Photovoltaic abnormal DC arc detection method and system

Abstract

The invention discloses a photovoltaic abnormal direct current arc detection method and system, relates to the technical field of arc fault detection, and aims to solve the problem of low accuracy of a current arc detection method, and the method comprises the following steps: S1, collecting direct current data of a photovoltaic system under different operation conditions, and forming a model training data set; s2, sequentially performing noise frequency elimination and normalization preprocessing on the model training data set to obtain a processed model training data set; s3, training the constructed one-dimensional convolutional neural network model based on the processed model training data set to obtain a KAN-based DC arc abnormal fault detection model; and S4, deploying and operating the DC arc abnormal fault detection model, and detecting whether a DC arc is generated in the photovoltaic system. According to the invention, the DC arc abnormal fault detection model based on the KAN neural network is adopted to detect whether the DC arc is generated in the system in real time, and the detection accuracy is high.

Description

Technical Field

[0001] The present invention relates to the technical field of arc fault detection, and in particular to a photovoltaic abnormal DC arc detection method and system. Background Art

[0002] Arc detection and interruption (AFCI) technology is integrated into the inverter and can immediately identify and disconnect the circuit when an arc occurs, thus resolving the fire hazard caused by DC arc faults in distributed photovoltaic systems. Typically, DC arc detection is based on monitoring and analyzing the current signal in the system, determining the occurrence of abnormal arcs in the system by changes in the energy characteristics of the current signal spectrum. Due to the complexity of photovoltaic system operating conditions and the volatility of system output, the accuracy and versatility of abnormal arc detection based on current signals still have significant room for improvement. In addition, the implementation of abnormal DC arc detection algorithms requires the support of a hardware computing platform. Depending on the size and complexity of the algorithm model, hardware with corresponding computing performance and storage space is required to support the deployment and operation of the model.

[0003] Although the Chinese patent publication number CN119250165A can realize the detection of fault arc, its specific implementation path of the detection is different from that of the present application. Summary of the Invention

[0004] The present invention solves the problem of low accuracy of current arc detection methods and proposes a photovoltaic abnormal DC arc detection method and system. The DC arc abnormal fault detection model based on the KAN neural network is used to detect in real time whether a DC arc is generated in the system, with high detection accuracy.

[0005] In order to achieve the above object, the present invention adopts the following technical solution: a photovoltaic abnormal DC arc detection method, comprising the following steps: S1, collects DC current data of the photovoltaic system under different operating conditions to form a model training data set; S2, performing noise frequency elimination and normalization preprocessing on the model training data set in sequence to obtain a processed model training data set; S3, training the constructed one-dimensional convolutional neural network model based on the processed model training data set to obtain a DC arc abnormal fault detection model based on KAN; S4, deploy and run the DC arc abnormal fault detection model to detect whether a DC arc occurs in the photovoltaic system.

[0006] First, current data with and without arcing under different test conditions are collected and a model training dataset is constructed. Then, the constructed model training dataset is subjected to noise frequency elimination and normalization processing. The processed model training dataset is then used to train the model and optimize parameter adjustment to generate a KAN-based DC arc anomaly fault detection model. Finally, the above model is deployed and operated to detect whether there is a DC arc in the photovoltaic system.

[0007] The present invention is further configured as follows: Step S1 includes: building an arc simulation experimental device, collecting current signals with and without arc under different photovoltaic system operating conditions, and digitally sampling the current signals, marking the current signals sampled each time, and forming a model training data set classified by arc and no arc.

[0008] In this technical solution, the corresponding current detection device is used to collect current signals with and without arcs, and data sampling and splitting are performed on them, so as to finally obtain the corresponding model training data set.

[0009] The present invention is further configured such that: the noise frequency elimination specifically includes: Random sampling is performed in the model training data set, and the sampled data is analyzed in the frequency domain to generate a spectrum diagram. The peak frequency in the sampled data spectrum that corresponds to the system switching frequency is selected as the system noise that needs to be eliminated, and a notch filter with the corresponding frequency is used to eliminate the system background noise.

[0010] In this technical solution, noise frequency is removed from all data in the model training data set.

[0011] The present invention is further configured as follows: the normalization preprocessing specifically includes: The data after noise frequency elimination is normalized between 0 and 1. After the normalization is completed, the data is verified twice.

