A method for detecting DC arc in photovoltaic system

Abstract

The present invention discloses a method for detecting direct current arcing of a photovoltaic system. Test data of the entire time-series continuous process from normal operation to arcing of the photovoltaic system are collected. The data reflect the change trend of the system before and after arcing. The change trend remains basically consistent under the influence of different external environments and working conditions, thereby achieving the effect of reducing the influence of external interference and realizing the portability of the detection scheme under different environments and working conditions. In addition, the present invention proposes an arcing detection framework based on time-frequency information on the basis of obtaining continuous data. On the one hand, the non-stationary characteristics of arcing spectrum features in time series are additionally considered, thereby providing a richer detection basis for arcing detection. On the other hand, a processing flow and a model combination architecture specifically for time-frequency domain combined information are proposed. First, time-frequency information is obtained based on a time-frequency analysis method, and then two cascaded models are used to serially process frequency domain features and time-series changes of frequency domain features, thereby finally obtaining a highly accurate arcing detection result.

Description

Technical Field

[0001] The present invention relates to the field of new energy technology, and in particular to a method for detecting direct current arcing in a photovoltaic system. Background Art

[0002] In the prior art, the methods for detecting arc faults can generally be divided into two categories: detection based on the electrical and physical characteristics of the arc, heat, noise, electromagnetic radiation, etc., and detection based on arc current or voltage data. However, for photovoltaic systems, the photovoltaic inverter itself is a switching power supply, which will cause frequency interference to the current signal, and the field environment is very complex, which is prone to various uncertain noise signals. These signals will not only interfere with the arc signal, but some signals may also be very similar to the arc signal, which increases the difficulty of detection. Detection only through the time domain or frequency domain characteristics of the arc current and / or voltage is still prone to misjudgment.

[0003] In addition, in the prior art,

[0004] Patent CN117148049B discloses a DC arc fault detection method, which collects DC current and calculates the corrected proportional coefficient based on its value, and then puts the corrected high-frequency AC signal into a neural network based on FFT for arc detection and judgment. This processing only optimizes the influence of the absolute value of the current, and the final detection is only based on the FFT result, with a single criterion.

[0005] Patent CN104092440B discloses a method, device, processor and system for detecting DC arc faults in a photovoltaic system. The method only uses the amplitude characteristics of discrete points of the current spectrum, has a single criterion and low detection accuracy. In order to improve the accuracy, six thresholds are set for judgment, but the operation complexity is increased. Only a small number of operating parameters are used to set different thresholds, and the adaptability to complex working conditions is not strong.

[0006] Patent CN103913663B discloses an online detection method and protection device for arc faults in a DC system, which performs targeted processing for different DC system working conditions to improve the recognition rate under various working conditions. Although the method takes into account the noise impact of different working conditions, it requires different calculations and processing for different noises, such as filtering out different switching frequency points, and also requires setting different thresholds. Not only is the calculation process of the feature quantity complicated, but the portability of different scenarios is also relatively poor.

[0007] Patent CN107340459B discloses a DC fault arc detection method and system, which combines time domain and frequency domain features, but processes and detects the time domain and frequency domain features separately, and the fast Fourier transform is only applicable to the processing of stationary signals, while the arcing characteristic signal is a non-stationary signal. Using only the result of the fast Fourier transform as the feature of the neural network to detect arcing is prone to misjudgment; the influence of different environments and working conditions is not considered, and the adaptability is not strong.

[0008] Patent CN107086855B discloses a method for detecting fault arcs in photovoltaic systems by integrating machine learning with multiple time-frequency features. The method combines time-frequency features into a machine learning model, but requires two transformations and integral operations during data processing, which takes a long time to calculate. At the same time, the hidden Markov model assumes that the state space is finite, and its limited expression ability cannot adapt well to the very complex photovoltaic system operating environment and different working conditions, resulting in weak detection accuracy and transferability.

[0009] In summary, the prior art has certain defects / problems. The present invention proposes a method for detecting DC arcing in a photovoltaic system to solve the defects / problems of the prior art. Summary of the invention

[0010] The purpose of the present invention is to solve the shortcomings of the prior art and to provide a method for detecting DC arcing in a photovoltaic system.

