A direct current fault arc detection method and system based on Monte Carlo simulation feature optimization

Abstract

The application relates to a DC fault arc detection method and system based on Monte Carlo simulation feature optimization, which comprises the following steps: collecting normal and fault arc data of different stages of a photovoltaic system inverter in a grid-connected state in a set time window, determining a band-pass filtering range by comparing the frequency spectrum of the arc at different stages and the normal current frequency spectrum through Fourier transform; collecting the electrical quantity of the DC side of the photovoltaic system in real time under normal grid-connected and simulated fault conditions, and then performing filtering processing; then, the optimal intrinsic mode function is determined through Fourier transform and empirical mode decomposition, combined with the variance contribution rate and the Pearson correlation coefficient, so that the time-frequency domain feature effect is increased; the Monte Carlo simulation method is adopted to find the optimal characteristic quantity, which is used for training a BP neural network model, and a genetic algorithm is used for optimization to obtain a DC fault arc detection model. Compared with the prior art, the application can improve the detection and recognition accuracy of fault arcs at different stages, and improve the anti-interference performance and detection speed.

Description

TECHNICAL FIELD

[0001] The present application relates to a direct current power system, in particular to a DC fault arc detection method and system based on Monte Carlo simulation feature optimization. BACKGROUND

[0002] With the development and application of clean energy becoming more and more widespread, in the process of energy transformation, distributed photovoltaic power generation has become one of the key trends. Distributed photovoltaic system has flexibility and environmental protection, which can better adapt to the energy demand of different regions. As an important category of photovoltaic inverters, string inverter has made a great contribution to the development of distributed photovoltaic power generation, and has become the type of inverter that photovoltaic enterprises focus on research and innovation. However, in the process of operation, the string on the DC side of the inverter may be loose, the insulation layer may be broken, and other reasons may cause DC arc fault.

[0003] In some large photovoltaic power stations, the power of a string inverter can reach hundreds of kilowatts. If the DC side fault arc cannot be detected in time and the inverter is not turned off quickly, it may cause damage to the inverter and cause serious fire accidents. Compared with parallel fault arc, when the inverter DC side appears series fault arc, it will cause the loop current to decrease rapidly, and it is difficult to detect the fault by setting protection devices or detection elements outside. Therefore, the harm of series fault arc is greater, and it is more likely to cause electrical fire.

[0004] At present, the research methods for detecting DC fault arc of photovoltaic power generation system mainly include: detection methods based on arc light, arc sound and electromagnetic radiation characteristics; detection methods based on arc current and voltage time-frequency domain characteristics. Among them, the detection method based on arc light, arc sound and electromagnetic radiation characteristics is to study the radiation characteristics of arc, and to detect faults according to the frequency spectrum characteristics, but it is greatly affected by the interference of complex environment, and is greatly limited by sensors; the detection method based on arc current and voltage time-frequency domain characteristics is the most commonly used detection method, but in photovoltaic power generation system, the scene is complex and the components are numerous, normal switch action and load mutation may cause misjudgment, and the threshold cannot change with the environment. In a word, the existing detection methods cannot quickly, stably and reliably detect DC fault arc, especially for the detection and identification of fault arc in different stages, which has the problem of low accuracy. SUMMARY

[0005] The purpose of the present application is to overcome the defects of the prior art and provide a DC fault arc detection method and system based on Monte Carlo simulation feature optimization, which can improve the detection and identification accuracy of fault arc in different stages, improve the anti-interference performance and detection speed.

[0006] The object of the application can be realized by the technical scheme: a DC fault arc detection method based on Monte Carlo simulation feature optimization, comprising the following steps:

[0007] S1, collecting normal and fault arc data of different stages in the grid-connected state of the inverter of the photovoltaic system in a set time window;

[0008] S2, Fourier transform is performed on the collected arc data, the frequency spectrum of different stages of the arc is compared with the normal current frequency spectrum, the main energy concentration frequency band of the arc is analyzed, and the band-pass filtering range is determined;

[0009] S3, real-time collection of DC side electrical quantities of the inverter of the grid-connected photovoltaic system in normal grid-connected and simulated fault conditions, and filtering processing according to the band-pass filtering range;

[0010] S4, Fourier transform and empirical mode decomposition (EMD) are performed on the filtered DC side electrical quantities, the optimal intrinsic mode function is determined by combining the variance contribution rate and the Pearson correlation coefficient, and the current frequency domain and time-frequency domain information is extracted;

[0011] S5, in the current frequency domain and time-frequency domain, the optimal feature quantity is searched for in a Monte Carlo simulation manner, so as to serve as a training set of a genetic algorithm optimized BP (Back Propagation, back propagation) neural network model, and a DC fault arc detection model is obtained through training;

[0012] S6, real-time collection of current signals and calculation of corresponding frequency domain and time-frequency domain feature quantities, input into the DC fault arc detection model, and real-time detection and judgment of the DC arc.

[0013] Further, the specific process of step S1 is: simulating a plurality of series DC fault arc experiments, collecting a plurality of sets of current data of normal arcs and arcs at different stages in a set time window through an oscilloscope; wherein the current data of arcs at different stages include: data in the set time window after the appearance of arc spikes, and data in the set time window after the disappearance of arc spikes.

