Photovoltaic arc detection method and system based on neural network

Abstract

The invention discloses a photovoltaic arc detection method and system based on a neural network. The method comprises the following steps: generating a data set; preprocessing the data set to obtain a training data set; a hybrid architecture detection model combining a CNN model and a Transform model is constructed, the CNN model processes the voltage time sequence data to extract feature information, the Transform model fuses the feature information at each moment to extract fused feature information, and a predicted value is calculated to adjust parameters of the hybrid architecture detection model to obtain an arc detection model; sampling voltage time sequence data of the photovoltaic system are collected and imported into the arc detection model, and the arc detection model detects whether an arc occurs or not and outputs detection information. According to the method, the accuracy of arc fault detection in the photovoltaic system is greatly improved by fusing the convolutional neural network CNN and the Transform technology, and the safety and stability of operation of the photovoltaic system are improved. The key features of the arc fault can be accurately captured from the complex electric signal, so that the detection is more efficient and stable.

Description

Technical Field

[0001] The present invention relates to a photovoltaic arc detection method and system based on a neural network, and belongs to the technical field of fault arc detection in a photovoltaic system. Background Art

[0002] The combination of the fields of electrical engineering and computer science has extensive applications in today's society. This phenomenon has achieved remarkable development in related fields of artificial intelligence applications, especially in the arc fault detection technology of the photovoltaic industry. The development background of this field mainly benefits from the increasing attention of the industry to the safety of power generation facilities such as photovoltaic power systems, and the continuous improvement of international safety standards.

[0003] Currently, the existing circuit protection devices are very mature in the detection technology of electrical fires caused by traditional electrical faults such as overload and short circuit, and can play a good preventive role in such electrical fires. However, due to the characteristics of randomness, concealment, complexity, etc. of the fault arcs generated by factors such as aging cables, loose terminal blocks, or worn insulation layers, the detection accuracy has always been a difficult problem to be solved. Currently, fuses, circuit breakers, and residual current operated protective devices are the main protection devices in low-voltage distribution networks, and these traditional protection devices are highly targeted at problems such as leakage and overcurrent overload. If a fault arc occurs in the line, it is equivalent to the appearance of a non-linear dynamic resistor, so the current amplitude will not suddenly jump, but may even decrease, and generally will not exceed the setting value of the traditional protection device. All in all, in a low-voltage distribution system, when a high-frequency fault arc occurs due to reasons such as line insulation aging and poor contact, the current amplitude is very small, and it is difficult to distinguish it from the normal load current. At this time, the traditional overcurrent protection device is difficult to detect it.

[0004] Therefore, the detection technology for fault arcs (Arc Fault Detection, AFD) has gradually attracted people's attention. The current mainstream arc fault detection technologies can be mainly divided into two categories: traditional methods based on hardware and intelligent methods based on software. Traditional methods based on hardware include current differential detection technology, ultrasonic detection technology, optical detection technology, voltage and current characteristic analysis technology, etc. Intelligent methods based on software include machine learning and deep learning, signal processing technology, and artificial neural networks.

[0005] However, with the widespread deployment and increasing complexity of photovoltaic systems, traditional hardware-based arc fault detection methods are facing more and more challenges. On the one hand, there are sensitivity and accuracy issues: traditional methods rely on hardware sensors and simple signal processing algorithms, which are prone to false alarms or missed detections in complex environments and are difficult to meet the requirements of high sensitivity and high accuracy. On the other hand, the working environment of photovoltaic systems is complex and variable. When traditional methods face factors such as environmental noise and equipment aging, their performance often deteriorates significantly. In contrast, software-based intelligent methods, by leveraging the powerful capabilities of deep learning, can effectively overcome the deficiencies of traditional methods and provide more powerful, sensitive, and reliable technical support for the safety monitoring of the photovoltaic industry.

