ[W2δ(W1z)] Formula 7;

[0040] Among them, δ is the ReLu activation function, W1 and W2 are channel descriptors;

[0041] The attention operation of the attention module is to multiply the weights obtained by the compression operation and the excitation operation by the original features to obtain the final output of the SE Model The calculation formula is:

[0042]

[0043] in, s is a scalar; u∈R H×W ;

[0044] The overall network structure of the improved DCGAN is as follows: for the generator, the input random noise is 1×100, and after a fully connected layer, a BN layer and a tanh activation layer, the output vector is 1×8192, which is reconstructed into a 8×8×128 matrix;

[0045] Then there are two upsampling convolution layers with 128 convolution kernels. The SE Model is integrated into the second and third convolution layers of the generative network. The SE Model makes the features extracted by the convolution layer more accurate and avoids image generation redundancy.

[0046] The last layer of the generated network forms a fake image of size 64×64×3 through convolution operation; the activation function in the generator is tanh;

[0047] The discriminant network uses three consecutive convolution-max pooling layers. The first convolution layer uses 64 convolution kernels, and the others all use 128 convolution kernels. The step size of the maximum pooling layer is (2,2). Finally, the discriminant results are output through two fully connected layers.

[0048] In the discriminative network, the convolution kernel size of all convolutional layers is 5×5, and except for the last layer which uses sigmoid as the activation function, the activation functions of all other convolutional layers are tanh.

[0049] Step S4 is a fault detection step, comprising the following steps:

[0050] Step S41: using the data acquisition technology of step S1 to obtain current data samples under normal conditions, and converting them into GASF images through step S2;

[0051] Step S42: using the normal GASF image samples and the augmented arc GASF feature image samples to train a convolutional neural network classification model CNN, the specific network structure of which is described in the following Table 1:

[0052]

[0053] Step S43: According to the data acquisition technology of step S1, the acquired suspected arc fault samples are input into the trained CNN classification model to complete the identification of the arc fault.

[0054] The photovoltaic array comprises an array structure formed by connecting a plurality of photovoltaic panel assemblies, and the photovoltaic array is connected to the grid for power generation through a three-phase inverter.

[0055] The diagnostic method collects the current signal at the bus by using a current transformer arranged at the DC bus of the photovoltaic array in conjunction with an oscilloscope.

[0056] The advantage of the present invention is that the inherent characteristics of arc faults are amplified by encoding one-dimensional fault time series signals into two-dimensional feature images, thereby improving the recognition accuracy of the detection algorithm. Arc fault samples are augmented by using a data enhancement algorithm, and the identification of series arc faults can be completed under limited fault sample conditions.

[0057] The innovation and advantages of the present invention also lie in:

[0058] (1) An innovative and lightweight startup algorithm was designed to achieve the initial screening of series arc fault data. This technology effectively filtered out most of the non-fault data, effectively improved the efficiency of the back-end recognition algorithm, and achieved the capture of the fault moment and the acquisition of series arc fault data.

[0059] (2) The Gram angle and field technology is innovatively used for series arc fault diagnosis. This technology is applied for the first time in the field of photovoltaic array arc fault diagnosis. By converting the one-dimensional time series signal into a two-dimensional image, the weak characteristics of the arc fault are effectively amplified, thereby improving the recognition accuracy of the back-end detection algorithm.

[0060] (3) The deep convolutional generative adversarial network (DCGAN) was innovatively improved by integrating the attention module into the generative network of DCGAN, which effectively improved the quality of augmented samples. The improved DCGAN not only solves the problem of sample imbalance in actual situations, but also reduces the distortion of generated images, provides a large amount of fault data support for the training of the diagnostic network, and further improves the recognition accuracy of the detection algorithm.

[0061] Compared with the existing detection algorithms, the photovoltaic array series arc fault diagnosis method proposed in the present invention has strong anti-interference ability, good adaptability and high accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0063] Attached Figure 1 is a schematic flow chart of step S1 of the present invention;

[0064] Attached Figure 2 is a schematic flow chart of step S2 of the present invention;

[0065] Attached Figure 3 It is a schematic diagram of the overall network structure of the improved DCGAN in step S3 of the present invention;

[0066] Attached Figure 4 It is a schematic diagram of the process of step S3 of the present invention;

[0067] Attached Figure 5is a schematic flow chart of step S4 of the present invention;

