Photovoltaic DC fault arc two-stage detection method and photovoltaic DC fault arc model construction method

Abstract

The invention provides a photovoltaic direct-current fault arc two-stage detection method and a model construction method thereof, and the method comprises the steps: firstly building a photovoltaic system direct-current series arc fault experiment platform, collecting data samples of four typical working conditions: a normal working signal, an inverter starting signal, a load change, and a series arc signal; and then the fault arc is input into a constructed arc detection model, the model adopts a two-stage structure, and accurate detection and classification of the fault arc are realized by combining two artificial intelligence technologies of multiple MLP and BLSTM. Wherein the change detection model based on the MLP is used as a trigger for detecting whether the output signal has any noise or distortion, and the BLSTM fully mines the front and back associated information of the time sequence data, so that the understanding ability of the model to the time sequence characteristics is remarkably enhanced. According to the method, the detection speed and accuracy of the direct current series arc fault are remarkably improved, and accurate distinguishing of the specific series arc fault, other possible photovoltaic system faults and background noise is achieved.

Description

Technical Field

[0001] The present invention belongs to the technical field of power system data processing and detection, and particularly relates to a two-stage detection method for photovoltaic DC fault arcs and a method for constructing a model thereof. Background Art

[0002] In recent years, as a typical representative technology of clean energy, photovoltaic power generation has played a key role in the global energy transformation process. However, with the continuous expansion of the installed capacity of photovoltaic power generation systems, the problems exposed during their operation have become increasingly prominent. Among them, DC series fault arcs, due to their strong intermittency, nonlinearity, and concealment, have become one of the main hidden dangers threatening the safe operation of photovoltaic systems. Such fault arcs may not only lead to a decrease in the power generation efficiency of photovoltaic systems but also cause serious accidents such as equipment damage and even fires. Therefore, developing an efficient and reliable fault arc detection method has important theoretical value and practical significance for improving the safety and stability of photovoltaic systems.

[0003] Traditional fault arc detection methods usually rely on spectral analysis, time-domain features, or statistical features, and are identified through rule thresholds or classification models. However, traditional arc fault detection methods usually rely on mathematical analytical operations to process arc signals, and the determinacy of their mathematical expressions also limits the generality of the methods. In contrast, data-driven arc fault diagnosis methods do not require the establishment of an accurate mathematical model and have stronger generalization ability. Patent CN117741370A proposed a scientific windowing method based on time-domain parameters such as autocorrelation coefficient, mean, and variance, combined with Fourier transform for frequency-domain feature extraction, and at the same time extracted high-frequency modal features. Through a weight fusion strategy and a support vector machine model, an accurate detection scheme for series arc faults was finally constructed. Patent CN116150588A proposed an innovative DC arc detection method and device for photovoltaic systems. This method comprehensively uses multi-dimensional feature analysis techniques in the time domain, wavelet domain, and frequency domain to achieve accurate discrimination of fault arc phenomena in photovoltaic systems.

[0004] However, when dealing with the complex and changeable operating environment of photovoltaic systems, the detection accuracy of these methods is easily interfered by factors such as noise and load fluctuations. In addition, the operating data of photovoltaic systems usually has strong temporal correlation and complex nonlinear characteristics. Existing single machine learning models have certain limitations in feature extraction and temporal dependence modeling, and it is difficult to fully mine the deep information in arc fault signals. Summary of the Invention