[0006] In the prior art, software-based intelligent methods generally include methods based on current or voltage thresholds, machine learning methods based on manually designed features, and methods based on convolutional neural networks or recurrent neural networks. Among them, the method based on current or voltage thresholds detects arc faults by setting thresholds for current or voltage. When the current or voltage exceeds the preset threshold, the system determines that an arc fault has occurred. Although this method is simple to operate and has a low implementation cost, its main limitation is its inability to adapt to changing environmental conditions. In practical applications, the current or voltage of a photovoltaic system may experience normal fluctuations due to factors such as weather changes and equipment aging, which may lead to false alarms or missed detections, thus affecting the accuracy and reliability of the detection. Machine learning methods based on manually designed features identify arc faults by manually designing features and using machine learning algorithms (such as support vector machines, decision trees, etc.). Although it is an improvement over the threshold method and can handle more complex data patterns, its accuracy and generalization ability are limited by the design of feature engineering. Manually designed features may not be able to fully capture all the key information of arc faults, especially when facing complex and variable arc patterns, and its performance may deteriorate significantly. Convolutional neural networks (CNNs) and recurrent neural networks (RNNs) are two commonly used deep learning models that are applied to arc fault detection to improve the detection accuracy. CNNs are good at processing spatial data and can effectively extract the spatial features of arc faults, but they may lose some temporal characteristics when processing temporal data, which limits their application effect in arc detection. RNNs can process sequential data and retain certain temporal information, making them more suitable for arc fault detection. However, traditional RNNs face problems of gradient vanishing or gradient explosion when processing long sequence data and are not ideal for capturing long-term dependencies. Although variants such as long short-term memory networks (LSTMs) and gated recurrent units (GRUs) have alleviated this problem to some extent, there is still room for improvement in complex arc fault detection scenarios. Summary of the Invention

[0007] The object of the present invention is to solve the deficiencies of the above-mentioned existing technologies. Aiming at the problems of large differences in different scenarios and environments, low generalization ability of manual feature engineering, and insufficient feature capture ability of simply using convolutional neural networks or recurrent neural networks in the existing DC arc detection technology, a photovoltaic arc detection method and system based on neural networks are proposed.

[0008] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0009] A photovoltaic arc detection method based on neural networks includes the following steps:

[0010] S1 Construct a data set. Obtain electrical data when the photovoltaic system is working normally and during a fault arc through an experimental platform, and generate a data set. The electrical data includes voltage time series data with arc markings.

[0011] S2 Data preparation and data preprocessing. Preprocess the data set to obtain a training data set.

[0012] S3 Model training. Construct a hybrid architecture detection model combining a CNN model and a Transformer model. The CNN model processes the voltage time series data to extract feature information, and the Transformer model fuses the feature information at each moment to extract fused feature information, calculates the predicted value, and adjusts the parameters of the hybrid architecture detection model to obtain an arc detection model.

[0013] S4 Arc detection. Collect the sampled voltage time series data of the photovoltaic system and import it into the arc detection model. The arc detection model detects whether an arc occurs and outputs detection information.

[0014] Preferably, in step S3, the error is calculated between the predicted value and the true value through a loss function, and the parameters of the arc detection model are adjusted in real time through the error. The error adjustment is repeated iteratively until the error converges to obtain a trained arc detection model.

[0015] Preferably, in step S3, the CNN model includes a one-dimensional convolutional layer, a batch normalization layer, and a max pooling layer. The one-dimensional convolutional layer extracts local features in the voltage time series data, the batch normalization layer normalizes each batch of input data, and the max pooling layer performs data pooling.

[0016] The Transformer model includes a Transformer layer for extracting fused feature information and a linear layer for calculating the predicted value.

[0017] Preferably, in step S1, the data set includes a training set for constructing the arc detection model, a validation set for optimizing the parameters of the arc detection model, and a test set for testing and evaluating the arc detection model.

[0018] Preferably, in step S2, the voltage time series data in the dataset is vlist, and the arc label is label,

[0019] vlist = [v1 v2,..., v m , label = 0 / 1, where V is the voltage value, m is the length of the time value, 0 means no arc, and 1 means there is an arc;

[0020] The voltage time series data in the dataset is segmented to generate a tensor dataset,

[0021] vlist_detect = [v1 v2,..., v n , vlist_detect is the segmented time series segment, n is the length of the segmentation time, vlists_detect = [[v 11 v 12 ,..., v 1n , [v 21 v 22 ,..., v 2n ,..., [[v t1 v t2 ,..., v tn , vlists_detect is the tensor dataset, labes = [0, 1,..., 0], labels is the arc label set, 0 means no arc, and 1 means there is an arc.