[0011] In order to achieve the above object, the present invention adopts the following technical solutions:

[0012] A method for detecting DC arcing in a photovoltaic system, comprising the following steps:

[0013] Step 1: Training data collection and processing;

[0014] Step 1.1: Collect the physical quantities related to arc detection in the continuous process from no arc to arc, perform analog-to-digital conversion and then divide the frames to obtain a discrete training data set;

[0015] Step 1.2: Filter the training data in step 1.1 to obtain a multi-frame signal that is continuous in time sequence;

[0016] Step 1.3: transform the multi-frame signal in step 1.2 to obtain the corresponding time-frequency spectrum;

[0017] Step 2: DC arc detection model training;

[0018] Step 2.1: Use the frequency domain feature network to extract frequency domain features;

[0019] Step 2.2: Use the time domain feature network or timing model to extract the timing changes of the frequency domain features.

[0020] Step 3: DC arc online detection.

[0021] Furthermore, the physical quantity is current or voltage.

[0022] Furthermore, the current refers to the AC component in the DC current.

[0023] Furthermore, the filtering adopts a threshold filtering method based on histogram statistics.

[0024] Furthermore, the step 3 comprises:

[0025] Step 3.2: Collect real-time data, which is a physical quantity related to arc detection;

[0026] Step 3.3: Filter the real-time data and calculate the time-frequency spectrum;

[0027] Step 3.4: Input the time-frequency spectrum into the trained DC arc detection model to calculate the arc probability of real-time data.

[0028] Furthermore, the step 3 also includes step 3.1, and the step 3.1 is before step 3.2, and the step 3.1 is to deploy the trained DC arc detection model to the embedded platform.

[0029] Furthermore, the transformation is short-time Fourier transform or wavelet transform.

[0030] Furthermore, the frequency domain feature network is a convolutional neural network or a feedforward neural network.

[0031] Furthermore, the time domain feature network is a long short-term memory network, or a simple recurrent neural network, or a gated recurrent unit.

[0032] Furthermore, the convolutional neural network is a one-dimensional convolutional neural network or a two-dimensional convolutional neural network.

[0033] Furthermore, the step 3.4 includes making an arcing judgment on each frame of real-time data, and determining that the current frame is arcing when the arcing probability of the current frame is greater than a first threshold, and determining that the current frame is operating normally when the arcing probability of the current frame is less than or equal to the first threshold; when the number of arcing frames is greater than a second threshold, it is determined that the photovoltaic system is arcing.

[0034] Furthermore, the first threshold range is [50%, 60%], and the second threshold range is [N / 2, n], wherein N is the number of detection frames, n is a positive integer, and the value range of n is (N / 2, N].

[0035] Compared with the prior art, the method for detecting DC arc in a photovoltaic system provided by the present invention has the following advantages:

[0036] The photovoltaic system DC arcing detection method of the present invention collects test data of the entire time-series continuous process from normal operation to arcing of the photovoltaic system. This data reflects the change trend of the system before and after arcing. The change trend remains basically consistent under the influence of different external environments and working conditions, thereby achieving the effect of reducing the influence of external interference and realizing the portability of the detection scheme under different environments and working conditions.

[0037] On the basis of acquiring continuous data, the present invention proposes an arc detection framework based on time-frequency information. On the one hand, it additionally considers the non-stationary characteristics of the arc spectrum characteristics in time series, providing a richer detection basis for arc detection; on the other hand, it proposes a processing flow and model combination architecture specifically for the combined information of time-frequency domain. First, the time-frequency information is acquired based on the time-frequency analysis method, and then two cascaded models are used to serially process the frequency domain characteristics and the time series changes of the frequency domain characteristics, and finally a highly accurate arc detection result is obtained. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the specific implementation methods of the present invention or the technical solutions in the prior art, the drawings required for use in the specific implementation methods or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some implementation methods of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0039] Figure 1 is a flow chart of a method for detecting DC arcing in a photovoltaic system according to Embodiment 1 of the present invention;

[0040] Figure 2 It is a framework diagram of the time-frequency arc detection model of the first embodiment provided by the present invention;

[0041] Figure 3 is a schematic diagram for explaining continuous data in time sequence according to the first embodiment of the present invention;

[0042] Figure 4 is a flowchart of a histogram filtering algorithm according to the first embodiment of the present invention;

[0043] Figure 5 is a short-time Fourier transform flow chart of the first embodiment provided by the present invention;

[0044] Figure 6 Schematic diagram of the structure of a single LSTM unit according to the first embodiment of the present invention;

[0045] Figure 7 This is a schematic diagram of the network structure of the combined arc detection model of the first embodiment provided by the present invention;

[0046] Figure 8 is a flow chart of online detection threshold determination according to the first embodiment of the present invention;

[0047] Fig. 9 It is an overall flow chart of the first embodiment provided by the present invention. DETAILED DESCRIPTION

[0048] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention. Example

[0049] See also Figure 1 - Fig. 9 The present invention provides a method for detecting DC arc in a photovoltaic system, which specifically comprises the following steps:

[0050] Step 1: Training data collection and processing.