[0014] Further, the specific process of step S2 is: after filtering out the DC component of the collected arc data time domain signal, the discrete Fourier transform is performed, and the formula is as follows:

[0015]

[0016] Wherein, X(k) is a complex number form of the frequency domain signal at frequency , x(n) is the nth sample of the time domain signal, and N is the sample number of the signal;

[0017] The amplitude information is extracted from the discrete Fourier transform result X(k), and the formula is as follows:

[0018] A(k)=|X(k)|

[0019] Where A(k) is the frequency f k The amplitude at , that is, the spectrum value of the normal and fault arc comparison;

[0020] Compare the spectrum diagrams of normal and fault arcs within the set frequency value, and select the common concentrated frequency band where the arc at different stages is greater than the normal spectrum as the bandpass filtering range.

[0021] Furthermore, the specific process of performing empirical mode decomposition on the filtered DC side electrical quantity in step S4 is as follows:

[0022] By iteratively decomposing the time domain signal x(t), a set of intrinsic mode functions (IMFs) and a residual term r(t) are obtained as follows:

[0023]

[0024] Among them, c i (t) is the i-th eigenmode function, I is the total number of decomposition levels;

[0025] The signal maximum point is defined as: {t j ,x j}, the minimum point is: {t k ,x k}, connect the adjacent maximum and minimum points to construct the upper and lower envelopes u of the signal up (t) and u down (t);

[0026] Calculate the smoothed signal m(t), the formula is:

[0027]

[0028] Extract the first intrinsic mode function c1(t), whose formula is:

[0029] c1(t)=x(t)-m(t)

[0030] Repeat the above process and iterate c1(t) until the preset stopping condition is met to obtain all the intrinsic mode functions, where the preset stopping condition is whether the residual term r(t) is a function that tends to be stable or noise.

[0031] Furthermore, the specific process of determining the optimal intrinsic mode function in step S4 is:

[0032] S41, calculating variance contribution rates of normal and different stage fault arc data under the same current level;

[0033] S42, calculating Pearson correlation coefficients of normal and different stage fault arc data under the same current level;

[0034] S43, selecting several IMF components with greater difference by comparing variance contribution rates of each order IMF component of normal and different stage fault arc, and determining an optimal IMF component by combining correlation coefficients of each order IMF component of normal signal and correlation coefficients of each order IMF component of original signal.

[0035] Further, the variance contribution rate of the arc data in the step S41 is specifically:

[0036]

[0037] wherein, VCR i is the variance contribution rate, Var(S) is the variance of the original time domain signal, Var(IMF i ) is the variance of the i-th IMF component, N is the number of sampling points in a period, c i (k) is the value of the i-th IMF at the k-th sampling point, and x(k) is the value of the original signal at the k-th sampling point.

[0038] Further, the specific process of the step S42 is:

[0039] Suppose that the normal, stage 1 arc and stage 2 arc signals are X1, X2 and X3 respectively, the sample covariance between normal and different stage fault arc is calculated as follows: and

[0040]

[0041] The standard deviation of the original signal is calculated as follows:

[0042]

[0043] The Pearson correlation coefficient between normal and different stage fault arc is calculated as follows:

[0044]

[0045] wherein, r1 is the Pearson correlation coefficient of normal and stage 1 fault arc data, and r2 is the Pearson correlation coefficient of normal and stage 2 fault arc data.

[0046] Furthermore, the specific process of finding the optimal characteristic quantity in step S5 is: for several groups of frequency domain and time-frequency domain information of normal and different stage arcs, based on the Monte Carlo simulation method, characteristic statistics are obtained through different frequency band combinations, and the characteristic statistics with the largest difference between normal and fault arcs are selected as frequency domain and time-frequency domain characteristic quantities. The characteristic statistics include: the mean, standard deviation, and root mean square value of different frequency band combinations in the spectrum, that is, the mean, standard deviation, and root mean square value of different frequency band combinations in the spectrum are obtained after Fourier transform of the optimal IMF component.

[0047] Furthermore, the specific process of training the DC fault arc detection model in step S5 is: obtaining several sets of frequency domain and time-frequency domain feature quantities of normal and different stage arcs as training samples of the BP neural network model. After the training is completed, the obtained model is optimized by genetic algorithm to construct the DC fault arc detection model.

[0048] A DC fault arc detection system based on Monte Carlo simulation feature optimization, comprising:

[0049] The sampling module is used to perform real-time sampling of the filtered electrical quantities on the DC side of the actual photovoltaic system inverter;

[0050] Bandpass filter module, used to filter the electrical quantities on the DC side of the actual photovoltaic system inverter;

[0051] Communication module, used to communicate with the host computer and transmit the current signal data obtained by sampling in the current cycle;

[0052] A Fourier transform and empirical mode decomposition module is used to perform Fourier transform and empirical mode decomposition on several sets of received current signal data to extract frequency domain and time-frequency domain feature information;

[0053] A feature extraction module is used to perform Monte Carlo simulation on the frequency domain and time-frequency domain feature information of several groups of current signals to extract the frequency domain and time-frequency domain feature quantities;

[0054] The model training module is used to use the frequency domain and time-frequency domain feature quantities obtained by the feature extraction module as BP neural network model training samples, and to construct a DC fault arc detection model through model training and optimization using a genetic algorithm;

[0055] The judgment module uses the DC fault arc detection model to judge the current signal obtained by the sampling module and detect the DC fault arc in real time.

