Photovoltaic DC arc fault detection method and system

Abstract

The invention discloses a photovoltaic DC arc fault detection method and system, and solves the problems of low DC arc fault detection precision, high detection cost and large occupied space of a convolutional network in the prior art, and the method comprises the steps: collecting current data under different working conditions during the operation of a simulated photovoltaic system; randomly generating convolution kernels according to set parameters, randomly generating convolution kernels according to the set parameters, and for each piece of training data and a first-order difference form thereof, performing convolution transformation by using all the generated convolution kernels one by one in a dual-channel parallel form, and constructing a feature vector; training a classifier by using a feature vector set generated after random convolution kernel transformation is applied to all the training data to obtain a time series data classification model based on convolution kernel transformation; and outputting an arc probability by using the time sequence data classification model, and judging whether a direct-current arc exists or not. The convolution structure is optimized, the calculation overhead of the random convolution kernel transformation algorithm is reduced, and the detection efficiency is improved while the detection precision is kept.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of abnormal direct current arc fault detection, and particularly relates to a photovoltaic direct current arc fault detection method and system. BACKGROUND

[0002] Arc fault detection and interruption (AFCI) is a technology integrated in the inverter, which can identify and cut off the circuit at the first time when the arc occurs, and solve the fire hazard caused by direct current arc fault in distributed photovoltaic systems. Generally, the detection of direct current arc is based on the monitoring and analysis of current signals in the system, and the occurrence of abnormal arc in the system is judged by the change of the frequency spectrum energy characteristics of the current signal. Due to the complexity of the operation condition of the photovoltaic system and the volatility of the system output, the detection accuracy and universality of the abnormal arc based on the current signal still have a lot of room for improvement. In addition, the implementation of the abnormal direct current arc detection algorithm needs the support of the hardware computing platform. According to the size and complexity of the algorithm model, the hardware with corresponding computing performance and storage space is needed to support the deployment and operation of the model.

[0003] At present, the direct current arc fault detection scheme based on deep learning model has obvious advantages in detection accuracy and generalization compared with traditional algorithms. For example, the existing patent CN118244060A, a photovoltaic series direct current arc fault detection method based on deep residual shrinkage network, selects current data under different working conditions for preprocessing to obtain a data set for training a deep residual shrinkage network model; the deep residual shrinkage network model is trained and parameter adjusted based on the data set to obtain a photovoltaic series direct current arc fault detection model; the real-time current collected is calculated and processed as the input of the photovoltaic series direct current arc fault detection model; the photovoltaic series direct current arc fault detection model outputs the arc probability and the non-arc probability corresponding to the collected real-time current, and adjusts the arc cumulative value based on the arc probability and the non-arc probability; the arc fault is determined based on the comparison result of the arc cumulative value and the preset arc detection threshold. Although the detection accuracy is improved, the convolutional network needs a deep model layer, which leads to a large number of model parameters and high space occupation after training, and deployment and operation require high computing performance and large storage space. Moreover, the detection model based on deep residual shrinkage network needs to be deployed and implemented on a Raspberry Pi carrier, which has high hardware cost. SUMMARY

[0004] The purpose of the present application is to solve the problems of low detection accuracy, high detection cost and large space occupation of convolutional network in the prior art, and to provide a photovoltaic direct current arc fault detection method and system. The improved random convolution kernel transformation algorithm is used for photovoltaic direct current arc fault detection, the convolution structure is optimized, the calculation overhead of the random convolution kernel transformation algorithm is reduced, the detection efficiency is improved while maintaining the detection accuracy, and the method is suitable for lightweight edge computing platform, which reduces the hardware cost.

[0005] In order to achieve the above object, the present application adopts the following technical solutions: Collecting current data under different working conditions of the analog photovoltaic system during operation as training data; Generating convolution kernels randomly according to the set parameters, and for each training data and its first-order differential form, performing convolution transformation on all generated convolution kernels one by one in a double-channel parallel form to construct a feature vector; Training a classifier using the feature vector set generated after applying the random convolution kernel transformation to all training data, and constructing a time series data classification model based on convolution kernel transformation; Outputting an arc probability using the time series data classification model, and determining whether there is a direct current arc according to the arc probability.