[0051] Step 1.1: Collect the physical quantities related to arc detection in the continuous process from no arc to arc, perform analog-to-digital conversion and then divide the frames to obtain a discrete training data set;

[0052] The AC component of the DC current of the continuous process of the photovoltaic system from the normal operating state to the arcing state under different arcing positions and different external working conditions in an operating environment is collected at a fixed sampling frequency FS (in other embodiments, the AC component of the DC current can be replaced by other physical quantities related to arcing detection, such as voltage, other forms of current, etc.), and the same condition can be repeatedly collected 3 to 6 times (or other values, such as 4 to 7 times, 5 to 8 times, 10 to 12 times, etc.) to eliminate sampling errors; wherein the same condition refers to the same environment, arcing position and external working condition. The sampling signal is converted to a digital signal after analog-to-digital conversion (ADC), and then framed to obtain a discrete training data set; the arcing position includes the DC input end of the photovoltaic inverter and between photovoltaic panels; the external working conditions include whether the photovoltaic system is connected to an external power optimizer, the current size of different photovoltaic strings, etc.; a special arcing fixture is required for data collection to realize artificial control of arc generation and automatic data marking.

[0053] In one embodiment, the signal sampling, analog-to-digital conversion, and framing process are as follows:

[0054] The inverter is connected to the data acquisition module and the arc-drawing fixture. The inverter is powered on and operates normally. Use the host computer software to click the "Start Acquisition" button to start acquisition. After waiting for a period of time (for example, 3S), use the arc-drawing fixture to draw the arc. At this time, it is in the arc-drawing stage; continue to wait for a period of time (for example, 3S), and click the "Stop Acquisition" button on the host computer to stop signal acquisition; the collected signal forms a complete analog signal data from normal operation to arcing. During the acquisition process, the acquisition module will convert the collected analog signal data (voltage or current) into digital signals and save them to storage media such as computers. The label of whether arcing is occurring is also saved synchronously; then this complete digital signal is framed according to the required window length for subsequent processing.

[0055] Different environments or working conditions have different effects on the noise of arc current, and the impact on the value of arc characteristics is also unknown. However, the difference and change trend between the normal operation data and the arc data under the same environment are relatively unchanged; Figure 3 As shown in the figure, the horizontal axis (Time) is time, and the vertical axis (Value) is the characteristic value of the data (for example, the AC component of the DC current signal, DC voltage, voltage, etc.), or the time series change of the frequency domain characteristics, or the change of the frequency domain characteristics of the data in time series; the vertical axis is divided into two intervals on the left and right, the left interval is the no arc interval, or the normal operation interval, and the right interval is the arc interval; in the process from normal operation to arcing, 4 data are collected in the normal operation stage (the AC component of the DC current ） , 4 data were collected during the arcing stage (the AC component of the DC current ）； The red dot indicates arcing, the black dot indicates no arcing, and the three lines of different colors (blue, black, and green) Data A, DataB, and DataC indicate that the data collected under different environments or working conditions are continuous in time series (actually discrete data quantities, not continuous analog quantities). Judging from the values ​​of each point alone, the impact of noise on the red point may be higher or lower than the impact on the black point, but the trend of each data from normal to arcing remains unchanged, but the degree of distinction will change to a certain extent.

[0056] Figure 3 In the embodiment, the number of data collected is 8; in other embodiments, other numbers of data may be collected; for example, 10 data are collected in total, 5 data are collected in the normal operation stage and the arc drawing stage; another example is 20 data are collected in total, 10 data are collected in the normal operation stage and the arc drawing stage; another example is 30 data are collected in total, 15 data are collected in the normal operation stage and the arc drawing stage; and so on. Generally speaking, the more data are collected, the more accurate the model training is, but the timeliness factor should be considered, and the value should not be too large. 。

[0057] Different from the arc detection data used in other arc detection schemes, the collected continuous data reflects the changing trend of the DC side current of the photovoltaic system from normal operation to arcing. By detecting the change of trend rather than the absolute state quantity, the influence of external working conditions and environment is weakened, and the adaptability of the detection model and the migration ability in different environments are enhanced. That is, only one site data needs to be collected to train the model, and the model can be directly migrated to other sites without additional changes. In addition, in the experiment, the collection time of each frame of data in the online detection stage is equal to the processing time of each frame of data and the arc detection calculation time, so as to ensure that the data in the online detection stage is consistent with the training stage, that is, each piece of data is completely continuous, further improving the accuracy and speed of the model in the online detection stage.