[0022] The present invention also proposes a photovoltaic arc detection system based on a neural network,

[0023] including a dataset construction module for constructing a dataset, a data preparation and preprocessing module for data preparation and preprocessing, and a model training and arc detection module for model training and arc detection.

[0024] The beneficial effects of the present invention are mainly reflected in:

[0025] 1. By fusing the convolutional neural network CNN and Transformer technologies, the accuracy of arc fault detection in the photovoltaic system is greatly improved, and the safety and stability of the operation of the photovoltaic system are improved.

[0026] 2. It can accurately capture the key features of arc faults from complex electrical signals, making the detection more efficient and stable.

[0027] 3. It contributes an important technical foundation to the safety monitoring and development of the photovoltaic industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Other features, objectives, and advantages of the present application will become more apparent by reading the detailed description of the non-limiting embodiments with reference to the following drawings:

[0029] Figure 1 It is the training flowchart of the hybrid architecture detection model in the present invention.

[0030] Figure 2 It is the detection flowchart of the arc detection model in the present invention.

[0031] Figure 3 It is the schematic diagram of the data processing flow of the arc detection model in the present invention.

[0032] Figure 4 It is the schematic diagram of the voltage time series data in the present invention. Detailed Embodiments

[0033] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the protection scope of the present invention.

[0034] The present application will be further described in detail below with reference to the drawings and embodiments. It can be understood that the specific embodiments described herein are only used to explain the related invention and are not a limitation to the invention. Additionally, it should be noted that only parts related to the relevant invention are shown in the drawings for the sake of convenience of description. It should be noted that, without conflict, the embodiments in the present application and the features in the embodiments can be combined with each other.

[0035] The present invention provides a photovoltaic arc detection method based on a neural network, including the following steps:

[0036] Construct a dataset, obtain electrical data when the photovoltaic system is operating normally and during a fault arc through an experimental platform, and generate a dataset. The electrical data includes voltage time series data with arc markings.

[0037] Data preparation and data preprocessing, preprocess the dataset to obtain a training dataset.

[0038] Model training, construct a hybrid architecture detection model combining a CNN model and a Transformer model. The CNN model processes the voltage time series data to extract feature information, the Transformer model fuses the feature information at each moment to extract fused feature information, calculates the predicted value, and adjusts the parameters of the hybrid architecture detection model to obtain an arc detection model.

[0039] Arc detection, collect the sampling voltage time series data of the photovoltaic system and import it into the arc detection model, and the arc detection model detects whether an arc occurs and outputs detection information.

[0040] Specific implementation process and principle description:

[0041] This method is mainly divided into two steps: training and detection. Among them, as Figure 1 shown, it is the training flowchart of the hybrid architecture detection model, and as Figure 2 shown, it is the detection flowchart of the arc detection model.

[0042] Refer to Figure 1 for specific description. Since there is no officially published public dataset in the field of arc fault detection, aiming at this limitation, the present invention builds a fault arc experimental platform for electrical data collection, selects various different types of loads, and respectively simulates the normal state and the fault state of the same load in the experimental platform, collects the electrical data in normal operation and fault arcs, and uses all the obtained sample current data as the experimental dataset for research. In the experiment, different load types are mainly considered and as many groups of data as possible are collected to present the diversity of arc currents, and a more comprehensive and more practical experimental dataset is constructed. Among them, the electrical data includes voltage time series data with arc markings.

[0043] After the construction of the dataset is completed, we can use it for the training of the arc detection model. Data preprocessing refers to the processing of data such as cleaning, transformation, and standardization before inputting the data into the model. Its main functions include: improving data quality: cleaning data can remove noise, correct errors, and handle missing values, thereby improving the quality of data and enabling the model to learn more meaningful features; standardization and normalization: standardizing or normalizing data can make different features have the same scale, avoid some features dominating model training due to large numerical ranges, and thus improve the convergence speed and stability of the model; feature engineering: through feature extraction and feature selection, more representative and discriminative features can be extracted from the original data, thereby improving the performance of the model; data augmentation: in fields such as curve graph processing, data augmentation techniques can generate more training samples, increase the diversity of data, and reduce overfitting of the model. Through reasonable data preprocessing, the accuracy of model training can be greatly improved and the time for model training can be reduced.