[0068] Attached Figure 6 It is a schematic diagram of the overall process of the present invention;

[0069] Attached Figure 7 is a schematic diagram of the photovoltaic array and data collection points;

[0070] Attached Figure 8 It is a schematic diagram comparing the GASF characteristic images under the PV array arc fault state and the normal working state;

[0071] Attached Fig. 9 This is a schematic diagram of the PSNR comparison of DCGAN generated images before and after improvement;

[0072] Attached Fig.10 It is a schematic diagram of the detection results of the present invention in the embodiment. DETAILED DESCRIPTION

[0073] As shown in the figure, a method for diagnosing a photovoltaic array series arc fault is provided, and the method comprises the following steps:

[0074] Step S1: collect the current signal at the DC bus of the photovoltaic array at a preset sampling rate, and calculate the margin factor C of the current signal X in a preset time window, and further obtain the absolute difference value C of the margin factor between two adjacent time windows. d , so as to obtain the fault timing signal;

[0075] Step S2, encoding the one-dimensional fault time series signal into a GASF feature image to amplify the intrinsic characteristics of the arc fault and form a GASF image data set;

[0076] Step S3, training an improved deep convolutional generative adversarial network with the GASF image dataset; the improved deep convolutional generative adversarial network is used to augment arc fault samples;

[0077] Step S4: training a convolutional neural network classification model CNN with the augmented faulty GASF feature image samples and normal GASF image samples, and using the convolutional neural network classification model to judge the acquired suspected arc fault samples to identify whether there is an arc fault.

[0078] The step S1 is a data collection step, which includes the following steps:

[0079] Step S11: Collect the current signal at the DC bus of the photovoltaic array at a preset sampling rate, and calculate the margin factor C of the current signal X in each time window of preset duration, which is specifically defined by the formula:

[0080] Where max and min represent the maximum and minimum values ​​of the current signal X, respectively, N is the length of X, and X i is a single data value in the current signal X;

[0081] Step S12: Using the time window of the preset time length in step S11 as the step length, sliding acquisition and calculating the margin factor C of the next window i+1 , further, calculate the absolute difference value C of the margin factor of two adjacent time windows d , expressed as the formula;

[0082] C d =|C i -C i+1 | Formula 2;

[0083] Among them C i is the margin factor value of the current window, C i+1 is the margin factor value of the next time window;

[0084] Step S13: C calculated each time d Compared with the fault judgment threshold Z, if C d ≥Z, it is determined that a suspected arc fault has occurred. At this time, the acquisition equipment is started to capture the first M data points and the last N data points including the time window as suspected fault samples that need to be further identified; if C d If it is less than Z, then continue to collect and calculate C with the preset time length of step S11 as the time window sliding d .

[0085] In step S11, the sampling rate is 200kHz and the preset duration is 0.25ms; in step S13, M and N are both 500; and Z is 0.22.

[0086] The step S2 is an image conversion step, comprising the following steps:

[0087] Step S21: using the segmented aggregation approximation method to reduce the dimension of the time series current signal X with a length of n, as shown in the following formula 3;

[0088]

[0089] Step S22: The signal after dimension reduction Normalize as shown in the following formula 4;

[0090]

[0091] in is the jth sequence value after normalization;

[0092] Step S23: encoding the normalized data into a polar coordinate system according to the following formula 5;

[0093]

[0094] Among them, φ j express The angle in polar coordinates, r j is the radius, t j is the timestamp, N is the total duration;

[0095] Step S24: Calculate φ at two different times j and, based on the cosine function, the Gram angle sum field (GASF) matrix is ​​constructed, as shown in the following formula 6;

[0096]

[0097] Step S25: Generate a GASF feature image according to the constructed GASF matrix, and use the GASF image to form a two-dimensional feature image data set.

[0098] The step S3 is a data augmentation step, comprising the following steps:

[0099] Step S31: after the arc fault dataset obtained in step S1 is converted into a GASF image through step S2, an improved deep convolutional generative adversarial network DCGAN is trained using the GASF image;

[0100] Step S31: Generate a sufficient number of arc fault GASF feature images using the trained improved DCGAN.