[0058] Step 1.2: Filter the training data in step 1.1 to obtain a multi-frame signal that is continuous in time sequence;

[0059] First, each frame of current data is filtered in the time domain to reduce the interference of burr signals (interference between the on-site environment or machines and wiring will cause the current ripple to become larger, which is reflected in the current signal as an obvious burr signal, and in the spectrum as a harmonic component in a frequency band similar to the arc characteristic frequency; these noises are very common and will greatly disrupt the accuracy of arc detection, so they need to be filtered out) to obtain a multi-frame current signal that is continuous in time sequence; Figure 4 For threshold filtering based on histogram statistics, the current ADC distribution histogram is calculated, sorted by data size, the proportion of glitch signals is obtained according to the histogram statistics of the current ADC data, the filtering threshold is determined, and truncated according to the threshold (truncated if exceeding the threshold, retained if below or equal to the threshold); this filtering method has clear and simple ideas, is easy to implement, and runs fast.

[0060] In other embodiments, when the physical quantity in step 1.1 is voltage, the above current data is correspondingly replaced with voltage data, and the processing method is similar to that of the current data.

[0061] Step 1.3: transform the multi-frame signal in step 1.2 to obtain the corresponding time-frequency spectrum;

[0062] like Figure 5As shown in the figure, short-time Fourier transform is performed on each frame of current signal after filtering to obtain a time-frequency spectrum, and the time-frequency spectrum is taken as the feature quantity of subsequent model detection; the basic idea of ​​short-time Fourier transform is to window and truncate the signal, so that this segment of the signal within a time window is approximately stable, and then Fourier transform is performed on each segment of the approximately stable signal, and finally the spectrum of each segmented signal is obtained, and the spectrum is stacked up in chronological order to obtain the time-frequency spectrum of the original signal; the time-frequency spectrum contains both frequency information (frequency domain features) and time series information (the time series changes of frequency domain features, or the changes of frequency information over time), which provides a richer, more accurate and reliable detection basis for subsequent arc detection. Among them, the frequency domain feature refers to the physical quantity detected during the period from no arc to arcing (such as the AC component of DC current, voltage, etc.), which is transformed by FFT to obtain the spectrum of current / voltage during this period; the time series change of the frequency domain feature refers to the current / voltage data within this period being windowed and truncated, for example, truncated into current 1, current 2, current 3, etc., and then each current (current 1, current 2, current 3, etc.) is transformed by FFT to obtain the spectrum, and then stacked up in time order, so that the time series change of the frequency domain feature can be obtained.

[0063] Figure 5 In the invention, the original signal refers to the signal to be short-time Fourier transformed, which is a frame of current signal after frame filtering in the present invention; the window function is a basic concept in short-time Fourier transform, which is used to truncate the original signal, that is, multiply the window function by the original signal sequence to obtain a section of signal within a fixed time window. Commonly used window functions include Hamming window, rectangular window, etc.; windowing refers to sliding multiplication of the window function by the original signal sequence to obtain a series of signals within fixed time windows.

[0064] It can be understood that when the physical quantity in step 1.1 is voltage, the above current signal is correspondingly replaced by a voltage signal.

[0065] Determine whether each frame of current signal has an arcing label, where 0 indicates that there is no arcing fault in the current frame, and 1 indicates that there is an arcing fault in the current frame;

[0066] At this point, the collection and processing of training data is completed.

[0067] Step 2: DC arc detection model training.

[0068] Construct a CNN and LSTM combined model to extract the time-frequency characteristics of the current signal; specifically:

[0069] Step 2.1: Use the frequency domain feature network to extract frequency domain features.

[0070] Specifically, Figure 7As shown in the figure, in order to process the time series data of the current signal, a time-distributed 2-dimensional convolution neural network (TD-2DCNN) is used. The input layer is used to input data. The input data is the time-frequency spectrum of the current signal, which includes two dimensions of time and frequency. The time dimension is retained and the frequency is split into two dimensions. The time-distributed 2-dimensional convolution neural network will fix the time dimension and use the convolution kernel with the same parameters to extract features for the frequency vector of each time step; TD-2DCNN contains two convolution layers and one pooling layer. The two convolution layers are the time-distributed convolution layer 1 and the time-distributed convolution layer 2. The time-distributed convolution layer 1 contains 32 convolution kernels (filter kernels) and the time-distributed convolution layer 2 contains 16 convolution kernels. The convolution kernels are all 3×3 convolution kernels, and the activation function is relu; the global maximum pooling operation is performed through the global maximum pooling layer 1 to extract the global features, and its output is used as the input of the long short-term memory network (LSTM).