[0012] In this technical solution, normalization processing is performed after noise frequency elimination to ensure that the data amplitude itself does not affect the arc feature judgment; then a secondary verification is performed to determine whether normal data is missing to ensure the security and integrity of the data.

[0013] The present invention is further configured as follows: the DC arc abnormal fault detection model includes three convolution-KAN layers, each convolution-KAN layer is connected to a maximum pooling layer, and the last pooling layer is connected to two KAN fully connected layers.

[0014] In this technical solution, the DC arc abnormal fault detection model is based on the KAN network. A learnable activation function is used at the edge of the network to replace the node parameters of the traditional neural network. Combined with the characteristics of the convolutional neural network, a convolution-KAN layer is designed to build the abnormal DC arc detection model network framework.

[0015] The present invention is further configured as follows: the size of a single input data of the DC arc abnormal fault detection model is 64x64; the DC arc abnormal fault detection model can output a 1x2 binary classification probability of arc presence / no arc.

[0016] The present invention is further configured as follows: Step S4 includes: The trained DC arc detection model is deployed to the photovoltaic system simulation test platform for operation. The data collected in real time by the photovoltaic system is input into the DC arc detection model after steps S1 and S2 to determine whether a DC arc occurs.

[0017] In this technical solution, the model is deployed using an STM32F1 series embedded controller with 256Kb Flash.

[0018] The present invention is further configured as follows: the DC arc abnormal fault detection model network adopts a batch training method, each batch contains 64 data; the training is carried out for 10 training cycles; and the cycle training result with the highest verification accuracy is selected as the final trained model.

[0019] In this technical solution, a detection model with higher accuracy can be obtained through the above-mentioned training.

[0020] The present invention is further configured as follows: the sampling rate of the digital sampling is 250KHz, and the amplitude is 0-4095.

[0021] A photovoltaic abnormal DC arc detection system is applicable to the above photovoltaic abnormal DC arc detection method, comprising: Acquisition module, collects DC current data and forms a model training data set; The preprocessing module performs noise frequency elimination and normalization preprocessing on the model training data set; A training optimization module performs model training to obtain a trained DC arc abnormal fault detection model; Run the module, deploy the operation model, and detect in real time whether a DC arc occurs in the system.

[0022] This technical solution mainly includes an acquisition module, a preprocessing module, a training optimization module and an operation module. After the acquisition module collects data and obtains the corresponding model training data set, it performs noise frequency elimination and normalization processing on the corresponding data set in the preprocessing module, and then trains the corresponding DC arc abnormal fault detection model in the training optimization module, and finally performs real-time detection in the operation module.

[0023] The present invention can bring the following beneficial effects: The present invention relates to a photovoltaic abnormal DC arc detection method, which adopts a DC arc abnormal fault detection model based on the KAN neural network to detect in real time whether a DC arc is generated in the system. By utilizing the high parameter efficiency and strong fitting ability of the KAN network, the number of model parameters is greatly reduced while having detection accuracy and generalization ability comparable to deep convolutional neural networks. This allows the abnormal DC arc detection model to be deployed and run on lighter edge computing platforms (such as the STM32 series) while having high detection capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 This is a schematic diagram of the operation flow of a photovoltaic abnormal DC arc detection method of the present application.

[0025] Figure 2 This is a graph showing the changes in the arc detection model training loss value for this application.

[0026] Figure 3 This is a graph showing the changes in the arc detection model training accuracy for this application. DETAILED DESCRIPTION

[0027] In order to make the objectives, technical solutions and advantages of the present invention more clear, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific implementation method described herein is only an optimal embodiment of the present invention, which is only used to explain the present invention and does not limit the scope of protection of the present invention. All other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0028] Example 1 This embodiment proposes a photovoltaic abnormal DC arc detection method, referring to Figure 1 , which mainly includes the following steps.

[0029] In step S1, direct current data of the photovoltaic system under different operating conditions are first collected to form a corresponding model training data set.

[0030] For the above-mentioned step S1, in more detail, first build a corresponding arc simulation experimental device, collect current signals with and without arc under different photovoltaic system operating conditions, digitally sample the current signals, and mark the current signals sampled each time, so as to finally obtain a model training data set including data with and without arc.