[0056] Compared with the prior art, the present invention has the following advantages:

[0057] The present invention collects normal and fault arc data at different stages of the inverter grid-connected state in a set time window in a photovoltaic system; combines Fourier transform to determine the bandpass filtering range; then collects DC side electrical quantities of the grid-connected photovoltaic system inverter under normal grid-connected and simulated fault conditions in real time, and performs filtering processing according to the bandpass filtering range; then performs Fourier transform and empirical mode decomposition on the filtered DC side electrical quantities, and combines the variance contribution rate and the Pearson correlation coefficient to determine the optimal intrinsic mode function to extract current frequency domain and time-frequency domain information; finally, Monte Carlo simulation is used to find the optimal feature quantity in the current frequency domain and time-frequency domain, which is used as a training set for the genetic algorithm to optimize the BP neural network model, and a DC fault arc detection model is obtained through training. Therefore, in order to address the problems that when the actual DC fault is in the stable burning stage, the characteristic quantity is not obvious and is easily confused with the fluctuation of normal current under external interference, Monte Carlo simulation feature optimization is used to extract frequency domain and time-frequency domain features. For the empirical mode decomposition time domain signal, the variance contribution rate and Pearson correlation coefficient are combined to determine the optimal intrinsic mode function to increase the time-frequency domain feature effect. It can effectively determine the optimal characteristic quantity for distinguishing normal and fault arcs, thereby improving the detection and identification accuracy of fault arcs at different stages, while improving the anti-interference performance and detection speed.

[0058] The present invention can effectively improve the anti-interference and universality of characteristic information to environmental noise by finding frequency domain and time-frequency domain feature quantities with minimum error in normal and fault current data sets.

[0059] The present invention adopts the Monte Carlo simulation method to extract the frequency domain and time-frequency domain features of normal and fault data sets under different working conditions as samples for training the BP neural network model, and uses the genetic algorithm to optimize the BP neural network model to avoid the problems of slow convergence of BP neural network training and easy falling into local optimal solutions, thereby further improving the detection speed and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 Schematic diagram of the method flow of the present invention;

[0061] Figure 2 Schematic diagram of the application process of the embodiment;

[0062] Figure 3 This is a schematic diagram of a photovoltaic system in which an oscilloscope is used to collect DC side electrical quantities after simulating an arcing experiment in the embodiment;

[0063] Figure 4 The time domain diagram and spectrum diagram of the normal and fault arc initiation stages collected by the oscilloscope in the embodiment;

[0064] Figure 5Time-domain and frequency spectrum graphs of normal and fault arc stable combustion stages collected by the oscilloscope in the embodiment;

[0065] Figure 6 A schematic diagram of the DC fault arc detection system in the embodiment;

[0066] Figures 7a to 7c Time-domain and frequency spectrum graphs of normal and fault different stages collected in the embodiment;

[0067] Figure 8 A flowchart of the empirical mode decomposition current time-domain signal in the embodiment;

[0068] Figure 9 IMF component graphs of the empirical mode decomposition current time-domain signal in the embodiment;

[0069] Figure 10 Monte Carlo simulation in the embodiment in the process of optimizing several groups of normal and different stage fault arc frequency domain characteristics;

[0070] Figure 11 The frequency band combination-error distribution graph after three times of characteristic optimization in the embodiment;

[0071] Figure 12 A comparison graph of normal and fault frequency domain characteristic quantities in the embodiment;

[0072] Figure 13 A BP neural network model training effect graph in the embodiment;

[0073] Figure 14 A genetic algorithm optimized BP neural network model verification effect graph in the embodiment. DETAILED DESCRIPTION

[0074] The application will be described in detail below in conjunction with the drawings and specific embodiments.

[0075] Embodiment

[0076] At present, the characteristic information of the fault arc is easily affected by the voltage, current level and arc triggering position, and the analysis of the current, voltage level and other normal and fault waveforms under different working conditions will find that the arc combustion stage and the normal waveform are very close, and the arc occurrence position is random in practice, and the current information of the arc is easier to analyze than the voltage, therefore, the DC fault arc detection often takes the current component on the DC side as the research object.

[0077] When a fault arc occurs, it is difficult to find a distinguishing degree from the time domain angle and is easily disturbed by external environmental noise because the waveform of the arc in the stable combustion stage is very close to the normal waveform, but the fault characteristic information can be found from the frequency domain to distinguish the normal waveform. Compared with the frequency spectrum of normal current, the energy of DC arc is mainly distributed in the (10-100) kHz frequency band, and the harmonic content of fault arc is more in this frequency band. The current level has a greater impact on the energy distribution of arc frequency, and when the current level is low, the energy of fault arc is distributed below 100 kHz, but when the current level increases, there is more high-frequency noise, and the arc energy is mainly concentrated in the low frequency band. Because the frequency domain characteristics of arcs at different stages are not the same, relying only on frequency domain characteristic quantities as features to detect fault arcs can easily lead to misjudgment, especially in the small arc stage before and after the arc, the frequency domain signal has small distinguishing degree, so the empirical mode decomposition (EMD) method can be considered for time-frequency analysis of the original signal, which is beneficial to the analysis of the arc time series signal which is nonlinear and non-stationary, but more intrinsic mode functions (IMF) will be obtained after decomposition, and the number of IMF components after decomposition of different signals may differ. Therefore, the variance contribution rate can be selected to select the optimal IMF component, but there may be a small variance contribution rate of the arc signal at a certain stage, resulting in errors and ultimately misjudgment. However, different loads, current levels and voltage levels may affect the arc, and the input voltage level of the inverter DC side in practice may exceed 1000V or more. In order to reduce the false positive rate and false negative rate of fault arc detection, machine learning algorithms are introduced into arc detection. However, the hardware requirements are high, and the implementation cost is high, and most of them are still in the theoretical research stage.