[0101] The improved feature of the deep convolutional generative adversarial network is that a channel attention module SE Model is introduced into the generator to optimize the deep convolutional generative adversarial network's focus performance on useful information of the current task, and the channel attention module includes a compression module, an excitation module and an attention module;

[0102] The compression operation of the compression module is to compress the global spatial information into the channel by using global average pooling, thereby obtaining a set of feature channel weight vectors with global information, and its mathematical expression is:

[0103]

[0104] The excitation operation of the excitation module uses the aggregate information of each channel to obtain the correlation between channels. The specific steps are: the channel feature vector obtained by the compression operation is reduced to T / r channels through a fully connected layer to reduce the amount of calculation, and then the number of channels is restored to T through a second fully connected layer, and finally the sigmoid function is used to obtain the normalized weight of each channel. The mathematical expression of this process is:

[0105] s=F ex (z,W)=σ[g(z,W)]=σ[W2δ(W1z)] Formula 7;

[0106] Among them, δ is the ReLu activation function, W1 and W2 are channel descriptors;

[0107] The attention operation of the attention module is to multiply the weights obtained by the compression operation and the excitation operation by the original features to obtain the final output of the SE Model The calculation formula is:

[0108]

[0109] in, s is a scalar; u∈R H×W ;

[0110] like Figure 3 As shown in the figure, the overall network structure of the improved DCGAN is as follows: for the generator, the input random noise with a dimension of 1×100 is output as a vector with a dimension of 1×8192 after a fully connected layer, a BN layer and a tanh activation layer, and is reconstructed into a matrix of 8×8×128;

[0111] Then there are two upsampling convolution layers with 128 convolution kernels. The SE Model is integrated into the second and third convolution layers of the generative network. The SE Model makes the features extracted by the convolution layer more accurate and avoids image generation redundancy.

[0112] The last layer of the generated network forms a fake image of size 64×64×3 through convolution operation; the activation function in the generator is tanh;

[0113] The discriminant network uses three consecutive convolution-max pooling layers. The first convolution layer uses 64 convolution kernels, and the others all use 128 convolution kernels. The step size of the maximum pooling layer is (2,2). Finally, the discriminant results are output through two fully connected layers.

[0114] In the discriminative network, the convolution kernel size of all convolutional layers is 5×5, and except for the last layer which uses sigmoid as the activation function, the activation functions of all other convolutional layers are tanh.

[0115] Step S4 is a fault detection step, comprising the following steps:

[0116] Step S41: using the data acquisition technology of step S1 to obtain current data samples under normal conditions, and converting them into GASF images through step S2;

[0117] Step S42: using the normal GASF image samples and the augmented arc GASF feature image samples to train a convolutional neural network classification model CNN, the specific network structure of which is described in the following Table 1:

[0118]

[0119] Step S43: According to the data acquisition technology of step S1, the acquired suspected arc fault samples are input into the trained CNN classification model to complete the identification of the arc fault.

[0120] The photovoltaic array comprises an array structure formed by connecting a plurality of photovoltaic panel assemblies, and the photovoltaic array is connected to the grid for power generation through a three-phase inverter.

[0121] The diagnostic method collects the current signal at the bus by using a current transformer arranged at the DC bus of the photovoltaic array in conjunction with an oscilloscope.

[0122] Example:

[0123] In this example, the overall process of the proposed photovoltaic array series arc fault diagnosis method is as follows: Figure 6 First, the series arc fault data is obtained according to step S1, and then converted into GASF feature images in step S2. Then, the improved DCGAN is trained using the arc GASF images, and the fault samples are augmented. Finally, a CNN classification model is trained to further classify and identify the suspected arc fault samples obtained in step S1.

[0124] The data samples obtained are from Figure 7 The photovoltaic array shown in FIG. 1 is composed of 24 photovoltaic modules connected into a 2×12 array structure. A three-phase inverter is used for grid-connected power generation. The specific parameters of the photovoltaic modules and inverters used are shown in Table 2 below. The arc fault generating device is connected in series to the DC bus of the photovoltaic array to simulate a series arc fault. The current transformer CPL8100A is clamped to the DC bus of the photovoltaic array, and the DSOX2024A oscilloscope is used to collect the current signal at the bus. The server parameters used for calculation are: I5-10200H CPU, 16G RAM.

[0125] Table 2

[0126]

[0127]

[0128] exist Figure 7 In the photovoltaic array shown, a total of 2,600 groups of data samples were collected, of which 200 groups of series arc fault data were used to train the improved deep convolutional generative adversarial network and generate 2,000 arc fault GASF feature images.