[0006] The photovoltaic direct current arc fault detection method provided by the present application collects current data under different working conditions of the analog photovoltaic system during operation through a data acquisition device integrated with a current sensor, and constructs a model training data set; then, a time series data classification model based on random convolution kernel transformation is trained based on the data set. When detecting photovoltaic direct current arc faults, an arc detection device is designed based on an STM32H7 platform, and the generated random convolution kernel weight and the trained time series data classification model are deployed to the chip end; the arc detection device is mounted to the busbar of the photovoltaic system inverter end to reduce the hardware cost.

[0007] As a preferred, for each training data, the first-order differential form thereof is calculated, for each training data and its first-order differential form, convolution transformation is performed on all generated convolution kernels one by one in a double-channel parallel form, the positive value proportion of each convolution result is calculated separately, all positive value proportions obtained from the training data and its first-order differential form are combined to form a vector, and the feature vector of the training data after random convolution kernel transformation is obtained.

[0008] As a preferred, the parameters for setting the random convolution kernel generation include the number of generated random convolution kernels and the convolution kernel length, and the setting of each convolution kernel weight; for a given convolution kernel, a bias value is extracted from the convolution output of a given expansion combination and a single randomly selected training sample; the padding of the convolution kernel is alternately performed in each convolution kernel / expansion combination, and zero values are added at the beginning and end of the input time series during padding.

[0009] As a preferred, the classifier is a fully connected neural network, the fully connected neural network includes one normalization layer and one fully connected layer, the input of the fully connected neural network is the feature vector of the training data, and the output is a binary classification.

[0010] Preferably, the value of n is determined according to the relationship between the length of the input time series data and the length of the convolution kernel, the values of the power function with 2 as the base and the exponent being a positive integer between 0 and n are calculated respectively, and all the calculation results are combined to form a set D, and the value of the convolution kernel expansion is obtained by randomly selecting from the set D.

[0011] Preferably, each weight of the convolution kernel is selected from -2 and 1, and the sum of all the weights of each convolution kernel is 0.

[0012] Preferably, for the original current data, the first-order difference is the difference between the current data at the last moment and the current data at the current moment.

[0013] A photovoltaic direct current arc fault detection system comprises a simulated photovoltaic system and arc drawing device and a direct current arc fault detection platform, the simulated photovoltaic system and arc drawing device simulates different system currents and system loads under different operating conditions by adjusting a current source and configuring an impedance network, collects operating data under variable system conditions, and the direct current arc fault detection platform collects real-time operating data and performs arc detection model reasoning, and prompts alarm information when an arc fault is detected.

[0014] Preferably, the simulated photovoltaic system and arc drawing device comprises a photovoltaic simulated direct current source, the photovoltaic simulated direct current source is connected to an impedance network, the impedance network is connected to an arc drawing machine, the arc drawing machine is connected to an inverter, the inverter and the arc drawing machine are connected to a data acquisition board through a digital signal, and the data acquisition board is connected to a fault detection unit.

[0015] Preferably, the direct current arc fault detection platform comprises an MCU, the MCU is connected to an external clock and a current sensor, the current sensor is connected to a DC direct current source, the DC direct current source is connected to a DC-DC voltage reduction unit, the DC-DC voltage reduction unit is connected to an LDO unit, the LDO unit is connected to the MCU, the MCU is connected to an indicator light and a wireless communication module.

[0016] Therefore, the present application has the following beneficial effects: the first-order difference form of the original signal is introduced through pre-digital filtering, the double-channel convolution kernel is designed to process the original current signal and the first-order difference signal in parallel, and the feature extraction capability of the algorithm is enhanced; the calculation cost of the random convolution kernel transformation algorithm is reduced by setting special convolution kernel weights and adjusting and optimizing the random rules of the convolution kernel parameters; the most advanced time series classification performance is maintained while maintaining high detection accuracy, the calculation amount is greatly reduced, and the operation efficiency of the algorithm is significantly improved. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 The step flow chart of the photovoltaic direct current arc fault detection method in the present application.

[0018] Figure 2 This is a schematic diagram illustrating the changes in accuracy and loss during the training process of the fully connected neural network in this invention.

[0019] Figure 3 This is a schematic diagram illustrating the changes in the training accuracy of the fully connected neural network in this invention.