[0044] After the collection of the dataset and the preprocessing of the data are completed, this part officially introduces the core process model training module of the present invention. The present invention mainly studies deep learning models, including CNN models and Transformer models. The training process and inference process of the present invention are as Figure 1 and Figure 2As shown in the figure. The training process requires preparing a voltage time series dataset. During the training process, the program will send multiple segments of time series data and their labels into the model at one time. The time series data is used to extract feature information through the CNN model, and the feature information at each moment is fused in the Transformer model to extract deeper feature information. Finally, through the calculation of the linear layer, the predicted value of the model is obtained. Thus, the parameter adjustment is realized to obtain the arc detection model with updated parameters.

[0045] As Figure 2 shown in the figure, when performing online arc detection on a photovoltaic system, voltage data is collected in real time and a certain length of historical voltage data is stored. This segment of voltage data is sent into the arc detection model. Through the CNN model and the Transformer model, an output of 0 or 1 can be obtained. 0 indicates no arc, and 1 indicates an arc. If an arc occurs at this time, the photovoltaic system control device will be notified to perform actions such as opening the circuit, playing a role in safety protection. If no arc occurs, according to the acquisition frequency set by the system, real-time voltage data will continue to be collected regularly, and the above reasoning process will be repeated.

[0046] In a specific embodiment, the predicted value and the true value are used to calculate the error through a loss function, and the parameters of the arc detection model are adjusted in real time through the error. The error adjustment is repeated iteratively until the error converges, and the trained arc detection model is obtained.

[0047] Generally, the error convergence tends to 0. Of course, the error convergence in this embodiment means that the parameters of the model converge to a valid interval, that is, the value that satisfies the judgment detection standard is sufficient. The valid interval is a threshold setting for the parameters. The error convergence to a certain extent is the adjustment of the intelligent model, and the convergence region of its various parameters is not limited here.

[0048] The CNN model includes a one-dimensional convolutional layer, a batch normalization layer, and a max pooling layer. The one-dimensional convolutional layer extracts local features in the voltage time series data, the batch normalization layer normalizes each batch of input data, and the max pooling layer performs data pooling; the Transformer model includes a Transformer layer for extracting fused feature information and a linear layer for calculating the predicted value.

[0049] Specifically, in the general process of deep learning, this predicted value will be used to calculate the error with the true value through a loss function, and the parameters of the entire neural network will be adjusted through this error, including the convolutional kernels of the convolutional layer, the projection matrix inside the Transformer model, the weight parameters of the linear layer, and so on. Repeating this process iteratively will gradually reduce the error in each iteration process. When the error gradually converges to near 0, the neural network will also tend to converge. At this time, the neural network has been trained and the trained network parameters can be saved for actual inference applications.

[0050] More specifically, as Figure 3 shown, the overall model is divided into a CNN model and a Transformer model. The CNN model internally contains a one-dimensional convolutional layer, a batch normalization layer, and a max pooling layer. The one-dimensional convolutional layer can effectively extract the features of time series data. Its principle is to perform a convolution operation on the time series data with a one-dimensional convolutional kernel. One-dimensional convolution slides a fixed-size convolutional kernel (filter) over the input sequence and performs a weighted sum (possibly plus a bias) on each subsequence covered by the convolutional kernel, thereby extracting local features in the sequence. This process can capture patterns and trends in sequence data and is suitable for feature extraction of one-dimensional data such as time series analysis and audio processing. Batch normalization mainly normalizes each batch of data input to the network layer to make it have a unified mean and variance, thereby accelerating the training process of the deep neural network and improving the stability of the model. This process helps to alleviate the problem of internal covariate shift during the training process and makes the model easier to optimize. Max Pooling is a pooling operation that is used for downsampling in a convolutional neural network. Its principle is to slide a fixed-size window over the feature map and select the maximum value within each window as the output of the window. In this way, max pooling reduces the subsequent computational amount by retaining the maximum feature value within the window and reducing the size of the feature map at the same time. The main role of max pooling is to improve the robustness of the network to input changes, reduce the number of parameters and computational amount, and help prevent model overfitting, thereby enhancing the generalization ability of the model.