[0078] Based on the above description, referring to Figure 1 The present scheme proposes a DC fault arc detection method based on Monte Carlo simulation feature optimization, comprising the following steps:

[0079] S1, collect normal and fault arc data at different stages of the inverter in the grid-connected state of the photovoltaic system within a set time window, specifically, simulate a number of series DC fault arc experiments, and collect a number of normal and different stage arc current data within a set time window through an oscilloscope; wherein the current data of different stage arcs includes: data within a set time window after the appearance of arc peak, data within a set time window after the disappearance of arc peak;

[0080] S2, Fourier transform is performed on the collected arc data, the frequency spectrum of the arc at different stages is compared with the frequency spectrum of the normal current, and the arc energy main concentration frequency band is analyzed to determine the band pass filtering range, specifically, after filtering out the direct current component of the time domain signal, the discrete Fourier transform is performed, and the formula is as follows:

[0081]

[0082] Where X(k) is the frequency domain signal at frequency The complex form of x(n) is the nth sample of the time domain signal, and N is the number of samples of the signal;

[0083] Extract the amplitude information from the discrete Fourier transform result X(k), and the formula is as follows:

[0084] A(k)=|X(k)|

[0085] Where A(k) is the frequency f k The amplitude at , that is, the spectrum value of the normal and fault arc comparison;

[0086] Comparing the spectrum of normal and fault arcs within 100kHz, the common concentrated frequency band where the arc at different stages is greater than the normal spectrum is selected as the range of the bandpass filter circuit;

[0087] S3. Real-time collection of DC side electrical quantities of the grid-connected photovoltaic system inverter under normal grid-connected and simulated fault conditions, and filtering according to the bandpass filtering range. Specifically, several groups of DC side electrical quantities under normal and fault conditions after bandpass filtering can be obtained through AD sampling, including: AC components of DC current;

[0088] S4. Perform Fourier transform and empirical mode decomposition on the filtered DC side electrical quantities. Combined with the variance contribution rate and Pearson correlation coefficient, determine the optimal intrinsic mode function to extract the current frequency domain and time-frequency domain information. Specifically, for the empirical mode decomposition time domain signal, it includes:

[0089] Using the EMD method, the time domain signal x(t) is iteratively decomposed to obtain a set of intrinsic mode functions (IMFs) and a residual term r(t), as shown below:

[0090]

[0091] Among them, c i (t) is the i-th eigenmode function, I is the total number of decomposition levels;

[0092] The signal maximum point is defined as: {t j ,x j}, the minimum point is: {t k ,x k}, connect the adjacent maximum and minimum points to construct the upper and lower envelopes u of the signal up (t) and u down (t);

[0093] Calculate the smoothed signal m(t), the formula is:

[0094]

[0095] Extract the first intrinsic mode function c1(t), whose formula is:

[0096] c1(t)=x(t)-m(t)

[0097] Repeat the above process and iterate c1(t) until the preset stopping condition is met to obtain all the intrinsic mode functions. The preset stopping condition is whether the residual term r(t) is a function that tends to be stable or noise.

[0098] For the sampled normal and fault arc time domain signals, a set of intrinsic mode functions is obtained through the above-mentioned empirical mode decomposition. Then, the variance contribution rate and Pearson correlation coefficient are combined to determine the optimal intrinsic mode function, including:

[0099] The variance contribution rate of normal and different stage fault arc data at the same current level is calculated as follows:

[0100] Calculate the variance of the original time-domain signal:

[0101]

[0102] Calculate the variance of each IMF component:

[0103]

[0104] Calculate the variance contribution rate:

[0105]

[0106] Among them, N is the number of sampling points in a cycle, c i (k) is the value of the i-th IMF at the k-th sampling point, and x(k) is the value of the original signal at the k-th sampling point;

[0107] Calculate the Pearson correlation coefficient of normal and different stage fault arc data at the same current level. Assuming that the normal, stage 1 arc, and stage 2 arc signals are X1, X2, and X3 respectively, the formula is as follows:

[0108] Calculate the sample covariance between normal and fault arcs at different stages and as follows:

[0109]

[0110] Calculate the standard deviation of the original signal as follows:

[0111]

[0112] The Pearson correlation coefficients between normal and different stage fault arcs are calculated, as follows:

[0113]

[0114] By comparing the variance contribution rates of the IMF components of the normal and different stage fault arcs, a number of IMF components with greater differences are preliminarily selected, and the correlation coefficients of the IMF components based on the normal signal and the correlation coefficients of the IMF components based on the original signal are combined to determine an optimal IMF component;

[0115] S5, in the current frequency domain and time-frequency domain, an optimal feature quantity is searched for by using Monte Carlo simulation, to be used as a training set of a BP neural network model optimized by a genetic algorithm, and a DC fault arc detection model is obtained by training. Specifically, for a number of groups of normal and different stage arc frequency domain and time-frequency domain information, a method based on Monte Carlo simulation is proposed, feature statistics are obtained by different frequency band combinations, and the feature statistics with the greatest difference between normal and fault arcs are selected as the frequency domain and time-frequency domain feature quantities. The feature statistics include the mean, standard deviation, and root mean square value of different frequency band combinations in the frequency spectrum, and the optimal IMF component is subjected to Fourier transform to obtain the mean, standard deviation, and root mean square value of different frequency band combinations in the frequency spectrum.