[0020] Figure 4 This is a schematic diagram simulating the architecture of a photovoltaic system and an arc-pulling device.

[0021] Figure 5 This is a schematic diagram of the architecture of a DC arc fault detection platform.

[0022] Figure 6 This is a diagram illustrating the results of a real-time test.

[0023] In the diagram: 1. Photovoltaic simulated DC power source; 2. Impedance network; 3. Arc pulling machine; 4. Inverter; 5. Data acquisition board; 6. Fault detection unit; 7. MCU; 8. Current sensor; 9. External clock; 10. DC power supply; 11. DC-DC step-down unit; 12. LDO unit; 13. Indicator light; 14. Wireless communication module. Detailed Implementation

[0024] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments: Example 1: This embodiment provides a method for detecting photovoltaic DC arc faults, such as... Figure 1 As shown, the operation process is as follows: Step 1, collect current data under different operating conditions of the simulated photovoltaic system as training data; Step 2, randomly generate convolution kernels according to set parameters, and for each training data and its first-order difference form, perform convolution transformation one by one using all generated convolution kernels in a dual-channel parallel manner to construct feature vectors; Step 3, use the set of feature vectors generated after applying random convolution kernel transformation to all training data to train a classifier and construct a time-series data classification model based on convolution kernel transformation; Step 4, collect photovoltaic system current data in real time, input it into the time-series data classification model, obtain the arc probability, and determine whether a DC arc exists based on the arc probability.

[0025] The core idea of ​​the Random Convolutional Kernel Transform (ROCKET) algorithm is to first perform convolution operations on time series data using a large number of randomly generated convolutional kernels to extract discriminative features, and then classify these features using a simple linear classifier, thereby efficiently handling time series classification problems. This application optimizes the original random convolutional kernel transform algorithm to address the limitations of arc fault data features and embedded edge computing resources. It proposes a photovoltaic DC arc fault detection method based on an improved random convolutional kernel transform algorithm, suitable for lightweight edge computing platforms. Specifically targeting the characteristics of large fluctuation amplitude and high noise levels in arc signals, it introduces the first-order difference form of the original signal through pre-processing digital filtering, designs a dual-channel convolutional kernel to process the original current signal and the first-order difference signal in parallel, enhancing the algorithm's feature extraction capability. By setting special convolutional kernel weights and adjusting and optimizing the random rules of convolutional kernel parameters (convolutional kernel expansion value, bias value, etc.), the computational overhead of the random convolutional kernel transform algorithm is reduced. While maintaining high detection accuracy, the algorithm's operating efficiency is significantly improved.

[0026] The following examples and specific application scenarios further illustrate the technical solution and effects of the present invention. The following examples are explanations of the present invention, but the present invention is not limited to the following examples.

[0027] like Figure 1 As shown, the photovoltaic DC arc fault detection process specifically includes the following steps: Step (1): Collect current data under different operating conditions of the simulated photovoltaic system as training data.

[0028] A simulated photovoltaic system was used to simulate photovoltaic operation. For each operating condition, both under normal system operation and under arcing conditions, current data of duration t seconds were collected according to the set sampling frequency and sampling time. The collected current data was used as training data to construct a training dataset.

[0029] Step (2): Randomly generate convolution kernels according to the set parameters. For each training data and its first-order difference form, use all the generated convolution kernels in parallel in a dual-channel manner to perform convolution transformation and construct feature vectors.

[0030] Set the parameters for the randomly generated convolutional kernel, following the rules for parameter setting: (a) The number of random convolution kernels generated is set to 1344 (84 x 16) by default in this embodiment.

[0031] (b) Set the kernel length. In this embodiment, the length of each kernel is set to 9.

[0032] (c) Set the weights of the convolution kernel for each kernel. In this embodiment, the weights of each item in the convolution kernel are taken from {-1, 2}, and the sum of the weights of all items is 0. For example, it can be set as: {2, 2, 2, -1, -1, -1, -1, -1, -1}, {2, 2, -1, 2, -1, -1, -1, -1, -1}, {-1, -1, 2, 2, 2, -1, -1, -1, -1}.

[0033] The above design can make the convolution operation in the form of addition, and the convolution kernel is only sensitive to the relative size of the input value, that is, the convolution output is invariant to the translation of the input.