[0005] In view of the above problems, the present invention proposes a method for constructing a two-stage detection model for photovoltaic DC fault arcs, including the following steps: S1. Based on the established DC series arc fault experimental platform for the photovoltaic system, collect data samples of four typical working conditions, including normal working signals, inverter startup signals, load change signals, and series arc signals, to form a fault arc data set. S2. Construct a two - stage detection model for photovoltaic DC fault arcs, MLP - BLSTM, where the MLP - BLSTM includes an MLP change detection model and a BLSTM classification model. The MLP change detection model is used to detect whether there is any noise or distortion in the output signal. If there is noise or distortion in the signal, the signal will be passed to the BLSTM model for classification according to the category to which the signal belongs. The BLSTM model is improved by combining BRNN and LSTM. The front - end uses CNN for feature extraction, and then the input gate, forget gate, and output gate of the BLSTM model are used to screen data, capture key information in the time series, and finally complete the fault classification of the data through a fully - connected layer. S3. Train and test the MLP - BLSTM model based on the fault arc data set in S1 to obtain the final two - stage detection model for photovoltaic DC fault arcs.

[0006] Preferably, the DC series arc fault experimental platform for the photovoltaic system includes a photovoltaic array, an arc generator, a signal acquisition device, and a test device. The photovoltaic array is composed of several series - connected photovoltaic modules. The arc generator includes an anode, a cathode, an insulating clamp, a fixed base, and an adjustment knob. The anode is fixed, and the cathode is movable. By adjusting the knob, the two electrodes are separated, so that the electric field between the electrodes can break down the air gap and maintain a high - energy discharge, thus generating a stable DC arc. The signal acquisition device uses a current transformer to obtain the current in the circuit and uses an oscilloscope to record experimental data. When sampling, the sampling rate, storage depth, and AC / DC components can be selected through settings. The test device is a load change and a photovoltaic inverter, and the inverter has an MPPT control function.

[0007] Preferably, the MLP change detection model includes an input layer, an output layer, and a hidden layer. Inside each neuron, weighted summation of the input data is achieved through weight connections. The output of the neuron is expressed as: (1) where is a constant, is always 1, and represent the input value and the bias weight respectively. To achieve a non - linear mapping between the input and output time series, the Sigmoid function is used, as shown in formula (2): (2) The neurons of each layer take the output of the previous layer as input, and the signals are transmitted sequentially until they finally reach the output layer.

[0008] Preferably, the BLSTM classification model is composed of a bidirectional recurrent neural network BRNN and LSTM; the forward calculation of BLSTM is represented by the following formula: (3) (4) Where, represents the first feature of the input sequence, represents the input of the h-th LSTM unit at time step t, is the activation function of the h-th unit at time step t, represents the connection weight from the input feature l to the hidden unit h, while represents the connection weight between the hidden unit h and The activation function describes the non-linear transformation of the hidden unit h.

[0009] Preferably, the gradient backpropagation calculation of the BLSTM is represented by the following formula: (5) (6) Where, O represents the model output, represents the weight from the hidden unit h to the output unit k, is the derivative of the activation function.

[0010] Preferably, the front end of the BLSTM model uses CNN for feature extraction. The convolutional layer and max pooling layer of CNN are used to extract high-level features. The data is processed by multiple filter sizes, and the outputs of all pooling layers are concatenated as the input of BLSTM to keep the features in chronological order.

[0011] In the second aspect of the present invention, a two-stage detection method for photovoltaic DC fault arcs is provided. The photovoltaic DC fault arc two-stage detection model constructed by the construction method described in the first aspect is deployed in the detection system, and includes the following processes: Obtain the current signals in the photovoltaic system working conditions in real time, including normal working signals, inverter startup signals, load change signals, and series arc signals; Input the working condition signals into the photovoltaic DC fault arc two-stage detection model for analysis and classification; The MLP change detection model continuously monitors the output signal in real time for any noise or distortion. If noise or distortion is present in the signal, the signal is passed to the BLSTM model for analysis and classification based on the category to which the signal belongs. If the signal is clean and undamaged, the system continues to operate without interruption.

[0012] In a third aspect of the present invention, there is provided a two-stage photovoltaic DC fault arc detection device, which includes at least one processor and at least one memory, and the processor and the memory are coupled to each other. The memory stores a computer executable program of the two-stage photovoltaic DC fault arc detection model constructed by the construction method as described in the first aspect. When the processor executes the computer executable program stored in the memory, the processor executes a two-stage photovoltaic DC fault arc detection method.