[0031] In this embodiment, the sampling rate of digital sampling is 250KHz, and the amplitude is 0-4095; the current signal collected each time is marked to determine whether there is an arc, and the data with and without arc are split into single data including 4096 sampling points to form a model training data set classified by whether there is an arc.

[0032] After completing step S1, step S2 is performed, that is, performing noise frequency elimination and normalization preprocessing on the model training data set obtained above, and finally obtaining a processed model training data set.

[0033] For the above-mentioned noise frequency elimination, the specific process is as follows: Random sampling is performed in the above-mentioned model training data set, and the sampled data is subjected to frequency domain analysis to generate the corresponding spectrum diagram. The peak frequency in the sampled data spectrum that corresponds to the system switching frequency is selected as the system noise that needs to be eliminated, and a notch filter of the corresponding frequency is designed and applied to eliminate the system background noise.

[0034] In this technical solution, noise frequency is removed from all data in the model training data set.

[0035] For the above normalization preprocessing, the specific process is as follows: After noise frequency elimination, the data is normalized between 0 and 1 to ensure that the data amplitude itself does not affect the arc feature judgment. After normalization, the data is double-checked to ensure data integrity.

[0036] In this technical solution, normalization processing is performed after the noise frequency is eliminated to ensure that the data amplitude itself does not affect the arc feature judgment; then a secondary verification is performed to determine whether normal data is missing. If normal data is missing, it returns to step S1. If normal data is not missing, it enters the subsequent step S3.

[0037] In step S3, a one-dimensional convolutional neural network model is trained according to the processed model training data set, and finally a DC arc abnormal fault detection model based on KAN is obtained.

[0038] More specifically, by using the model training data set to train the one-dimensional convolutional neural network framework and combining it with the model parameter tuning algorithm, a DC arc abnormal fault detection model can be obtained.

[0039] In this technical solution, the proposed photovoltaic system DC arc abnormal fault detection model is based on the Kolmogorov-Arnold (KAN) network. A learnable activation function is used at the network edge to replace the node parameters of the traditional neural network. Combined with the characteristics of the convolutional neural network, a convolution-KAN layer is designed to build the DC arc abnormal fault detection model network framework.

[0040] The DC arc fault detection model consists of three convolutional-KAN layers with a 3x3 kernel size and a convolution stride of 1. Each convolutional-KAN layer is followed by a maximum pooling layer with a pooling window size of 2x2. The final pooling layer is followed by two fully connected KAN layers. All parameters of the KAN layer are learnable B-spline functions.

[0041] The above model has a single input data size of 64x64; the model outputs a 1x2 binary classification probability of arc / no arc.

[0042] The DC arc abnormal fault detection model network based on KAN uses batch training, with each batch containing 64 data. The training is carried out for 10 training cycles. The cycle training result with the highest verification accuracy is selected as the final trained model. The loss value change graph and accuracy change graph of the model training can be referred to Figure 2 and Figure 3 .

[0043] Step S4: deploy and run the above-mentioned DC arc abnormal fault detection model, which can detect whether a DC arc occurs in the photovoltaic system.

[0044] For step S4, more specifically, the model trained in step S3 is deployed to a photovoltaic system simulation test platform for operation. The data collected in real time by the photovoltaic system can be processed in steps S1 and S2 and input into the above-mentioned model for arc detection, thereby determining whether a DC arc is generated.

[0045] In this embodiment, the trained model is deployed to a photovoltaic system simulation test platform for operation. The model is deployed using an STM32F1 series embedded controller with 256KB Flash memory, and the data collection method is the same as step S1.

[0046] For a photovoltaic abnormal DC arc detection method of the present invention, first, current data with and without arcs under different test conditions are collected and a model training data set is constructed; then, the constructed model training data set is subjected to noise frequency elimination and normalization processing; then, the processed model training data set is used to train the model and optimize parameter adjustment to generate a DC arc abnormal fault detection model based on KAN; finally, the above model is deployed and operated, and the model is used to detect whether a DC arc is generated in the photovoltaic system.