[0034] (d) Set the convolution kernel dilation value, which is randomly selected from set D. Set D includes 0th power of 2, 1st power of 2, …, nth power of 2, where n is determined by the relationship between the length of the input time series data and the length of the convolution kernel.

[0035] Specifically: first, calculate the value of the logarithmic function of the ratio of the difference between the length of the input time series data and 1 to the difference between the length of the convolution kernel and 1, and the final result obtained by rounding down the value of the logarithmic function is the minimum value in 4, which is n.

[0036] The purpose of such setting is to ensure that the convolution kernel contains the maximum effective length after expansion, which is equal to the length of the input time series. For example, for an input sequence length that is 4 times the length of the convolution kernel, n = 2, and the convolution kernel dilation value is randomly selected from [1, 2, 4]; for an input sequence length that is more than 16 times the length of the convolution kernel, n = 4, and the convolution kernel dilation value is randomly selected from [1, 2, 8, 16].

[0037] (e) Set the bias value, which is extracted from the convolution output, for calculating the positive value proportion (PPV) feature.

[0038] Specifically, for a given convolution kernel and dilation combination, the bias value is extracted from the convolution output of the combination and a single randomly selected training sample. For example, for the combination wd of the convolution kernel w = [2, 2, 2, -1, -1, -1, -1, -1, -1] and the convolution kernel dilation value d = 2, a training sample X0 is randomly selected, the convolution output X0*wd is calculated, and the values at the positions of 1 / 4, 1 / 2, and 3 / 4 in the sorted output are randomly extracted as the bias of the combination.

[0039] (f) Set the padding method: the padding method is alternately used in each convolution kernel / dilation combination, so that half of the combinations use padding and the other half do not use padding.

[0040] This embodiment uses the standard zero-padding method, which adds zero values ​​at the beginning and end of the input time series, so that the convolution operation starts from the alignment of the "middle" element of the convolution kernel with the first element of the time series and ends with the alignment with the last element of the time series.

[0041] After setting the convolution kernel parameters, the feature vector is generated using the convolution kernel, specifically including: (g) For each training data point in the training dataset, compute its first-order difference form. For example, for the delayed current signal data X = [x1, x2, ..., x...], compute its first-order difference form. T The first-order difference X' represents the current data x from the previous moment. t+1 Current data x at the current moment t The difference, where t equals 1, 2, ..., T-1.

[0042] (g) Randomly generate convolution kernels according to the set parameters for each training data X. i The difference form X' of the training data i It employs a dual-channel parallel approach to perform convolution transformations one by one using all generated convolution kernels.

[0043] (h) Extract the features after the convolution kernel transformation to obtain the feature vector.

[0044] For each training data point in the training dataset and its difference form after convolution, calculate its positive percentage value (PPV). For each training data point, combine all the positive percentage values ​​obtained from the training data point and its first-order difference form into a vector to obtain the feature vector of that data point after transformation by a random convolution kernel.

[0045] That is, each training data point and its difference form, after passing through a convolution kernel, will yield features. In this embodiment, there are 1344 convolution kernels, meaning that each training data point has 1344 positive percentage values ​​(PPVs), and the difference form of each training data point also has 1344 positive percentage values ​​(PPVs). These 1344+1344 positive percentage values ​​(PPVs) constitute the feature vector of this training data point.

[0046] For example: For a training data point of length m and its first-order difference form, after convolution transformation using 1344 convolution kernels, an output matrix of size [m, 1344, 2] is obtained; the positive value proportion of each convolution result is calculated separately, resulting in 1344 x 2 = 2688 positive value proportions; these 2688 positive value proportions are grouped into a vector of [2688, 1], which is the feature vector obtained after the training data point is transformed by random convolution kernels.

[0047] Step (3): Use the set of feature vectors generated after applying random convolution kernel transformation to all training data to train a classifier and build a time series data classification model based on convolution kernel transformation.

[0048] For each piece of original training data in the training set, a feature vector is generated after transformation by the random convolution kernel in step (2). The feature vectors generated by each piece of training data are combined into a set to obtain a feature vector set. The classifier is trained using the feature vector set to obtain a time series data classification model based on convolution kernel transformation.