[0013] In a fourth aspect of the present invention, there is provided a computer-readable storage medium, which stores a computer program or instruction of the two-stage photovoltaic DC fault arc detection model constructed by the construction method as described in the first aspect. When the program or instruction is executed by a processor, the processor executes a two-stage photovoltaic DC fault arc detection method.

[0014] Compared with the prior art, the present invention has the following beneficial effects: The present invention proposes a two-stage photovoltaic DC fault arc detection method that combines two intelligent algorithms. For the first time, two different machine learning algorithms are used in combination during a two-stage process to detect and classify DC series arc faults. The MLP-based change detection model can effectively identify any abnormal signals and trigger the CNN for further analysis, while also combining the classification ability of the BLSTM model.

[0015] The present invention combines the advantages of the two models to construct an integrated classification model that can make full use of the respective characteristics of the CNN and the BLSTM. This model first extracts as much feature information as possible from the data through the convolutional layer of the CNN, and then uses these features as the input of the BLSTM. This design enables the data to maintain the timing relationship in a bidirectional time series, thereby more effectively capturing the timing characteristics of the arc fault signal.

[0016] In the scenario of the present invention, the RNN is vulnerable to the problem of gradient explosion or vanishing when dealing with long sequences, and its performance is poor. Therefore, the LSTM is introduced, which effectively alleviates the problem of the RNN in long-term dependence modeling through mechanisms such as the input gate, forget gate, and output gate. Although the LSTM and the RNN can utilize the information of the previous state, they can only extract data from the past time steps. Therefore, the BRNN is further improved, and the improved bidirectional structure can process both forward and backward information simultaneously. Description of the Drawings

[0017] Figure 1 This is the overall technical schematic diagram of the photovoltaic DC fault arc two - stage detection method that integrates two intelligent algorithms in the present invention.

[0018] Figure 2 This is the structural diagram of the photovoltaic system DC series arc fault experimental platform provided by the embodiment of the present invention.

[0019] Figure 3 This is the MLP network architecture diagram provided by the embodiment of the present invention.

[0020] Figure 4 This is the BLSTM network structure diagram provided by the embodiment of the present invention.

[0021] Figure 5 This is the operation flow chart of the MLP change detection model provided by the embodiment of the present invention.

[0022] Figure 6 This is the framework diagram of the classification model based on BLSTM provided by the embodiment of the present invention.

[0023] Figure 7 This is the recognition rate curve graph of the arc detection model based on MLP - BLSTM in the test set provided by the embodiment of the present invention.

[0024] Figure 8 This is the loss function curve graph of the arc detection model based on MLP - BLSTM in the test set provided by the embodiment of the present invention.

[0025] Figure 9 This is the structural sketch of the photovoltaic DC fault arc two - stage detection device of the present invention. Detailed implementation manners

[0026] The present invention innovatively proposes a photovoltaic DC fault arc two - stage detection method that integrates two intelligent algorithms. First, a photovoltaic system DC series arc fault experimental platform is built, and data samples of four typical working conditions including normal working signals, inverter startup signals, load changes, and series arc signals are collected systematically and then input into the arc detection model constructed by the present invention. This detection model adopts a two - stage structure, combines two artificial intelligence technologies of MLP and BLSTM, and realizes the accurate detection and classification of fault arcs. Among them, the MLP - based change detection model is used as a trigger to detect whether there is any noise or distortion in the output signal, and BLSTM fully excavates the front - rear correlation information of time - series data, significantly enhancing the model's understanding ability of time - series features.