[0047] In this technical solution, a model training data set is first constructed. Current data with and without arcing under different test conditions are selected, and all data are divided into training set, validation set, and test set in a ratio of 7:2:1. Then, the training data is preprocessed. A notch filter is used to filter out two switching frequencies in the spectrum (inverter and amplifier circuit), and all data is normalized between [0, 1]. Model training and parameter adjustment optimization are performed based on the training data set to obtain a DC arc abnormal fault detection model. Finally, the model is deployed and operated to detect in real time whether a DC arc is generated in the system.

[0048] In response to the limitations and optimization space in the hardware deployment and operation of the abnormal DC arc fault detection scheme in the photovoltaic system, the present invention designs and trains an abnormal DC arc detection model based on the idea of KAN neural network. It well combines the parameter efficiency of KAN neural network under the lightweight network structure with the excellent characteristics of convolutional neural network in feature extraction. The model test results have excellent accuracy, strong generalization ability and small space occupation; it locates and filters out the fixed switching frequency generated by the operation of the photovoltaic circuit system, thereby reducing the influence of system noise on the judgment of arc spectrum characteristics and improving the accuracy of the arc detection model; the KAN network can achieve high fitting of arc data characteristics with a relatively small number of network layers, so the model has a small number of parameters (only occupies 200KB of storage space), fast inference speed, and low computing resource requirements, which is suitable for deployment and operation on low-cost edge chips.

[0049] Example 2 This embodiment provides a photovoltaic abnormal DC arc detection method, which includes the following steps.

[0050] In step S1, direct current data of the photovoltaic system under different operating conditions are first collected to form a corresponding model training data set.

[0051] For the above-mentioned step S1, in more detail, first build a corresponding arc simulation experimental device, collect current signals with and without arc under different photovoltaic system operating conditions, digitally sample the current signals, and mark the current signals sampled each time, so as to finally obtain a model training data set including data with and without arc.

[0052] In this embodiment, the sampling rate of digital sampling is 250KHz, and the amplitude is 0-4095; the current signal collected each time is marked to determine whether there is an arc, and the data with and without arc are split into single data including 4096 sampling points to form a model training data set classified by whether there is an arc.

[0053] After completing step S1, step S2 is performed, that is, performing noise frequency elimination and normalization preprocessing on the model training data set obtained above, and finally obtaining a processed model training data set.

[0054] For the above-mentioned noise frequency elimination, the specific process is as follows: Random sampling is performed in the above-mentioned model training data set, and the sampled data is subjected to frequency domain analysis to generate the corresponding spectrum diagram. The peak frequency in the sampled data spectrum that corresponds to the system switching frequency is selected as the system noise that needs to be eliminated, and a notch filter of the corresponding frequency is designed and applied to eliminate the system background noise.

[0055] In this technical solution, noise frequency is removed from all data in the model training data set.

[0056] For the above normalization preprocessing, the specific process is as follows: After noise frequency elimination, the data is normalized between 0 and 1 to ensure that the data amplitude itself does not affect the arc feature judgment. After normalization, the data is double-checked to ensure data integrity.

[0057] In this technical solution, normalization processing is performed after the noise frequency is eliminated to ensure that the data amplitude itself does not affect the arc feature judgment; then a secondary verification is performed to determine whether normal data is missing. If normal data is missing, it returns to step S1. If normal data is not missing, it enters the subsequent step S3.

[0058] In step S3, a one-dimensional convolutional neural network model is trained according to the processed model training data set, and finally a DC arc abnormal fault detection model based on KAN is obtained.

[0059] More specifically, by using the model training data set to train the one-dimensional convolutional neural network framework and combining it with the model parameter tuning algorithm, a DC arc abnormal fault detection model can be obtained.

[0060] In this technical solution, the proposed photovoltaic system DC arc abnormal fault detection model is based on the Kolmogorov-Arnold (KAN) network. A learnable activation function is used at the network edge to replace the node parameters of the traditional neural network. Combined with the characteristics of the convolutional neural network, a convolution-KAN layer is designed to build the DC arc abnormal fault detection model network framework.

[0061] The DC arc fault detection model consists of three convolutional-KAN layers with a 3x3 kernel size and a convolution stride of 1. Each convolutional-KAN layer is followed by a maximum pooling layer with a pooling window size of 2x2. The final pooling layer is followed by two fully connected KAN layers. All parameters of the KAN layer are learnable B-spline functions.