[0116] The obtained frequency domain and time-frequency domain feature quantities of a number of groups of normal and different stage arcs are used as training samples of a BP neural network model, and after training, the obtained model is optimized by a genetic algorithm to construct a DC fault arc detection model.

[0117] S6, real-time current signal is collected and corresponding frequency domain and time-frequency domain feature quantities are calculated, which are input into the DC fault arc detection model for real-time detection and judgment of the DC arc.

[0118] The embodiment applies the above technical solution, as shown in Figure 2 , mainly including:

[0119] I. The oscilloscope is used to collect 10ms time window normal and fault arc data of different stages of the inverter in the grid-connected state of the photovoltaic system;

[0120] II. The collected current signal is subjected to Fourier transform, the frequency spectrum of the arc at different stages is compared with the normal current frequency spectrum, the arc energy is mainly concentrated in the frequency band, and the band-pass filtering range is determined;

[0121] III. The real-time DC side electrical quantity of the inverter in the normal grid-connected and simulated fault working condition of the grid-connected photovoltaic system is collected;

[0122] Fourth, perform Fourier transform and empirical mode decomposition on the sampled current signal, use the variance contribution rate and Pearson correlation coefficient method to determine the optimal intrinsic mode function, and extract the current frequency domain and time-frequency domain information;

[0123] Fifth, Monte Carlo simulation is used to find the optimal feature quantity in the current frequency domain and time-frequency domain. This feature quantity is used as the training set for the genetic algorithm to optimize the BP neural network model. After the training is completed, a DC fault arc detection model is constructed.

[0124] 6. Real-time sampling is used to obtain current signals and calculate the frequency domain and time-frequency domain feature quantities, which are then input into the trained detection model to perform real-time judgment on DC arcs.

[0125] Combine Figure 3 As shown, in step 1: several DC fault arc experiments are simulated, the oscilloscope sampling frequency f = 4MHz, the time window is 10ms, the number of sampling points: N = 40000, AC coupling and other parameters are set, and normal and fault arc data of several groups of photovoltaic systems at different stages of the inverter grid-connected state under three current levels are collected.

[0126] Combine Figure 4 、 Figure 5 As shown, in step 2: the DC component of the collected current data is filtered out and then Fourier transformed to calculate the spectrum value of each group. Since the energy of the fault arc is mainly concentrated below 100kHz, in order to enhance the contrast of the frequency domain information, this embodiment expresses the spectrum in a logarithmic form with a base of 10 to observe the frequency bands where the energy of the normal, arc starting stage and stable combustion stage is mainly concentrated. Figure 4 It can be seen that above 50kHz, the frequency domain energy of the normal and fault initiation stages is basically mixed together due to factors such as environmental noise and inverter control strategy. And as the current level decreases, the distinction between [20-50]kHz becomes less and less obvious. However, among the three current levels, the distinction below 20kHz is the most obvious. Figure 5 As can be seen from the figure, above 30kHz, the frequency domain energy during the stable arc burning phase is essentially mixed with normal ambient noise. As the current level decreases, only information below 20kHz has good discrimination, but the frequency domain energy decreases as the current level decreases. In summary, this embodiment considers a bandpass filter circuit with a range of [2-20]kHz.

[0127] This embodiment also provides a DC fault arc detection system, combined with Figure 6As shown, mainly includes sampling module, band pass filter module, communication module, Fourier transform and empirical mode decomposition module, feature extraction module, model training module and judgment module. In step three, step four: through the sampling module, a large number of photovoltaic system inverter grid-connected state of different stages of normal and fault arc data are obtained, the frequency spectrum value of each group of data is calculated after Fourier transform, and the time domain and frequency spectrum diagram of a group of normal, arc starting stage and stable combustion stage are shown in FIG. 7. Compared with the normal current frequency spectrum diagram, the energy of the arc starting and stable combustion stage increases obviously below 20 kHz, and the energy of the starting stage is greater than that of the combustion stage, mainly concentrated in the area below 16 kHz. The current data are obtained by the sampling module and the normal and fault current data set is constructed. Since the string inverter needs multiple string inputs, this embodiment realizes the sampling of multiple string access signals through a current transformer, which contains the current data of the normal operation of the inverter under various working conditions, and the data of different combinations of the number and direction of strings in the current transformer, and the data when accessing different strings during the fault arc experiment. In the current spectrum under different working conditions, the characteristic frequency band of the fault arc is found, and the frequency domain characteristic quantity is extracted, so that the real-time fault arc detection can be realized more accurately. In step four: the current data are obtained by the sampling module, and the current signal is decomposed by empirical mode decomposition, and the specific decomposition process is shown in FIG. Figure 8 As shown in Figure 9 As shown, after decomposition, a large number of IMF components are obtained. If all the components are selected as time-frequency domain features, firstly, the calculation time of the arc detection circuit will be increased, resulting in an increase in arc detection time, and secondly, some unnecessary components will confuse the characteristic information of normal current and arc stage current, reducing the recognition accuracy. Therefore, in order to obtain the optimal IMF component, this scheme proposes an IMF component determination method based on Pearson correlation analysis combined with the variance contribution rate of each component.