[0071] In other embodiments, when the physical quantity in step 1.1 is voltage, the input data of the input layer correspondingly becomes a time-frequency spectrum of a voltage signal, including two dimensions of time and frequency.

[0072] In other embodiments, the number of convolution kernels, the size of the convolution kernel, and the type of activation function can be different from Figure 7 The embodiments are different; for example, in other embodiments, the number of convolution kernels can be 4, 8, 64, 128, etc., the size of the convolution kernel can be 5×5, 7×7, etc.; the activation function type can be tanh, sigmoid, leakyrelu, etc.

[0073] Step 2.2: Use the time domain feature network to extract the temporal changes of frequency domain features.

[0074] Please continue reading Figure 7The long short-term memory network (LSTM) processes the frequency domain features extracted by the aforementioned two-dimensional convolutional neural network (TD-2DCNN) in the time dimension. The long short-term memory network (LSTM) contains two bidirectional LSTM layers, namely bidirectional LSTM layer 1 and bidirectional LSTM layer 2. Bidirectional LSTM layer 1 and bidirectional LSTM layer 2 contain 32 and 16 LSTM units respectively, and the activation function is relu; a layer of random dropout is added between the two bidirectional LSTM layers. The random dropout operation will randomly close some of the first layer LSTM units. The output of the element (bidirectional LSTM layer 1) is used to reduce the risk of overfitting and enhance the generalization ability of the network; for sequence data, layer normalization is performed after the convolution and LSTM layers to improve network stability and accelerate model convergence; the output layer is a fully connected layer with two neurons, and the activation function is Softmax, which is used for binary classification tasks; TD-2DCNN and LSTM are combined for end-to-end training to obtain a trained DC arc detection model; the whole process is convenient and simple, and the combined model takes into account both time-frequency and frequency domain characteristics, with high model accuracy and good detection reliability.

[0075] In other embodiments, the number of LSTM units may be 4, 8, 64, 128, etc.; the activation function type may also be different, for example, the activation function type may be tanh, sigmoid, leakyrelu, etc.

[0076] In summary, the detection model uses convolutional neural network (CNN) and long short-term memory network (LSTM) ） The target task is a binary classification task of whether or not an arc is generated. The time-frequency spectrum of the current signal is used as the model input, and the arc probability is the model output. A network framework of a combined model of CNN and LSTM is constructed, in which CNN is used to extract the features of the frequency dimension of the current signal (frequency domain features), and LSTM is used to extract the features of the time dimension of the current signal (time domain features). Finally, the frequency domain features and time domain features (the temporal changes of the frequency domain features) of the current signal are combined, and the arc probability calculation is realized through a fully connected layer, thereby completing the training of the DC arc detection model.

[0077] Among them, convolutional neural networks are generally composed of network layers such as convolution and pooling, and are mainly used to process more complex data with more implicit features such as images or sequences; convolution uses convolution kernels to calculate information within the spatial neighborhood of the data to extract data features. Convolution kernels of different sizes have different receptive fields and different types of features extracted; pooling downsamples the data, which is also a further extraction and screening of features, reducing the amount of calculation while better selecting useful features; activation functions are also used in the network layer to introduce nonlinearity and perform nonlinear transformations on the input information, so that the neural network can arbitrarily approximate any nonlinear function.

[0078] Among them, Figure 6 As shown, the long short-term memory network (LSTM ） It is a type of recurrent neural network suitable for processing sequence data. Each network layer consists of a stack of LSTM units as shown in the figure below. The input of each unit is the feature vector x at the current moment. t and the state vector c at the previous moment t-1 And the hidden state vector h t-1 , the output is the current state vector c t And the hidden state vector h t , through three gates, different degrees of retention and forgetting of input information can be achieved; the sequentially stacked LSTM units can process information at different times to continuously update the intermediate state and obtain the final output.

[0079] Step 3: DC arc online detection.