[0027] To achieve the above - mentioned purpose, the technical process of a photovoltaic DC fault arc two - stage detection method of the present invention includes: S1. Build a photovoltaic system DC series arc fault experimental platform; S2. Collect data samples of four typical working conditions including normal working signals, inverter startup signals, load change signals, and series arc signals to form a fault arc data set; S3. Build a DC series fault arc detection model for a photovoltaic system based on MLP-BLSTM, and train the model using the fault arc data set. The MLP-BLSTM includes a change detection model based on MLP and a classification model based on BLSTM; S4. Use a fault arc data test set to pass through the trained MLP change detection model to detect whether there is any noise or distortion in the output signal; S5. Pass the signal with noise or distortion through the BLSTM classification model to accurately determine the type of noise.

[0028] The following will be described in detail in combination with the specific embodiments and drawings of the present invention. It should be emphasized that the described embodiments of the present invention are only partial examples of the present invention and do not exhaust all possible application scenarios. For those of ordinary skill in the art, other embodiments derived from the technical solutions disclosed by the present invention without creative work shall be regarded as within the protection scope of the present invention.

[0029] Figure 1 The figure is a flowchart of a two-stage detection method for DC fault arcs in a photovoltaic system that integrates two intelligent algorithms. The method includes the following steps: Step S1: Build a DC series arc fault experiment platform for a photovoltaic system to obtain the current signal in the circuit. The acquisition platform includes a photovoltaic array, an arc generator, a signal acquisition device, and a test device; More specifically, as Figure 2 shown, the photovoltaic array is composed of 8 series-connected photovoltaic modules. The arc generator is designed and manufactured based on the UL1699B standard and consists of an anode (a flat copper rod with a diameter of 6 mm), a cathode (a pointed copper rod with a diameter of 6 mm), an insulating clamp, a fixed base, and an adjustment knob. Among them, the anode is fixed, and the cathode is movable. By adjusting the knob, the two electrodes are separated, so that the electric field between the electrodes can break down the air gap and maintain a high-energy discharge, thereby generating a stable DC arc. The signal acquisition device uses a current transformer to obtain the current in the line and uses a ZDS1000 oscilloscope to record the experimental data. The highest sampling rate is 1 GSa / s. The sampling rate, storage depth, and AC / DC components can be selected by setting during sampling. The test device is a load change and a photovoltaic inverter. The inverter has an MPPT control function, and its switching frequency is 20 kHz.

[0030] Step S2: Collect data samples of four typical working conditions including normal working signals, inverter startup signals, load changes, and series arc signals to form a fault arc data set; more specifically, set the experimental platform to the normal working state (no arc generation), set the oscilloscope sampling frequency to 500 kHz, the acquisition time to 280 ms, the wire length to 50 m, start the photovoltaic power supply and load equipment, and turn on the acquisition system after the system stabilizes, record and save the current signal. The acquisition steps for the other three working conditions are similar, only need to adjust the inverter and load settings according to the experimental requirements, and use the adjustment knob to control the arc generation device to stably generate an arc (arc gap is 1 mm). During the experiment, record the current signal under each working condition and save it as a.CSV format file, which can be exported through a USB flash drive.

[0031] Step S3: Construct a DC series fault arc detection model for the photovoltaic system based on MLP-BLSTM, and train the model using the fault arc data set. The MLP-BLSTM includes an MLP change detection model and a BLSTM classification model; More specifically, as Figure 3 shown, the MLP change detection model includes an input layer, an output layer, and a hidden layer. Inside each neuron, weighted summation of the input data is achieved through weight connections. The output of the neuron is expressed as: (1) where, is a constant, is always 1, and represent the input value and the bias weight respectively. To achieve the non-linear mapping between the input and output time series, the Sigmoid function is adopted, as shown in formula (2): (2) The neurons in each layer take the output of the previous layer as the input, and the signal is transmitted sequentially until it finally reaches the output layer. By repeatedly adjusting the weights and biases of the network, MLP can optimize the output until it meets specific performance criteria. During the training process, the error is adjusted through the backpropagation algorithm, gradually reducing the error and optimizing the network parameters. Finally, MLP can accurately learn the mapping relationship between the input and output and assign values to the weights.