[0049] In this embodiment, a fully connected network is used as a classifier for training, which is used to determine the presence or absence of an arc. Specifically, a small fully connected neural network is used, consisting of a normalization layer (nn.BatchNorm1d) and a fully connected layer (nn.linear). The input to the fully connected neural network is a 1x2688 vector, and the output is a 1x2 binary classification. The changes in accuracy and loss (i.e., the changes in loss value during training) of the fully connected network are as follows: Figure 2 As shown, the training loss and validation loss gradually decrease with increasing training epochs, with the most significant decrease occurring in the first five training epochs. The change in training accuracy during the training of a fully connected network is shown in the figure. Figure 3 As shown, the validation accuracy gradually increases with the increase of training cycles. Although the increase is not significant in the first 10 training cycles, it eventually stabilizes at around 0.996.

[0050] Step (4): Collect photovoltaic system current data in real time, input the time series data classification model to obtain the arc probability, and determine whether there is a DC arc based on the arc probability.

[0051] In this embodiment, when using the trained time-series data classification model to determine DC arcs, a total of 1,000 samples are collected. If 1,000 data samples are not collected, the collection continues until 1,000 data samples are collected before proceeding with the subsequent judgment process. These 1,000 data samples are used as a single piece of raw data.

[0052] After the raw data is input into the time series data classification model based on convolution kernel transformation, its first difference is calculated first. Then, an improved random convolution kernel is applied to the raw data and its first difference form to transform them. The positive value proportion (PPV) of the convolution result is calculated. Based on the positive value proportion calculation result, a feature vector is generated. The feature vector is input into the classifier, and the arc probability F of the data sample is output. If the arc probability is greater than 0.5, it is considered that a DC arc has been detected; otherwise, the data sample is re-acquired.

[0053] The photovoltaic direct-current arc fault detection method provided by the embodiment has the following beneficial effects: in the prior art, a direct-current arc fault detection model represented by deep learning is usually based on a deep convolutional network architecture. In order to achieve good detection precision and generalization performance, the convolutional network usually needs to have a relatively deep model layer, which results in a large number of model parameters, high space occupation after training is completed, and the deployment and operation require an edge computing device with high operation performance and large storage space, and it is difficult to be deployed on a resource-limited edge device (such as an MCU or an edge gateway). However, the improved random convolution and algorithm is adopted in the present application, the special convolution kernel weight is set, the random convolution kernel parameter rule (convolution kernel expansion value, bias value, etc.) is adjusted and optimized, the calculation overhead of the random convolution kernel transformation algorithm is reduced, and the STM32H7 embedded platform can be deployed and run, the process of real-time data acquisition and arc fault detection model inference calculation is realized, and the inference speed meets the real-time requirements (<2.5s or before the arc energy exceeds 750J) of the UL1996B standard and the national standard GBT 39750-2021.

[0054] Embodiment two The embodiment provides a photovoltaic direct-current arc fault detection system based on the embodiment one, for realizing the photovoltaic direct-current arc fault detection method provided by the embodiment one, for realizing real-time detection of whether a direct-current arc occurs in a distributed photovoltaic system, and realizing real-time, rapid and accurate judgment of whether an abnormal direct-current arc fault occurs in a running distributed photovoltaic system. In view of the actual needs and cost control requirements of the photovoltaic system abnormal direct-current arc fault detection scheme in the deployment and operation process, the following problems are solved: (1) how to use a computing resource-limited edge embedded chip to build a photovoltaic system arc fault detection scheme and realize high-precision arc fault detection; and (2) how to realize lightweight deployment of a high-precision arc fault detection algorithm to an STM32 series chip and real-time operation.

[0055] Specifically, a photovoltaic direct-current arc fault detection system comprises: an analog photovoltaic system and arc device and a direct-current arc fault detection platform. The analog photovoltaic system and arc device simulate different system currents and system loads under different operating conditions by adjusting a current source and configuring an impedance network, collect operating data under variable system conditions, and the direct-current arc fault detection platform collects real-time operating data and performs arc detection model inference, and prompts alarm information when an arc fault is detected.