[0062] The above model has a single input data size of 64x64; the model outputs a 1x2 binary classification probability of arc / no arc.

[0063] The DC arc abnormal fault detection model network based on KAN uses batch training, with each batch containing 64 data. The training is carried out for 10 training cycles. The cycle training result with the highest verification accuracy is selected as the final trained model. The loss value change graph and accuracy change graph of the model training can be referred to Figure 2 and Figure 3 .

[0064] Step S4: deploy and run the above-mentioned DC arc abnormal fault detection model, which can detect whether a DC arc occurs in the photovoltaic system.

[0065] For step S4, more specifically, the model trained in step S3 is deployed to a photovoltaic system simulation test platform for operation. The data collected in real time by the photovoltaic system can be processed in steps S1 and S2 and input into the above-mentioned model for arc detection, thereby determining whether a DC arc is generated.

[0066] In this embodiment, the trained model is deployed to a photovoltaic system simulation test platform for operation. The model is deployed using an STM32F1 series embedded controller with 256KB Flash memory, and the data collection method is the same as step S1.

[0067] For a photovoltaic abnormal DC arc detection method of the present invention, first, current data with and without arcs under different test conditions are collected and a model training data set is constructed; then, the constructed model training data set is subjected to noise frequency elimination and normalization processing; then, the processed model training data set is used to train the model and optimize parameter adjustment to generate a DC arc abnormal fault detection model based on KAN; finally, the above model is deployed and operated, and the model is used to detect whether a DC arc is generated in the photovoltaic system.

[0068] In response to the limitations and optimization space in the hardware deployment and operation of the abnormal DC arc fault detection scheme in the photovoltaic system, the present invention designs and trains an abnormal DC arc detection model based on the idea of KAN neural network. It well combines the parameter efficiency of KAN neural network under the lightweight network structure with the excellent characteristics of convolutional neural network in feature extraction. The model test results have excellent accuracy, strong generalization ability and small space occupation; it locates and filters out the fixed switching frequency generated by the operation of the photovoltaic circuit system, thereby reducing the influence of system noise on the judgment of arc spectrum characteristics and improving the accuracy of the arc detection model; the KAN network can achieve high fitting of arc data characteristics with a relatively small number of network layers, so the model has a small number of parameters (only occupies 200KB of storage space), fast inference speed, and low computing resource requirements, which is suitable for deployment and operation on low-cost edge chips.

[0069] In this technical solution, a model training data set is first constructed. Current data with and without arcing under different test conditions are selected, and all data are divided into training set, validation set, and test set in a ratio of 7:2:1. Then, the training data is preprocessed. A notch filter is used to filter out two switching frequencies in the spectrum (inverter and amplifier circuit), and all data is normalized between [0, 1]. Model training and parameter adjustment optimization are performed based on the training data set to obtain a DC arc abnormal fault detection model. Finally, the model is deployed and operated to detect in real time whether a DC arc is generated in the system.

[0070] Different from Example 1, this embodiment further proposes a photovoltaic abnormal DC arc detection system, which includes an acquisition module, a preprocessing module, a training optimization module, and an operation module.

[0071] Among them, the acquisition module is connected to the preprocessing module, the preprocessing module is connected to the training optimization module, and the training optimization module is connected to the operation module.

[0072] For the acquisition module, its main function is to collect the corresponding DC current data, and then mark these current data to form the corresponding model training data set.

[0073] The preprocessing module mainly performs the functions of eliminating noise frequency and normalizing the preprocessing of the model training data set formed above.

[0074] In more detail, the above-mentioned preprocessing module mainly includes a noise removal unit and a normalization unit. The noise removal unit is connected to the normalization unit. In the preprocessing module, a secondary verification unit can also be provided. The secondary verification unit can be connected to the normalization unit. The noise frequency elimination operation can be performed in the noise removal unit, the normalization operation can be performed in the normalization unit, and the data can be secondary verified in the secondary verification unit.

[0075] For the training optimization module, its main function is to train the constructed one-dimensional convolutional neural network model and optimize the parameters, and finally obtain a DC arc abnormal fault detection model.