[0032] More specifically, as Figure 4As shown, the BLSTM classification model BLSTM consists of a bidirectional recurrent neural network (BRNN) and LSTM. RNN is an important evolution of ANN, designed specifically for processing sequence and time series data, capable of encoding dependencies between inputs. However, RNN is vulnerable to the problem of gradient explosion or vanishing when dealing with longer sequences, resulting in poor performance. Therefore, LSTM is introduced. Through mechanisms such as input gates, forget gates, and output gates, LSTM effectively alleviates the problem of RNN in long-term dependency modeling.

[0033] Although LSTM and RNN can utilize information from previous states, they can only extract data from past time steps. Therefore, BRNN is further improved, and the improved bidirectional structure can process both forward and backward information simultaneously. By combining BRNN and LSTM, BLSTM is formed. The forward calculation of BLSTM can be expressed by the following formula: (3) (4) Where, represents the l-th feature of the input sequence, represents the input to the h-th LSTM cell at time step t, is the activation function of the h-th cell at time step t, represents the connection weight from input feature l to hidden unit h, while represents the connection weight between hidden unit h and The activation function describes the non-linear transformation of hidden unit h.

[0034] The gradient backpropagation calculation of BLSTM is expressed by the following formula: (5) (6) Where, O represents the model output, represents the weight from hidden unit h to output unit k, is the derivative of the activation function. Formulas (5) and (6) detail the calculation process of gradient backpropagation in BLSTM. By combining the advantages of LSTM and BRNN, BLSTM can exhibit stronger modeling capabilities and context capture capabilities when dealing with complex sequence and time series tasks.

[0035] Step S4: Use the trained MLP change detection model with the fault arc data test set to detect whether there is any noise or distortion in the output signal; More specifically, as Figure 5As shown, the MLP change detection model first uses training data containing a large amount of signal datasets. This method is highly compatible with existing classification methods. For the input signal, the system will analyze it. If there is noise or distortion in the signal, the signal will be triggered and passed to the BLSTM model for analysis and classification according to the category to which the signal belongs. If the signal is clean and undamaged, the system will continue to run without interruption. Only the noisy or distorted signals need to pass through the BLSTM classification model to accurately determine the type of noise, significantly improving the efficiency and reliability of the detection system.

[0036] Step S5: The signal containing noise or distortion passes through the BLSTM-based classification model to accurately determine the type of noise; More specifically, as Figure 6 shown, the BLSTM classification model consists of three stages: convolutional neural network (CNN) feature extraction, BLSTM feature learning, and fully connected stage. The convolutional layer and max-pooling layer of the CNN are used to extract high-level features, and the data is processed through various filter sizes. The outputs of all pooling layers are concatenated as the input of the BLSTM to keep the features in chronological order. The BLSTM filters the data through the input gate, forget gate, and output gate to capture the key information in the time series. Finally, these features are fed into the fully connected layer to complete the classification of the data, including types such as series arc, inverter startup, and load change. By combining the efficient feature extraction ability of the CNN with the time series modeling advantage of the BLSTM, the classification model proposed in the present invention can achieve efficient and accurate fault classification in the complex signal environment of the photovoltaic system. The bidirectional data processing and long-term dependence modeling ability of the model further improve its reliability and applicability in practical applications.

[0037] More specifically, to comprehensively evaluate the performance of the MLP-BLSTM algorithm, the five-fold cross-validation method is adopted. The specific method is to randomly divide the data into 5 parts, select one part as the test set each time, and the remaining four parts as the training set for 5 times of training and testing. Finally, the stability and generalization ability of the model are evaluated through the average value of the 5 recognition rates. This method avoids the evaluation bias caused by uneven dataset division or accidental factors, thus obtaining more representative performance indicators. The training accuracy and loss function of the MLP-BLSTM model proposed in the present invention are as Figure 7 and Figure 8 shown. As the number of iterations increases, the accuracy gradually improves and starts to converge at the 40th iteration, finally stabilizing at 99% and reaching up to 100% at most. The loss function shows a downward trend and tends to be stable after about 50 iterations, finally converging to about 0.01. The experimental results show that the photovoltaic DC series fault arc detection method proposed in the present invention shows strong performance in the classification task of arc signals, and can accurately identify, diagnose, and classify DC series arc faults with a high accuracy of 99%.