[0051] The Transformer model contains a Transformer and linear layers inside, which are mainly responsible for processing the feature data extracted and compressed by CNN and MaxPooling to capture the long-term dependencies and complex patterns in the sequence. Its main principles are as follows: Self-Attention, the Encoder of the Transformer calculates the importance of each element in the input sequence for other elements in the sequence through the self-attention mechanism. The self-attention mechanism endows the model with the ability to capture long-distance dependencies, which means that even if the elements in the sequence are far apart, the model can learn the relationship between them. Positional Encoding, since the self-attention mechanism itself does not have the ability to capture the sequence order, the Encoder of the Transformer adds positional encoding to retain the position information of the elements in the sequence. The positional encoding is usually a vector of the same dimension as the sequence, which is added to each element of the input sequence. Multi-Head Attention, the Encoder of the Transformer uses the multi-head attention mechanism to process data in parallel. Each "head" captures the information of different parts of the input sequence, improving the model's ability to capture multiple patterns. The role of the Encoder of the Transformer in the model is to use its powerful sequence modeling ability to further analyze and understand the features extracted by CNN and Max Pooling, so as to provide a richer and deeper feature representation for the final tasks (such as classification, detection, etc.).

[0052] In a specific embodiment, the data set includes a training set for building an arc detection model, a validation set for optimizing the parameters of the arc detection model, and a test set for testing and evaluating the arc detection model.

[0053] The sample current data collected by the faulty arc experimental platform is composed into an experimental data set, and the experimental data set is divided into three categories: a training set, a validation set, and a test set. Among them, the training set is used to build a faulty arc detection algorithm model, the validation set is used to tune the parameters of the faulty arc detection algorithm model, and the test set is used to evaluate the performance effect of the optimal parameter detection algorithm model. The positive and negative sample ratios in the training set, validation set, and test set are all 1:10. Such a division of the experimental data set provides a reliable basis for verifying the training detection algorithm model. The test set obtained from both series loads and parallel operation of multiple loads can improve the generalization ability of the algorithm model. This experimental data set has good data transplantation ability, can better simulate real-life scenarios, and has practical guiding significance.

[0054] The creation of the data set and the data loader can be achieved through the following code:

[0055] segment_length = 128

[0056] train_dataset = ArcingDataset(train_data, train_labels, segment_length = segment_length)

[0057] test_dataset = ArcingDataset(test_data, test_labels, segment_length = segment_length)

[0058] train_loader = DataLoader(train_dataset, batch_size = 32, shuffle = true)

[0059] test_loader = DataLoader(test_dataset, batch_size = 32, shuffle = false).

[0060] In a specific embodiment, as Figure 4 shown, the voltage time series data in the dataset is vlist, the arc label is label, vlist = [v1 v2,..., v m , label = 0 / 1, where V is the voltage value, m is the length of the time value, 0 means no arc, and 1 means there is an arc; the voltage time series data in the dataset is segmented to generate a tensor dataset, vlist_detect = [v1v2,..., v n , vlist_detect is the segmented time series segment, n is the length of the segmentation time, and n is much smaller than m. vlists_detect = [[v 11 v 12 ,..., v 1n , [v 21 v 22 ,..., v 2n ,..., [[v t1 v t2 ,..., v tn , vlists_detect is the tensor dataset, labes = [0, 1,..., 0], labels is the arc label set, 0 means no arc, and 1 means there is an arc.

[0061] Specifically, assume that the data we collect are multiple one-dimensional arrays, and the element value of each array represents the voltage value at that moment. At the same time, each one-dimensional array has a corresponding label, where 0 indicates that there is no arc in the voltage sequence, and 1 indicates that there is an arc in the voltage sequence. That is: vlist = [v1 v2,..., v m , label = 0 / 1.

[0062] Since the length of the collected data is very long, in actual arc detection, we usually sample at a certain frequency and then perform arc detection at time intervals such as 0.5 s. The arc detection system usually only stores the voltage values sampled within this time window, while this system performs arc detection based on the voltage sequence within this time window.