[0128] Among them, the variance contribution rate of any one group of current data is shown in Table 1:

[0129] Table 1 Variance contribution rate of each order IMF component

[0130]

[0131] As shown in Table 1, under three working conditions, the variance contribution rate of IMF1 component is the largest, and the variance contribution rates of the remaining IMF components are small, indicating that the main energy distribution is concentrated in the IMF1 component. However, the difference between the variance contribution rates of IMF1 of normal and arc starting and arc combustion stages is the smallest, and the difference between the variance contribution rates of IMF2 and IMF3 is the largest, indicating that when a fault occurs, the energy of IMF2 and IMF3 components will increase. In order to further determine which component between IMF2 and IMF3 can better reflect the characteristic information of different stages of fault arc.

[0132] Solve the Pearson correlation coefficient under different working conditions and find the IMF component with the smallest correlation with the original signal and the largest difference between other working conditions. Perform the following steps:

[0133] Taking the normal signal’s various order IMF components as the benchmark, the correlation coefficients of the arcing and arcing’s various order IMF components relative to the normal signal are calculated as shown in Table 2. Taking the normal, arcing and arcing’s original signals as the benchmark, the correlation coefficients of the various order IMF components relative to the original signal are calculated as shown in Table 3.

[0134] Table 2 Correlation coefficients based on normal signal IMF components of various orders

[0135] Quantity normal Arc starting Arcing IMF1 1.0000 -0.6884 -0.9146 IMF2 1.0000 0.0912 0.0553 IMF3 1.0000 0.0204 0.0018 IMF4 1.0000 0.0656 0.1093 IMF5 1.0000 0.2643 0.3395 IMF6 1.0000 0.3152 -0.3632

[0136] Table 3 Correlation coefficients of IMF components of various orders based on the original signal

[0137]

[0138] From the analysis in Table 2, we can see that the correlation coefficients of the IMF2 and IMF3 components of the arcing and arcing signals are all less than 0.1 compared to the normal signal, indicating that the normal and fault data in these two components have a certain degree of distinction. If the three IMF1 components are highly correlated, the waveforms of the corresponding normal, arcing, and arcing signals will be similar, which may lead to misjudgment.

[0139] From the analysis in Table 3, we can see that, based on the original signals of the three signals, the IMF2 component of the normal signal is relatively small, and the IMF2 components of arcing and arcing are the most different from those of the normal signal. Therefore, the IMF2 component is selected as the time-frequency domain feature information to distinguish normal and fault currents.

[0140] In step five: through the sampling module, a large number of photovoltaic system inverter grid-connected state of different stages of normal and fault arc data are obtained; after the filtering module, higher frequency components are filtered out to avoid interference caused by environmental noise; then after the Fourier transform and empirical mode decomposition module, a number of groups of normal and fault arc frequency domain and time-frequency domain information in different stages are obtained, including: the spectral energy distribution below 20 kHz and the most different IMF2 component between normal and fault arcs. Directly using the frequency domain and time-frequency domain information as the input of the model training module may lead to too long training time and easy overfitting, ultimately affecting the detection accuracy of the arc judgment module. Therefore, for the frequency domain and time-frequency domain information, feature selection and dimension reduction are needed to select the most representative features to reduce the data dimension. The scheme proposes a feature optimization method based on Monte Carlo simulation, which randomly rearranges the order of the normal and fault data sets, selects statistical quantities for analysis and comparison, and thus obtains the optimal feature frequency band that distinguishes the two.

[0141] The specific steps of selecting the frequency domain feature quantity are as follows:

[0142] (1) Select a random number to rearrange the order of each group of normal and fault data sets;

[0143] (2) Select a statistical quantity, make an error function for comparison, and take the error function as: err=(N_X>F_X) number / group number, where: N_X refers to the statistical quantity of a certain frequency band, how many groups are greater than the fault, i.e. the error number, and finally divided by the number of groups to get the error percentage; The statistical quantity includes the following: assuming that the current collected in an arc detection circuit is: i(t), which contains N frequency points in the frequency domain: x1, x2,…x i ,…,x N .

[0144] Table 4 Feature statistical quantity formula

[0145]

[0146] (3) Start from the first point and traverse to solve the error function;

[0147] (4) Complete the error function solution, rearrange the order of the normal and fault data sets, solve the error again, and finally get all the error function results.