[0129] 2000 groups of normal data samples are used to train a CNN classification model together with the augmented fault samples. Finally, 100 groups of normal state and series arc fault data samples are used to test the accuracy of the diagnosis model; 100 groups of illumination mutation and MPPT action data samples are used to test the anti-interference ability of the detection model. In this example, a 200kHz sampling rate and a 5ms time window are used to collect data. The distribution of data samples is shown in Table 3 below.

[0130] Table 3

[0131]

[0132] Take 100 groups of normal state, arc fault, MPPT action and illumination mutation data samples respectively, calculate the maximum value of the absolute difference value of the margin factor of each group of samples, and if the value is greater than or equal to 0.22, the sample is judged as a suspected arc fault sample. The accuracy of the proposed startup algorithm is shown in Table 4 below.

[0133] Table 4

[0134]

[0135] It can be seen from Table 4 that the startup algorithm proposed in this example can detect arc faults 100%. For normal conditions and interference conditions, there will be a certain probability of false startup. In actual detection, some non-arc fault samples are effectively screened out, which reduces the execution times of the detection algorithm to a certain extent.

[0136] The Gram angle and field technique used in this example is as follows Figure 8 As shown, by converting one-dimensional data into two-dimensional space, the weak characteristics of arc faults are effectively amplified, and the recognition of normal samples and faulty samples is improved.

[0137] The effect of the improved deep convolutional generative adversarial network mentioned in this example is as follows Fig. 9 As shown in the figure, by integrating the channel attention mechanism in the generation network to avoid image generation redundancy, the quality of augmented samples is improved and the distortion of generated images is reduced.

[0138] Peak signal-to-noise ratio (PSNR) is an important indicator for evaluating image quality. Fig. 9 It can be seen that the PSNR of DCGAN has been in a fluctuating state within 2000 generations, and its stability is poor. On the contrary, the PSNR of the images generated by the improved DCGAN is significantly larger, that is, the improvement measures proposed by the diagnostic method of the present invention effectively reduce the distortion of the image, and after the 255th generation, the PSNR of the generated image gradually stabilizes. In short, after the improvement, the image quality generated by the model and the stability of the model are improved.

[0139] The fault detection technology mentioned in this example has the following detection results: Fig.10 The anti-interference analysis is shown in Table 5. The trained CNN classification model is used to test the test samples and interference shown in Table 3, and the results are as follows: Fig.10 From the confusion matrix shown in the figure, it can be found that the overall recognition rate of the test set is 99%, 2 fault samples are misjudged as normal, and the normal samples can be 100% recognized. According to the detection results shown in Table 5, the proposed algorithm can identify 100% of interference events such as MPPT action and illumination mutation without misjudgment, which fully verifies that the proposed algorithm has good anti-interference ability.

[0140] Table 5

[0141]