[0063] Then, the data fed into the neural network each time is a short voltage sequence, vlist_detect = [v1 v2,..., v n , where the length of n is much smaller than m. Therefore, we can clip the collected long sequence according to the length of n to form multiple voltage sequences, such as: vlist_detect = [[v 11 v 12 ,..., v 1n , [v 21 v 22 ,..., v 2n ,..., [[v t1 v t2 ,..., v tn , and the corresponding labels are: labes = [0, 1,..., 0].

[0064] We design such a training data format based on the characteristics of convolutional neural networks and other considerations of actual arc detection requirements:

[0065] In the photovoltaic DC detection system deployed in the actual environment, real-time voltage data will be sampled at a certain frequency, and these voltage data will be arranged according to the time series, resulting in a very long voltage time-domain data. Convolutional neural networks have strong feature capture capabilities. For abnormal waveforms in time-domain data, convolutional neural networks can quickly locate the abnormal waveforms (usually the positions where arcs occur). Therefore, the training data needs to be designed as a voltage sequence.

[0066] In the actually deployed photovoltaic system, it is possible to sample the voltage at a certain frequency and store the historical voltage data in the system's memory, and then send the voltage time-domain data of the most recent time window into the arc detection device for detection each time, so as to obtain whether an arc has occurred within this time window.

[0067] The following is the PyTorch code implementation for splitting data while assembling it into a tensor dataset:

[0068]

[0069]

[0070] The dataset loading algorithm is divided into two parts. Arcing Dataset is used to read and convert the original data format, which is a long voltage sequence and the label of whether there is an arc. Arcing Dataset will evenly crop a long sequence into a set of voltage sequences with the same length, which can effectively expand the dataset and improve the generalization ability of the model. Data Loader is used to convert the dataset obtained by Arcing Dataset into a data loader suitable for PyTorch. Data Loader can specify the batch size for each time sending data into the model. For example, setting the batch size to 32 means that 32 pieces of data can be sent into the model for training or inference at the same time, effectively improving the training efficiency.

[0071] The present invention also proposes a photovoltaic arc detection system based on a neural network, including a dataset construction module for constructing a dataset, a data preparation and preprocessing module for data preparation and preprocessing, and a model training and arc detection module for model training and arc detection.

[0072] Specifically, due to the lack of an officially published public dataset in the field of arc fault detection, to address this limitation, the present invention builds a fault arc experimental platform for current data acquisition. Select multiple different types of loads and simulate the normal state and fault state of the same load on the experimental platform respectively, collect the current data during normal operation and fault arcs, and use all the obtained sample current data as the experimental dataset for research. In the experiment, different load types are emphasized and as many groups of data as possible are collected to present the diversity of arc currents, construct a more comprehensive and practical experimental dataset, and divide the experimental dataset into a training set, a validation set, and a test set according to an appropriate ratio. Then, preprocess the experimental dataset to provide more powerful data support for subsequent research on the arc fault detection network algorithm model.

[0073] Based on the experimental platform, we conducted an arc detection experiment to obtain a dataset. The sample current data collected by the faulty arc experimental platform was formed into an experimental dataset, which was divided into three categories: training set, validation set, and test set. Among them, the training set was used to construct the faulty arc detection algorithm model, the validation set was used to optimize the parameters of the faulty arc detection algorithm model, and the test set was used to evaluate the performance of the optimal parameter detection algorithm model. The positive and negative sample ratios in the training set, validation set, and test set were all 1:10. Such a division of the experimental dataset provided a reliable basis for validating and training the detection algorithm model. The test set obtained from scenarios including both series loads and parallel operation of multiple loads could improve the generalization ability of the algorithm model. This experimental dataset had good data transplantation ability, could better simulate real-life scenarios, and had practical guiding significance.

[0074] Data preparation and data preprocessing module. After completing the construction of the dataset, it can be used for the training of the arc detection model. Data preprocessing refers to the processes of cleaning, transforming, and normalizing data before inputting it into the model. Its main functions include: improving data quality: cleaning data can remove noise, correct errors, and handle missing values, thereby improving data quality and enabling the model to learn more meaningful features; standardization and normalization: standardizing or normalizing data can make different features have the same scale, avoiding some features dominating model training due to large numerical ranges, and thus improving the convergence speed and stability of the model; feature engineering: through feature extraction and feature selection, more representative and discriminative features can be extracted from the original data, thereby improving the performance of the model; data augmentation: in fields such as image processing, data augmentation techniques (such as rotation, flipping, cropping, etc.) can generate more training samples, increase data diversity, and reduce model overfitting. Through reasonable data preprocessing, the accuracy of model training can be greatly improved and the model training time can be reduced.