[0056] Specifically, as Figure 4As shown, the simulated photovoltaic system and arc-pulling device include a photovoltaic simulated DC source 1, an impedance network 2, an arc-pulling machine 3, an inverter 4, a data acquisition board 5, and a fault detection unit 6. The photovoltaic simulated DC source is connected to the impedance network, the impedance network is connected to the arc-pulling machine, the arc-pulling machine is connected to the inverter, the data acquisition board is connected between the arc-pulling machine and the inverter, and the data acquisition board is connected to the fault detection unit. The fault detection unit and the data acquisition board are connected via digital signals, and the photovoltaic simulated DC source, impedance network, arc-pulling machine, and inverter are connected via a DC bus.

[0057] like Figure 5 As shown, the DC arc fault detection platform includes an MCU7, a current sensor8, an external clock9, a DC power supply10, a DC-DC step-down unit11, an LDO unit12, an indicator light13, and a wireless communication module14. The MCU is connected to the current sensor, external clock, indicator light, and wireless communication module. The current sensor is connected to a 12V DC power supply. The DC power supply is connected to a 12V to 5V DC-DC step-down unit. The DC-DC step-down unit is connected to the LDO unit, and the LDO unit is connected to the MCU. The LDO unit is used to stabilize the output DC voltage and ensure current stability when the input-output voltage difference is extremely low (typically below 400mV). The MCU uses an STM32H743ZI2 chip, and the current sensor is a Hall effect current sensor.

[0058] This embodiment implements the deployment and real-time operation of an improved random convolution kernel transform algorithm based on the STM32H7 platform. Specifically, based on the above photovoltaic DC arc fault detection system, the steps for photovoltaic DC arc fault detection, combined with the photovoltaic DC arc fault detection method in Embodiment 1, mainly include: collecting current data under different operating conditions of the simulated photovoltaic system through a data acquisition device with an integrated current sensor, and constructing a model training dataset; training a time-series data classification model based on random convolution kernel transform based on this dataset; designing a DC arc fault detection platform based on the STM32H7 platform and deploying the generated random convolution kernel weights and the trained feature classifier to the chip; and connecting the DC arc fault detection platform to the inverter bus of the photovoltaic system for operation.

[0059] Specifically, this manifests as follows: Step (1): Construct a simulated photovoltaic system and arc-pushing device, the composition of which is as follows: Figure 4 As shown.

[0060] Step (2): Simulate different system currents and system loads under different operating conditions by adjusting the current source and configuring the impedance network, and collect operating data of the system under varying operating conditions. For each operating condition, collect current data for 2 seconds under normal system operation and under the condition of arc generation by the arc-drawing machine, with a data acquisition frequency of 250kHz.

[0061] Table 1 System operating conditions covered by training data Current range 15-80A Off-grid On-grid, off-grid Operating conditions High voltage ride through, low voltage ride through, phase shift, constant power / constant current switching, etc. Step (3): Generate random convolution kernels and apply convolution kernels to transform the training data. Based on the actual edge algorithm running speed, the hyperparameter convolution kernel number is set to 1344, and the set number of convolution kernels is randomly generated according to the preset convolution kernel parameter rule. Convolution transformation is applied to all training data and feature vectors are extracted.

[0062] Step (4): Train the classifier. This case uses a small fully connected neural network, which consists of a batch normalization layer and a fully connected layer. The network input is a 1x2688 size vector size data, and the output is a 1x2 binary classification.

[0063] Step (5): Build a DC arc fault detection platform based on STM32H743ZI2 chip. The current sensor selects an open-loop Hall effect sensor, so that it can be mounted to the running photovoltaic system bus without disconnecting the connection; the system real-time running data acquisition and arc detection model reasoning, when arc fault is detected, the alarm information is prompted through wireless transmission and external indicator light.

[0064] Step (6): The convolution kernel and the trained classifier are deployed to the MCU in the form of code. The convolution transformation and the classifier function are rewritten in C language, and the convolution kernel weight and the classifier neural network weight are preloaded to the on-chip flash memory. The convolution kernel weight parameter occupies about 40kB, and the classifier model parameter occupies about 90kB.

[0065] Step (7): The device is mounted to the photovoltaic system DC bus for operation. The current data acquisition, operation process and arc judgment logic are shown in Figure 1 .