[0148] Take the standard deviation of the optimal frequency band in the normal and fault frequency domain as an example, select 100 groups of detection period data, which contain 50 groups of normal data and 50 groups of fault current, and use the Monte Carlo simulation method to optimize the frequency domain features. The process is shown in Figure 10Because the switching frequency 16 kHz occupies a large component, it may interfere with the characteristic frequency band optimization, therefore, only the components below 16 kHz are optimized for the characteristic frequency band. The frequency band points below 16 kHz are combined into: 1-2, 1-3, 1-4, …, 2-1, …, 48-49, a total of 1176 frequency band points, each point represents the standard deviation error obtained by averaging the error results after 1000 times of disordering sequence. In order to avoid the randomness of one optimization result, the above process is repeated three times, and the frequency band error distribution diagram is shown in Figure 11 . Specifically, the optimal frequency band distribution is shown in Table 5:

[0149] Table 5 Optimal frequency band distribution table

[0150] Number of optimization searches Frequency band combination Corresponding frequency band point Corresponding frequency band / kHz Minimum error 1 805 22：29 (7.15-9.425) 0.0083 2 806 22：30 (7.15-9.75) 0.01 3 829 23：27 (7.475-8.775) 0.009

[0151] After three Monte Carlo simulation characteristic optimizations, the best frequency band combination is distributed around 800, which indicates that using Monte Carlo simulation for multiple optimizations may cause some errors, but it can always be kept within a certain deviation range. Finally, take the frequency band combination 805 with the smallest error in the three optimizations, that is, for the frequency domain standard deviation characteristic quantity, the optimal frequency band is concentrated in (7.15-9.425) kHz, and the minimum error is 0.0083. Finally, select more characteristic quantities to optimize using the Monte Carlo simulation method to obtain the best characteristic quantity for distinguishing normal and fault in the frequency domain (see Figure 12 ), including but not limited to the four types of characteristic quantities shown in Figure 12 .

[0152] Among them, the specific steps of selecting frequency domain characteristic quantities are as follows:

[0153] (1) Calculate the IMF2 component of the normal and fault arc data sets, and perform Fourier transform;

[0154] (2) Repeat the above process, and finally obtain the following types of error minimum time-frequency domain characteristic quantities:

[0155] Table 6 Summary of time-frequency domain characteristic quantities

[0156] Frequency domain feature Optimal frequency band / kHz Minimum error Standard deviation (6.825-15.6) 0.0015 mean (6.175-10.075) 0.0010 RMS value (6.5-12.350) 0.0009 Discrete values (6.175-10.725) 0.0009

[0157] Thus, the frequency domain and time-frequency domain characteristic quantities of normal and fault data sets under different working conditions are extracted by using the Monte Carlo simulation method, as samples for training the artificial intelligence neural network model. The model output value of normal data is set to 0, and the model output value of fault data is set to 1000, which is convenient for the calculation of the judgment module and the subsequent adjustment of the detection threshold size. The detection effect of the training set, test set and validation set is shown in Figure 13To avoid the problem that the neural network model is easy to fall into local optimal solution, the genetic algorithm is used to optimize it, and 23 groups of data are verified, and the comparison of verification effect is as follows Figure 14 .

[0158] In step six, the current signal of a period is obtained by the sampling module, the higher frequency components that will cause interference are filtered out through the band-pass filtering module, and the low frequency components with higher contrast are retained; the frequency domain and time-frequency domain information of the current period are obtained through the Fourier transform and empirical mode decomposition module, the optimal feature quantity is extracted through the feature extraction module, and finally the judgment module is used to judge whether the model output value is continuous multiple times (set number) greater than the set detection threshold value, if yes, it is judged that arc fault occurs, otherwise it is judged that arc fault does not occur.

[0159] As can be seen from the above, the current data collected under different current levels, line impedance, control strategies and external sudden disturbances are studied to effectively determine the optimal feature quantity to distinguish normal and fault arc. And by finding the minimum error frequency domain and time-frequency domain feature quantity in the normal and fault current data set, the anti-interference and universality of the feature information to the environment noise are finally improved. In addition, the genetic algorithm is used to optimize the BP neural network model, which avoids the problems of slow convergence and easy to fall into local optimal solution of BP neural network training. The scheme can improve the accuracy of the detection of DC arc fault of the inverter in the photovoltaic system and reduce the false positive rate of the normal operation of the actual photovoltaic system. It has the advantages of small external interference and fast detection speed.

Claims

1. A DC arc fault detection method based on Monte Carlo simulation feature optimization, characterized in that: The following steps are involved: S1. Collect normal and fault arc data at different stages of the inverter grid-connected state in the photovoltaic system within a set time window; S2. Perform Fourier transform on the collected arc data, compare the spectrum of the arc at different stages with the normal current spectrum, analyze the frequency band where the arc energy is mainly concentrated, and determine the bandpass filter range; S3, real-time collection of DC side electrical quantities of the grid-connected photovoltaic system inverter under normal grid connection and simulated fault conditions, and filtering according to the bandpass filter range; S4. Perform Fourier transform and empirical mode decomposition on the filtered DC side electrical quantities, and determine the optimal intrinsic mode function by combining the variance contribution rate and the Pearson correlation coefficient to extract the current frequency domain and time-frequency domain information; S5. Using Monte Carlo simulation, find the optimal feature quantity in the current frequency domain and time-frequency domain to serve as the training set for the genetic algorithm to optimize the BP neural network model, and obtain the DC fault arc detection model through training; S6. Real-time acquisition of the current signal and calculation of the corresponding frequency domain and time-frequency domain feature quantities, which are input into the DC fault arc detection model to perform real-time detection and judgment of the DC arc; Among them, the specific process of finding the optimal characteristic quantity in step S5 is: for several groups of frequency domain and time-frequency domain information of normal and different stage arcs, based on the Monte Carlo simulation method, characteristic statistics are obtained through different frequency band combinations, and the characteristic statistics with the largest difference between normal and fault arcs are selected as frequency domain and time-frequency domain characteristic quantities. The characteristic statistics include: the mean, standard deviation, and root mean square value of different frequency band combinations in the spectrum, that is, the mean, standard deviation, and root mean square value of different frequency band combinations in the spectrum are obtained after Fourier transform of the optimal IMF component.