[0038] As shown Figure 9 in the figure, the present invention also provides a photovoltaic DC fault arc two-stage detection device, which includes at least one processor and at least one memory, and also includes a communication interface and an internal bus; a computer execution program is stored in the memory; a computer execution program of the photovoltaic DC fault arc two-stage detection model constructed by the above-mentioned construction method is stored in the memory; when the processor executes the computer execution program stored in the memory, the processor can execute a photovoltaic DC fault arc two-stage detection method. The internal bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, an Extended Industry Standard Architecture (EISA) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For the convenience of representation, the bus in the drawings of this application is not limited to only one bus or one type of bus. The memory may include a high-speed RAM memory, and may also include a non-volatile storage NVM, such as at least one disk memory, and may also be a USB flash drive, a mobile hard disk, a read-only memory, a magnetic disk or an optical disc, etc.

[0039] The device can be provided as a terminal, a server or other forms of devices. In an exemplary embodiment, the electronic device can be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors or other electronic components to execute the above method.

[0040] The present invention also provides a computer-readable storage medium, in which a computer execution program of the photovoltaic DC fault arc two-stage detection model constructed by the above-mentioned construction method is stored. When the computer execution program is executed by a processor, the processor can execute a photovoltaic DC fault arc two-stage detection method.

[0041] Specifically, a system, a device or an equipment equipped with a readable storage medium can be provided. A software program code for implementing the functions of any one of the above embodiments is stored on the readable storage medium, and the computer or processor of the system, the device or the equipment is made to read and execute the instructions stored in the readable storage medium. In this case, the program code read from the readable medium itself can implement the functions of any one of the above embodiments. Therefore, the machine-readable code and the readable storage medium storing the machine-readable code constitute a part of the present invention.

[0042] The above are only the preferred embodiments of the present application and are not intended to limit the present application. For those skilled in the art, various modifications and changes can be made to the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included within the protection scope of the present application.

[0043] Although the specific implementation manners of the present invention have been described above, they do not limit the protection scope of the present invention. Those skilled in the art should understand that various modifications or deformations that can be made without creative efforts on the basis of the technical solution of the present invention are still within the protection scope of the present invention.

Claims

1. A method for constructing a two-level detection model for photovoltaic DC fault arc, characterized in that: The following steps are involved: S1, based on the built photovoltaic system DC series arc fault experimental platform, collects data samples of four typical working conditions including normal working signal, inverter start-up signal, load change signal, and series arc signal to form a fault arc data set; S2, constructing a photovoltaic DC fault arc two-level detection model MLP-BLSTM, wherein the MLP-BLSTM includes an MLP change detection model and a BLSTM classification model; The MLP change detection model is used to detect whether the output signal has any noise or distortion. If there is noise or distortion in the signal, a trigger signal is passed to the BLSTM model to classify the signal according to the category it belongs to. The BLSTM model is improved by combining BRNN and LSTM. The front end uses CNN for feature extraction, and then the data is filtered through the input gate, forget gate and output gate of the BLSTM model to capture the key information in the time series. Finally, the fault classification of the data is completed through the fully connected layer. S3, based on the fault arc dataset in S1, the MLP-BLSTM model is trained and tested to obtain the final photovoltaic DC fault arc two-level detection model.