[0066] Step (8): Adjust the simulation photovoltaic system operating condition setting, collect 10000 times of data under normal and arc conditions respectively and perform inference judgment, and the results are shown in Figure 6 (confusion matrix). The results show that the photovoltaic DC arc fault detection method provided in the embodiment has a running test accuracy of more than 99.6% and a false positive rate of less than one thousandth. Compared with the scheme based on a multi-layer convolutional neural network (CNN), the parameter amount is greatly reduced, and the accuracy is higher than that of a small CNN network with a parameter amount of about 200k.

[0067] The above-described embodiments are only a preferred scheme of the present application, and do not limit the present application in any form. Other variants and modifications can be made without exceeding the technical solutions recited in the claims.

Claims

1. A method for detecting photovoltaic DC arc faults, characterized in that, include: Collect current data under different operating conditions of the simulated photovoltaic system as training data; Convolutional kernels are randomly generated according to the set parameters. For each training data and its first-order difference form, all generated convolutional kernels are used one by one in a dual-channel parallel manner to perform convolution transformation and construct feature vectors. A classifier is trained using the set of feature vectors generated after applying random convolution kernel transformation to all training data, thus constructing a time series data classification model based on convolution kernel transformation. The probability of arcing is output using the time-series data classification model to determine whether a DC arc exists.

2. The photovoltaic DC arc fault detection method according to claim 1, characterized in that, For each training data point, its first-order difference form is calculated. Each training data point and its first-order difference form are processed in parallel using dual-channel convolution kernels. The positive value proportion of each convolution result is calculated separately. All positive value proportions obtained from the training data point and its first-order difference form are combined into a vector to obtain the feature vector of the training data point after transformation by random convolution kernel.

3. The photovoltaic DC arc fault detection method according to claim 1, characterized in that, The parameters for setting the randomly generated convolutional kernels include: the number of random convolutional kernels and the kernel length, and setting the weight of each convolutional kernel; for a given convolutional kernel, a bias value is extracted from the convolutional output of a given dilation combination and a single randomly selected training sample; the kernel padding is performed alternately in each convolutional kernel / dilation combination, and zero values ​​are added at the beginning and end of the input time series during padding.

4. A photovoltaic DC arc fault detection method according to claim 1, 2, or 3, characterized in that, The classifier is a fully connected neural network, which includes a normalization layer and a fully connected layer. The input of the fully connected neural network is the feature vector of the training data, and the output is a binary classification.

5. The photovoltaic DC arc fault detection method according to claim 3, characterized in that, The value of n is determined based on the relationship between the length of the input time series data and the length of the convolution kernel. The values ​​of the power functions with base 2 and positive integers between 0 and n as exponents are calculated respectively. All calculation results are grouped into a set D. The convolution kernel expansion value is obtained by randomly selecting a value from set D.

6. A photovoltaic DC arc fault detection method according to claim 3 or 5, characterized in that, Each weight in the convolution kernel takes a value between -2 and 1, and the sum of all weights in each convolution kernel is 0.

7. The photovoltaic DC arc fault detection method according to claim 2, characterized in that, For the raw current data, its first-order difference is the difference between the current data at the previous moment and the current data at the current moment.

8. A photovoltaic DC arc fault detection system, employing the photovoltaic DC arc fault detection method according to any one of claims 1-7, characterized in that, The system includes a simulated photovoltaic system and an arc-pulling device, as well as a DC arc fault detection platform. The simulated photovoltaic system and arc-pulling device simulate different system currents and system loads under different operating conditions by adjusting the current source and configuring the impedance network, and collect operating data under varying system conditions. The DC arc fault detection platform collects real-time operating data and performs arc detection model inference, and provides alarm information when an arc fault is detected.

9. A photovoltaic DC arc fault detection system according to claim 8, characterized in that, The simulated photovoltaic system and arc-pulling device include a photovoltaic simulated DC source, which is connected to an impedance network. The impedance network is connected to an arc-pulling machine, which is connected to an inverter. The inverter and the arc-pulling machine are connected to a data acquisition board via digital signals. The data acquisition board is connected to a fault detection unit.

10. A photovoltaic DC arc fault detection system according to claim 8 or 9, characterized in that, The DC arc fault detection platform includes an MCU, which is connected to an external clock and a current sensor. The current sensor is connected to a DC power supply, which is connected to a DC-DC step-down unit. The DC-DC step-down unit is connected to an LDO unit, which is connected to the MCU. The MCU is connected to an indicator light and a wireless communication module.

Images