[0076] In the above-mentioned training optimization module, it mainly includes a training unit and a parameter optimization unit. The training unit is connected to the parameter optimization unit. The training unit can train the above-mentioned model. The parameter optimization unit can optimize the model parameters after the training unit is completed, further ensuring that it can obtain the best model.

[0077] The operation module mainly performs the function of deploying and running the above-mentioned model, which can detect in real time whether a DC arc is generated in the photovoltaic system.

[0078] This technical solution mainly includes an acquisition module, a preprocessing module, a training optimization module and an operation module. After the acquisition module collects data and obtains the corresponding model training data set, it performs noise frequency elimination and normalization processing on the corresponding data set in the preprocessing module, and then trains the corresponding DC arc abnormal fault detection model in the training optimization module, and finally performs real-time detection in the operation module.

[0079] Finally, it should be noted that the above embodiments are intended only to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the embodiments, it should be understood by those skilled in the art that modifications or equivalent substitutions to the technical solutions of the present invention do not depart from the spirit and scope of the technical solutions of the present invention and are intended to be encompassed by the claims of the present invention.

Claims

1. A photovoltaic abnormal DC arc detection method, characterized in that: The following steps are involved: S1, collects DC current data of the photovoltaic system under different operating conditions to form a model training data set; S2, performing noise frequency elimination and normalization preprocessing on the model training data set in sequence to obtain a processed model training data set; S3, training the constructed one-dimensional convolutional neural network model based on the processed model training data set to obtain a DC arc abnormal fault detection model based on KAN; S4, deploy and run the DC arc abnormal fault detection model to detect whether a DC arc occurs in the photovoltaic system.

2. A photovoltaic abnormal DC arc detection method according to claim 1, characterized in that: The step S1 includes: building an arc simulation experimental device, collecting current signals with and without arc under different photovoltaic system operating conditions, digitally sampling the current signals, and labeling the current signals sampled each time to form a model training data set classified by arc and no arc.

3. A photovoltaic abnormal DC arc detection method according to claim 1 or 2, characterized in that: The noise frequency elimination specifically includes: Random sampling is performed in the model training data set, and the sampled data is analyzed in the frequency domain to generate a spectrum diagram. The peak frequency in the sampled data spectrum that corresponds to the system switching frequency is selected as the system noise that needs to be eliminated, and a notch filter with the corresponding frequency is used to eliminate the system background noise.

4. A photovoltaic abnormal DC arc detection method according to claim 3, characterized in that: The normalization preprocessing specifically includes: The data after noise frequency elimination is normalized between 0 and 1. After the normalization is completed, the data is verified twice.

5. The photovoltaic abnormal DC arc detection method according to claim 1, characterized in that: The DC arc abnormal fault detection model includes three convolution-KAN layers, each convolution-KAN layer is connected to a maximum pooling layer, and the last pooling layer is connected to two KAN fully connected layers.

6. A photovoltaic abnormal DC arc detection method according to claim 5, characterized in that: The size of a single input data of the DC arc abnormal fault detection model is 64x64; the DC arc abnormal fault detection model can output a 1x2 binary classification probability of arc presence / no arc.

7. A photovoltaic abnormal DC arc detection method according to claim 4, characterized in that: The step S4 includes: deploying the trained DC arc detection model to the photovoltaic system simulation test platform for operation, and inputting the data collected in real time by the photovoltaic system into the DC arc detection model after the steps S1 and S2 to determine whether a DC arc is generated.

8. A photovoltaic abnormal DC arc detection method according to claim 5, characterized in that: The DC arc abnormal fault detection model network adopts a batch training method, with each batch containing 64 data; the training is carried out for 10 training cycles; and the cycle training result with the highest verification accuracy is selected as the final trained model.

9. A photovoltaic abnormal DC arc detection method according to claim 2, characterized in that: The sampling rate of the digital sampling is 250KHz, and the amplitude is 0-4095.

10. A photovoltaic abnormal DC arc detection system, applicable to a photovoltaic abnormal DC arc detection method according to any one of claims 1 to 8, characterized in that: include: Acquisition module, collects DC current data and forms a model training data set; The preprocessing module performs noise frequency elimination and normalization preprocessing on the model training data set; A training optimization module performs model training to obtain a trained DC arc abnormal fault detection model; Run the module, deploy the operation model, and detect in real time whether a DC arc occurs in the system.

Images