Claims

1. A photovoltaic array series arc fault diagnosis method, characterized by: The diagnostic method comprises the following steps: Step S1: collect the current signal at the DC bus of the photovoltaic array at a preset sampling rate, and calculate the margin factor C of the current signal X in a preset time window, and further obtain the absolute difference value C of the margin factor between two adjacent time windows. d , so as to obtain the fault timing signal; Step S2, encoding the one-dimensional fault time series signal into a GASF feature image to amplify the intrinsic characteristics of the arc fault and form a GASF image data set; Step S3, training an improved deep convolutional generative adversarial network with the GASF image dataset; the improved deep convolutional generative adversarial network is used to augment arc fault samples; Step S4, training a convolutional neural network classification model CNN with the augmented faulty GASF feature image samples and normal GASF image samples, and using the convolutional neural network classification model to judge the acquired suspected arc fault samples to identify whether there is an arc fault; The step S1 is a data collection step. The steps include: Step S11: Collect the current signal at the DC bus of the photovoltaic array at a preset sampling rate, and calculate the margin factor C of the current signal X in each time window of preset duration, which is specifically defined by the formula: Where max and min represent the maximum and minimum values ​​of the current signal X, respectively, N is the time length of X, and X i is a single data value in the current signal X; Step S12: Using the time window of the preset time length in step S11 as the step length, sliding acquisition and calculating the margin factor C of the next window i+1 , further, calculate the absolute difference value C of the margin factor of two adjacent time windows d , expressed as; C d =|C i -C i+1 |Formula 2; Among them C i is the margin factor value of the current window, C i+1 is the margin factor value of the next time window; Step S13: C calculated each time d Compared with the fault judgment threshold Z, if C d ≥Z, it is determined that a suspected arc fault has occurred. At this time, the acquisition equipment is started to capture the first M data points and the last N data points including the time window as suspected fault samples that need to be further identified; if C d If it is less than Z, then continue to collect and calculate C with the preset time length of step S11 as the time window sliding d ; The step S3 is a data augmentation step, comprising the following steps: Step S31: after the arc fault dataset obtained in step S1 is converted into a GASF image through step S2, an improved deep convolutional generative adversarial network DCGAN is trained using the GASF image; Step S31: Generate a sufficient number of arc fault GASF feature images using the trained improved DCGAN; the improved feature of the deep convolutional generative adversarial network is that a channel attention module SE Model is introduced into the generator to optimize the deep convolutional generative adversarial network's focus performance on useful information of the current task, and the channel attention module includes a compression module, an excitation module and an attention module; The compression operation of the compression module is to compress the global spatial information into the channel by using global average pooling, thereby obtaining a set of feature channel weight vectors with global information, and its mathematical expression is: The excitation operation of the excitation module uses the aggregated information of each channel to obtain the correlation between channels; the specific steps are: the channel feature vector obtained by the compression operation is reduced to T / r channels through a fully connected layer to reduce the amount of calculation, and then the number of channels is restored to T through a second fully connected layer, and finally the sigmoid function is used to obtain the normalized weight of each channel; the mathematical expression of this process is: s=F ex (z,W)=σ[g(z,W)]=σ[W2δ(W1z)]Formula 7; Among them, δ is the ReLu activation function, W1 and W2 are channel descriptors; The attention operation of the attention module is to multiply the weights obtained by the compression operation and the excitation operation by the original features to obtain the final output of the SE Model The calculation formula is: in, s is a scalar; u∈R H×W ; The overall network structure of the improved DCGAN is as follows: for the generator, the input random noise is 1×100, and after a fully connected layer, a BN layer and a tanh activation layer, the output vector is 1×8192, which is reconstructed into a 8×8×128 matrix; Then there are two upsampling convolution layers with 128 convolution kernels. The SE Model is integrated into the second and third convolution layers of the generative network. The SE Model makes the features extracted by the convolution layer more accurate and avoids image generation redundancy. The last layer of the generated network forms a fake image of size 64×64×3 through convolution operation; the activation function in the generator is tanh; The discriminant network uses three consecutive convolution-max pooling layers. The first convolution layer uses 64 convolution kernels, and the others all use 128 convolution kernels. The step size of the maximum pooling layer is (2,2). Finally, the discriminant results are output through two fully connected layers. In the discriminative network, the convolution kernel size of all convolutional layers is 5×5, and except for the last layer which uses sigmoid as the activation function, the activation functions of all other convolutional layers are tanh.

2. A photovoltaic array series arc fault diagnosis method according to claim 1, characterized in that: In step S11, the sampling rate is 200kHz and the preset duration is 0.25ms; in step S13, M and N are both 500; and Z is 0.

22.

3. A photovoltaic array series arc fault diagnosis method according to claim 1, characterized in that: The step S2 is an image conversion step, comprising the following steps: Step S21: using the segmented aggregation approximation method to reduce the dimension of the time series current signal X with a length of n, as shown in the following formula 3; Step S22: The signal after dimension reduction Normalize as shown in the following formula 4; in is the jth sequence value after normalization; Step S23: encoding the normalized data into a polar coordinate system according to the following formula 5; Among them, φ j express The angle in polar coordinates, r j is the radius, t j is the timestamp, N is the total duration; Step S24: Calculate φ at two different times j and, based on the cosine function, the Gram angle sum field (GASF) matrix is ​​constructed, as shown in the following formula 6; Step S25: Generate a GASF feature image according to the constructed GASF matrix, and use the GASF image to form a two-dimensional feature image data set.

4. A photovoltaic array series arc fault diagnosis method according to claim 1, characterized in that: The photovoltaic array comprises an array structure formed by connecting a plurality of photovoltaic panel assemblies, and the photovoltaic array is connected to the grid for power generation through a three-phase inverter.

5. A photovoltaic array series arc fault diagnosis method according to claim 1, characterized in that: The diagnostic method collects the current signal at the bus by using a current transformer arranged at the DC bus of the photovoltaic array in conjunction with an oscilloscope.

Images