2. A photovoltaic DC fault arc two-stage detection model construction method as claimed in claim 1, characterized in that: The photovoltaic system DC series arc fault experimental platform includes a photovoltaic array, an arc generator, a signal acquisition device and a test device; The photovoltaic array is composed of several photovoltaic modules connected in series, and the arc generator includes an anode, a cathode, an insulating clamp, a fixed base and an adjustment knob; the anode is fixed and the cathode is movable, and the two electrodes are pulled apart by adjusting the knob so that the electric field between the electrodes can break through the air gap and maintain high-energy discharge, thereby generating a stable DC arc; the signal acquisition device uses a current transformer to obtain the current in the circuit, and uses an oscilloscope to record experimental data. When sampling, the sampling rate, storage depth and AC / DC components can be selected by setting; the test equipment is a load change and photovoltaic inverter, and the inverter has an MPPT control function.

3. A photovoltaic DC fault arc two-stage detection model construction method as claimed in claim 1, characterized in that: The MLP change detection model includes an input layer, an output layer, and a hidden layer. Each neuron is connected by weights to achieve the weighted summation of the input data. The output of the neuron is expressed as: （1） in, is a constant, Always 1, and Represent the input value and bias weight respectively; in order to realize the nonlinear mapping between input and output time series, the Sigmoid function is used, as shown in formula (2): （2） The neurons in each layer take the output of the previous layer as input, and the signal is passed sequentially until it finally reaches the output layer.

4. A photovoltaic DC fault arc two-stage detection model construction method as claimed in claim 1, characterized in that: The BLSTM classification model consists of a bidirectional recurrent neural network BRNN and LSTM; the forward calculation of BLSTM is expressed by the following formula: （3） （4） in, represents the first feature of the input sequence, represents the input of the h-th LSTM unit at time step t, is the activation function of the h-th unit at time step t, represents the connection weight from the input feature l to the hidden unit h, and represents the hidden unit h and The connection weights between Describes the nonlinear transformation of the hidden unit h.

5. A photovoltaic DC fault arc two-stage detection model construction method as claimed in claim 4, characterized in that: The gradient back calculation of the BLSTM is expressed by the following formula: （5） （6） Among them, O represents the model output, represents the weight from hidden unit h to output unit k, K and H represent the maximum hidden unit and the maximum output unit respectively, is the derivative of the activation function.

6. A photovoltaic DC fault arc two-stage detection model construction method as claimed in claim 1, characterized in that: The BLSTM model front end uses CNN for feature extraction, uses the convolution layer and maximum pooling layer of CNN to extract high-level features, processes data through multiple filter sizes, and concatenates the outputs of all pooling layers as the input of BLSTM to keep the features in time order.

7. A two-stage detection method for photovoltaic DC fault arc, characterized in that: The photovoltaic DC fault arc two-level detection model constructed by the construction method described in any one of claims 1 to 6 is deployed in a detection system, and includes the following processes: Real-time acquisition of current signals in photovoltaic system working conditions, including normal working signals, inverter start-up signals, load change signals, and series arc signals; The operating condition signal is input into the photovoltaic DC fault arc two-stage detection model for analysis and classification; The MLP change detection model detects in real time whether the output signal has any noise or distortion; if there is noise or distortion in the signal, a trigger signal is passed to the BLSTM model for analysis and classification based on the category the signal belongs to; if the signal is clean and uncorrupted, the system continues to operate uninterrupted.

8. A photovoltaic DC fault arc two-stage detection device, characterized in that: The device includes at least one processor and at least one memory, the processor and the memory are coupled; the memory stores a computer execution program of a photovoltaic DC fault arc two-level detection model constructed by the construction method according to any one of claims 1 to 6; when the processor executes the computer execution program stored in the memory, the processor executes a photovoltaic DC fault arc two-level detection method.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program or instruction of a photovoltaic DC fault arc two-level detection model constructed by the construction method according to any one of claims 1 to 6, and when the program or instruction is executed by a processor, the processor executes a photovoltaic DC fault arc two-level detection method.

Images