[0080] Step 3.1: Deploy the trained DC arc detection model to the embedded platform;

[0081] In one embodiment, the model (trained DC arc detection model) is optimized before deployment to reduce the model size and improve the reasoning speed. After the optimized model is deployed to the embedded platform, real-time data can be collected for online arc detection. In some embodiments, the model does not need to be optimized and can be directly deployed to the embedded platform.

[0082] Step 3.2: Collect real-time data;

[0083] In the online detection stage, real-time data is acquired frame by frame to ensure that the single-frame data acquisition time is equal to the single-frame data processing and detection calculation time to ensure the continuity of real-time data. Among them, real-time data is a physical quantity related to arc detection, such as the AC component in the DC current, voltage, etc.; it can be understood that the physical quantity here is the same as the physical quantity in step 1.1, that is, if the physical quantity in the previous step 1.1 is the AC component in the DC current, the real-time data in step 3.2 is also the AC component in the DC current; if the physical quantity in the previous step 1.1 is the voltage, the real-time data in step 3.2 is also the voltage.

[0084] Step 3.3: Filter the real-time data and calculate the time-frequency spectrum;

[0085] The processing flow of real-time data is similar to that of the training data in the aforementioned step 1. For example, the current frame data is first subjected to threshold filtering based on histogram statistics, and then a short-time Fourier transform is performed to calculate the time-frequency spectrum.

[0086] Step 3.4: Input the time-frequency spectrum into the trained DC arc detection model to calculate the arc probability of real-time data;

[0087] like Figure 8 As shown, the time-frequency spectrum of real-time data is put into the trained DC arc detection model to calculate the arc probability. If the arc probability of the current frame is greater than the first threshold (ProbThresh), the current frame is determined to be an arc, and the arc of the current frame is set to 1; if the arc probability of the current frame is less than or equal to the first threshold (ProbThresh), the current frame is determined to be operating normally, and the arc of the current frame is set to 0; when the number of arc frames in continuous detection frames (for example, 20 frames, or other values) is greater than the second threshold (CountThresh), it is determined to be an arc in the photovoltaic system, and an arc fault is reported. In one implementation, a protection operation is also performed. Among them, the first threshold range can be [50%, 60%], and can also be selected as [50%, 100%] according to the required range). The second threshold is the ratio of the number of detection frames. For example, if the number of detection frames is 20, the second threshold range may be [10, 19], [10, 15], or [12, 15], and so on. In one embodiment, the number of detection frames is a positive integer N, and the second threshold range is [N / 2, n], where n is a positive integer and the value range of n is (N / 2, N].

[0088] It can be understood that in step 3.4, the number of detection frames may be the same as or different from the number of detection frames in step 1.2.

[0089] In other embodiments, the aforementioned step 3.1 is not necessary. Example

[0090] The second embodiment is substantially the same as the first embodiment, except that:

[0091] The data used in the detection scheme in step 1.1 of Example 1 is the AC component of the continuous DC current from normal operation to arcing. The continuous data and / or its selection method in Example 2 may be different, such as the length of the continuous time (i.e. the length of the collection time, such as 4S, 8S, 10S, etc.), the collection of continuous voltage data or DC current data, and other physical quantities that may be related to arcing detection. Example

[0092] The third embodiment is substantially the same as the first embodiment, except that:

[0093] The original data is analyzed in the time and frequency domain by using time and frequency domain analysis methods such as wavelet transform to analyze the non-stationary signal and obtain the time-frequency spectrum. Example

[0094] The fourth embodiment is substantially the same as the first embodiment, except that:

[0095] When extracting frequency domain features, other networks such as one-dimensional convolutional neural networks, feedforward neural networks or other feature extraction methods are used; among them, compared with two-dimensional convolutional neural networks, one-dimensional convolutional neural networks do not need to split the frequency into two dimensions. Example

[0096] The fifth embodiment is substantially the same as the first embodiment, except that:

[0097] Other recurrent neural networks or timing models are used to extract time domain features, such as simple recurrent neural networks, gated recurrent units, etc. Example

[0098] The sixth embodiment is substantially the same as the first embodiment, except that:

[0099] Embodiment 1 adopts the specific structure of the combined network of convolutional neural network and long short-term memory network, including the number of hidden layers, activation function, random inactivation threshold, etc., which have many different combinations. Embodiment 6 can also adopt other network structures to achieve similar effects.

[0100] In the embodiment of the present invention, it can be understood that the physical quantity in step 1.1 corresponds to the data in the subsequent steps. For example, if the physical quantity in step 1.1 is current, the data in the subsequent steps are all corresponding current data; if the physical quantity in step 1.1 is voltage, the data in the subsequent steps are all corresponding voltage data.