2. The method for detecting DC arc faults based on Monte Carlo simulation feature optimization according to claim 1, characterized in that: The specific process of step S1 is: simulating several series DC fault arc experiments, and collecting several groups of normal and different stage arc current data within a set time window through an oscilloscope; wherein the arc current data at different stages includes: data within the set time window after the arc spike appears, and data within the set time window after the arc spike disappears.

3. The DC arc fault detection method based on Monte Carlo simulation feature optimization according to claim 1, characterized in that: The specific process of step S2 is: after filtering out the DC component from the collected arc data time domain signal, the signal is subjected to discrete Fourier transform, and the formula is as follows: ， in, is the frequency domain signal at frequency plural form of place, It is the time domain signal samples, is the number of samples of the signal; The result after discrete Fourier transform The amplitude information is extracted from the equation: ， in, is the frequency The amplitude at , that is, the spectrum value of the normal and fault arc comparison; Compare the spectrum diagrams of normal and fault arcs within the set frequency value, and select the common concentrated frequency band where the arc at different stages is greater than the normal spectrum as the bandpass filtering range.

4. The method for detecting DC arc faults based on Monte Carlo simulation feature optimization according to claim 3, characterized in that: The specific process of performing empirical mode decomposition on the filtered DC side electrical quantity in step S4 is as follows: Decompose the time domain signal by iterative , we get a set of intrinsic mode functions (IMFs) and a residual term , as shown below: ， in, It is The eigenmode functions, is the total number of decomposition levels; The signal maximum point is defined as: , the minimum point is: , connect adjacent maximum and minimum points to construct the upper and lower envelopes of the signal and ; Calculate smoothed signal , the formula is: ， Extract the first eigenmode function , the formula is: ， Repeat the above process for Iterate until the preset stopping condition is met to obtain all the intrinsic mode functions, where the preset stopping condition is the residual term Whether it is a stationary function or noise.

5. The DC arc fault detection method based on Monte Carlo simulation feature optimization according to claim 4, characterized in that: The specific process of determining the optimal intrinsic mode function in step S4 is: S41. Calculate the variance contribution rate of normal and fault arc data at different stages under the same current level; S42. Calculate the Pearson correlation coefficient of normal and fault arc data at different stages at the same current level; S43. By comparing the variance contribution rates of the various order IMF components of normal and fault arcs at different stages, several groups of IMF components with relatively large differences are preliminarily selected. The optimal IMF component is determined by combining the correlation coefficients of the various order IMF components based on the normal signal and the correlation coefficients of the various order IMF components based on the original signal.

6. The method for detecting DC arc faults based on Monte Carlo simulation feature optimization according to claim 5, characterized in that: The variance contribution rate of the arc data in step S41 is specifically: ， ， ， in, is the variance contribution rate, is the variance of the original time domain signal, is the variance of the ith IMF component, N is the number of sampling points in a period, It is The IMF The value of the sampling point, The original signal is The value of the sampling point.

7. The method for detecting DC arc faults based on Monte Carlo simulation feature optimization according to claim 5, characterized in that: The specific process of step S42 is as follows: Assume that the normal, stage 1 arc and stage 2 arc signals are , calculate the sample covariance between normal and different stage fault arcs and : ， ， Compute the standard deviation of the original signal: ， Calculate the Pearson correlation coefficient between normal and fault arcs at different stages: ， in, is the Pearson correlation coefficient of normal and stage 1 fault arc data, is the Pearson correlation coefficient of normal and stage 2 fault arc data.

8. The method for detecting DC arc faults based on Monte Carlo simulation feature optimization according to claim 1, characterized in that: The specific process of training the DC fault arc detection model in step S5 is as follows: obtaining several sets of frequency domain and time-frequency domain feature quantities of normal and different stage arcs as training samples of the BP neural network model. After the training is completed, the obtained model is optimized by genetic algorithm to construct the DC fault arc detection model.

9. A DC arc fault detection system based on Monte Carlo simulation feature optimization, using the DC arc fault detection method based on Monte Carlo simulation feature optimization as claimed in any one of claims 1 to 8, characterized in that: include: The sampling module is used to perform real-time sampling of the filtered electrical quantities on the DC side of the actual photovoltaic system inverter; Bandpass filter module, used to filter the electrical quantities on the DC side of the actual photovoltaic system inverter; Communication module, used to communicate with the host computer and transmit the current signal data obtained by sampling in the current cycle; A Fourier transform and empirical mode decomposition module is used to perform Fourier transform and empirical mode decomposition on several sets of received current signal data to extract frequency domain and time-frequency domain feature information; A feature extraction module is used to perform Monte Carlo simulation on the frequency domain and time-frequency domain feature information of several groups of current signals to extract the frequency domain and time-frequency domain feature quantities; The model training module is used to use the frequency domain and time-frequency domain feature quantities obtained by the feature extraction module as BP neural network model training samples, and to construct a DC fault arc detection model through model training and optimization using a genetic algorithm; The judgment module uses the DC fault arc detection model to judge the current signal obtained by the sampling module and detect the DC fault arc in real time.

Images