[0075] From the above description, it can be found that by fusing the convolutional neural network CNN and Transformer technology, the accuracy of arc fault detection in the photovoltaic system has been significantly improved, enhancing the safety and stability of the photovoltaic system operation. It can accurately capture the key features of arc faults from complex electrical signals, making the detection more efficient and stable. It has contributed an important technical foundation to the safety monitoring and development of the photovoltaic industry.

[0076] The term "comprising" or any other similar term is intended to cover non-exclusive inclusion, so that a process, method, article, or device / equipment including a series of elements not only includes those elements but also other elements not explicitly listed, or also includes elements inherent in these processes, methods, articles, or devices / equipment.

[0077] So far, the technical solution of the present invention has been described in conjunction with the preferred embodiments shown in the accompanying drawings. However, it is easy for those skilled in the art to understand that the protection scope of the present invention is obviously not limited to these specific embodiments. Without departing from the principle of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the protection scope of the present invention.

Claims

1. A photovoltaic arc detection method based on a neural network, characterized in that It includes the following steps: S1 Construct a data set, obtain electrical data when the photovoltaic system is working normally and during a fault arc through an experimental platform, and generate a data set. The electrical data includes voltage time series data with arc markings; S2 Data preparation and preprocessing, preprocess the data set to obtain a training data set; S3 Model training, construct a hybrid architecture detection model combining a CNN model and a Transformer model. The CNN model processes the voltage time series data to extract feature information, and the Transformer model fuses the feature information at each moment to extract fused feature information, calculates the predicted value, and adjusts the parameters of the hybrid architecture detection model to obtain an arc detection model; S4 Arc detection, collect the sampled voltage time series data of the photovoltaic system and import it into the arc detection model. The arc detection model detects whether an arc occurs and outputs detection information.

2. The photovoltaic arc detection method based on a neural network according to claim 1, wherein: In step S3, the predicted value and the true value calculate the error through a loss function, and the parameters of the arc detection model are adjusted in real time through the error. The error adjustment is repeated iteratively until the error converges to obtain a trained arc detection model.

3. The photovoltaic arc detection method based on a neural network according to claim 2, wherein: In step S3, the CNN model includes a one-dimensional convolutional layer, a batch normalization layer, and a max pooling layer. The one-dimensional convolutional layer extracts local features in the voltage time series data, the batch normalization layer normalizes each batch of input data, and the max pooling layer performs data pooling; The Transformer model includes a Transformer layer for extracting fused feature information and a linear layer for calculating the predicted value.

4. The photovoltaic arc detection method based on a neural network according to claim 2, wherein: In step S1, the data set includes a training set for constructing the arc detection model, a validation set for optimizing the parameters of the arc detection model, and a test set for testing and evaluating the arc detection model.

5. The photovoltaic arc detection method based on a neural network according to claim 1, wherein: In the step S2, the voltage time series data in the dataset is vlist, and the arc label is label, where vlist = [v1 v2,..., v m , label = 0 / 1, where V is the voltage value, m is the length of the time value, 0 means no arc, and 1 means there is an arc; Split the voltage time series data in the dataset to generate a tensor dataset, vlist_detect = [v1 v2,..., v n , vlist_detect is the segmented time series segment, n is the length of the segmentation moment, vlists_detect = [[v 11 v 12 ,..., v 1n , [v 21 v 22 ,..., v 2n ,..., [[v t1 v t2 ,..., v tn , vlists_detect is the tensor dataset, labes = [0, 1,..., 0], labels is the arc label set, 0 means no arc, 1 means there is an arc.

6. An arc detection system for the photovoltaic arc detection method based on a neural network according to any one of claims 1 to 5, wherein: It includes a data set construction module for constructing a data set, a data preparation and preprocessing module for data preparation and preprocessing, and a model training and arc detection module for model training and arc detection.

Images