[0101] The DC arcing detection method of the present invention collects the AC component data of the DC current signal of the entire time-series continuous process from normal operation to arcing of the photovoltaic system. This data reflects the change trend of the system before and after arcing. The change trend remains basically consistent under the influence of different external environments and working conditions, thereby achieving the effect of reducing the influence of external interference and realizing the portability of the detection scheme under different environments and working conditions.

[0102] The present invention proposes an arc detection framework based on time-frequency information on the basis of obtaining continuous data. On the one hand, it additionally considers the non-stationary characteristics of the arc spectrum characteristics in time series, providing a richer detection basis for arc detection; on the other hand, it proposes a processing flow and model combination architecture specifically for the combined information of time-frequency domain. First, the time-frequency information is obtained based on the time-frequency analysis method, and then two cascaded models are used to serially process the time domain and frequency domain characteristics, and finally a highly accurate arc detection result is obtained. The present invention detects the trend of data changes and only needs to collect data from one field to train the model to adapt to different photovoltaic fields, thus achieving self-adaptation to different field environments and improving the transferability of the model; on this basis, through the application of a combined neural network model with strong expression ability, the high accuracy and strong reliability of DC arc detection in photovoltaic systems are achieved.

[0103] The following are some of the terms involved in the present invention:

[0104] DC arcing: DC arcing refers to the arcing phenomenon formed in the air or vacuum in a DC circuit due to excessive electric fields between conductors or other factors. DC arcing may cause serious consequences such as damage to electrical equipment and fire, so arc detection is very important in DC applications such as photovoltaic systems.

[0105] Fourier transform: Fourier transform is a mathematical transform used to convert signals from the time domain to the frequency domain. It decomposes the original signal into a combination of sine waves and cosine waves of different frequencies, thereby analyzing the spectral characteristics of the signal. Fourier transform is widely used in signal processing, image processing and other fields.

[0106] Short-time Fourier transform: Short-time Fourier transform is a method of Fourier transforming a signal in different time windows to obtain the time-frequency representation of the signal. STFT can analyze the local characteristics of the signal in time and frequency and is suitable for processing non-stationary signals.

[0107] Time-spectrum: The time-spectrum is a graph that describes the changes of a signal in two dimensions, time and frequency, and is usually obtained through methods such as short-time Fourier transform. The time-spectrum provides the time-frequency distribution of a signal and can reveal the frequency components of the signal at different time points.

[0108] Harmonics: Harmonics are components in a signal whose frequencies are integer multiples of the fundamental frequency. Harmonics often occur in nonlinear systems or devices (such as switching power supplies) and may cause power quality problems.

[0109] Switching power supply: A switching power supply is a power supply that controls the current switching state through electronic switching elements (such as transistors and MOSFETs) to achieve power conversion and regulation. Switching power supplies have the advantages of high efficiency and small size, and are widely used in various electronic devices.

[0110] Glitch signal: Glitch signal refers to a short, sudden abnormal signal in the signal, usually manifested as a sharp pulse. Glitch signal may be caused by external interference, equipment failure, etc., which will affect signal processing and system stability.

[0111] Histogram statistics: Histogram statistics is a method of describing data characteristics by the frequency distribution of statistical data in different intervals. Histograms can reveal the distribution pattern, central tendency and dispersion of data, and are a commonly used tool in data analysis.

[0112] Convolutional Neural Network: Convolutional Neural Network is a deep learning model that is particularly good at processing image data. CNN extracts local features of the image through convolutional layers, reduces feature dimensions through pooling layers, and performs classification or regression tasks through fully connected layers. CNN has significant effects in image recognition, object detection and other fields.

[0113] Recurrent Neural Network: Recurrent Neural Network is a deep learning model suitable for processing sequence data. RNN transfers information between time steps through a cyclic structure, which can capture dependencies in the sequence and is widely used in natural language processing, time series prediction and other fields.

[0114] Activation function: The activation function is a function used to introduce nonlinearity in a neural network. Common activation functions include Sigmoid, ReLU, Tanh, etc. The activation function enables the neural network to fit complex nonlinear relationships and improve the expressiveness of the model.

[0115] Dropout: Dropout is a regularization technique that prevents the model from overfitting by randomly setting the output of some neurons to zero during training. Dropout effectively improves the generalization ability of the model and reduces the risk of overfitting.

[0116] Layer Normalization: Layer normalization is a regularization technique that normalizes the output of each layer to reduce internal covariate shift. Layer normalization is particularly effective in models such as RNN, helping to stabilize the training process and improve model performance.

[0117] Model quantization: Model quantization is a model compression technology that reduces the model's storage and computing resource consumption by converting model weights and activation values ​​from high precision (such as 32-bit floating points) to low precision (such as 8-bit integers). In practical applications, model quantization helps to improve inference speed and reduce power consumption.

[0118] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for detecting DC arcing in a photovoltaic system, characterized in that: The specific steps include: Step 1: Training data collection and processing; Step 1.1: Collect the physical quantities related to arc detection in the continuous process from no arc to arc, perform analog-to-digital conversion and then divide the frames to obtain a discrete training data set; Step 1.2: filtering the training data in step 1.1 to obtain a multi-frame signal that is continuous in time sequence; the filtering adopts a threshold filtering method based on histogram statistics, calculates the data distribution histogram in the discrete training data set, sorts the data according to the size, obtains the proportion of burr signals according to the histogram statistics of the data, determines the filtering threshold, and performs truncation according to the threshold; Step 1.3: transform the multi-frame signal in step 1.2 to obtain the corresponding time-frequency spectrum; Step 2: DC arc detection model training; Step 2.1: Use the frequency domain feature network to extract frequency domain features; Step 2.2: Use the time domain feature network or time series model to extract the time series changes of the frequency domain features; In the step 2.1 and the step 2.2, the DC arc detection model adopts a combined detection model of a frequency domain feature network and a time domain feature network, and the target task is a binary classification task of whether an arc occurs; the time-frequency spectrum of the current or voltage signal is used as the input of the DC arc detection model, and the arc probability is used as the output of the DC arc detection model; by constructing a combined detection model network framework of a frequency domain feature network and a time domain feature network, wherein the frequency domain feature network is used for frequency domain feature extraction of the frequency dimension of the current or voltage signal, and the time domain feature network is used for time domain feature extraction of the current or voltage signal, and the output of the frequency domain feature network is used as the input of the time domain feature network, and finally the frequency domain features of the current or voltage signal and the time series changes of the frequency domain features are combined, and the arc probability calculation is realized through a fully connected layer, thereby completing the training of the DC arc detection model; Step 3: Online detection of DC arc.

2. A method for detecting DC arc in a photovoltaic system according to claim 1, characterized in that: The physical quantity is current or voltage.

3. A method for detecting DC arc in a photovoltaic system according to claim 2, characterized in that: The current refers to the AC component in the DC current.

4. A method for detecting DC arc in a photovoltaic system according to claim 1, characterized in that: The step 3 comprises: Step 3.2: Collect real-time data, which is a physical quantity related to arc detection; Step 3.3: Filter the real-time data and calculate the time-frequency spectrum; Step 3.4: Input the time-frequency spectrum into the trained DC arc detection model to calculate the arc probability of real-time data.

5. A method for detecting DC arc in a photovoltaic system according to claim 4, characterized in that: The step 3 also includes step 3.1, and the step 3.1 is before step 3.2, and the step 3.1 is to deploy the trained DC arc detection model to the embedded platform.

6. A method for detecting DC arc in a photovoltaic system according to claim 1, characterized in that: The transformation is short-time Fourier transformation or wavelet transformation.

7. A method for detecting DC arc in a photovoltaic system according to claim 1, characterized in that: The frequency domain feature network is a convolutional neural network or a feedforward neural network.

8. A method for detecting DC arc in a photovoltaic system according to claim 1, characterized in that: The time domain feature network is a long short-term memory network, or a simple recurrent neural network, or a gated recurrent unit.

9. A method for detecting DC arc in a photovoltaic system according to claim 7, characterized in that: The convolutional neural network is a one-dimensional convolutional neural network or a two-dimensional convolutional neural network.

10. A method for detecting DC arc in a photovoltaic system according to claim 4 or 5, characterized in that: The step 3.4 includes making an arcing judgment on each frame of real-time data, and determining that the current frame is arcing when the arcing probability of the current frame is greater than a first threshold, and determining that the current frame is operating normally when the arcing probability of the current frame is less than or equal to the first threshold; when the number of arcing frames is greater than a second threshold, determining that the photovoltaic system is arcing.

11. A method for detecting DC arc in a photovoltaic system according to claim 10, characterized in that: The first threshold range is [50%, 60%], and the second threshold range is [N / 2, n], where N is the number of detection frames, n is a positive integer, and the value range of n is (N / 2, N].

Images