A fault identification method and system for photovoltaic power generation system

Abstract

The present invention belongs to the field of photovoltaic power generation safety technology, and specifically relates to a fault identification method and system for photovoltaic power generation systems. The method comprises: collecting real-time current signals from the photovoltaic power generation system and dividing them into multiple analysis windows; constructing a baseline model for characterizing normal operating conditions to detect abnormal disturbances and extract residual signals from the real-time collected current signals; and performing structural feature analysis on the residual signals from two dimensions, namely, energy distribution form and time domain oscillation form, to achieve fault identification of DC arcing faults. The present invention can intelligently adapt to the dynamic operating conditions of the photovoltaic system. By performing deep coupling analysis on the signal structure, it fundamentally reduces the false alarm rate and achieves high-reliability fault identification.

Description

Technical Field

[0001] The present invention relates to the field of photovoltaic power generation safety technology. More specifically, the present invention relates to a fault identification method and system for a photovoltaic power generation system. Background Art

[0002] The safe and stable operation of photovoltaic power generation systems is fundamental to their large-scale deployment. Among the numerous safety hazards, series arcing faults on the DC side have become the primary cause of fires in photovoltaic power plants due to their insidious nature, highly concentrated energy, and extremely high temperatures.

[0003] In the existing technology, the commonly used technical route for arc fault detection is: extracting the frequency domain characteristics of the current or voltage signal through time-frequency analysis tools such as Fourier transform or wavelet transform, mainly analyzing the distribution of its energy in different frequency bands, and then using machine learning classifiers to learn and judge these features to identify faults.

[0004] However, such methods generally face the dilemma of balancing sensitivity and reliability in practical applications. On the one hand, under the dynamic operating conditions of photovoltaic systems, some strong normal operating interference signals, especially the switching noise generated by the inverter when operating at different powers and loads, may have energy amplitudes and main frequency band distributions that highly overlap with the characteristics of weak arcing signals occurring in the early stages or at remote locations. These two signals show great similarity in the traditional energy feature dimension, making classifiers that rely solely on energy features extremely easy to confuse and difficult to effectively distinguish. On the other hand, traditional methods usually set a fixed energy threshold or train a globally invariant classification model to judge faults. However, the background noise level of photovoltaic systems under normal conditions is itself dynamically changing. For example, it will fluctuate significantly with changes in light intensity, ambient temperature, and load. This fixed diagnostic standard cannot adapt to the dynamically changing background. Summary of the Invention

[0005] In order to solve the technical problems in existing photovoltaic power generation systems, such as the lack of in-depth signal morphology recognition capability in DC arc fault diagnosis and the high false alarm rate caused by the inability to adapt to dynamic working conditions, the present invention provides solutions in the following aspects.

[0006] In a first aspect, the present invention provides a fault identification method for a photovoltaic power generation system, comprising: collecting real-time current signals of the photovoltaic power generation system and dividing them into multiple analysis windows; constructing a baseline model for characterizing normal operating conditions based on multiple analysis windows under historical normal operating conditions; processing the newly collected analysis window based on the baseline model to detect abnormal disturbances existing in the newly collected analysis window, and extracting a residual signal characterizing the abnormal disturbance, wherein the residual signal is the portion of the newly collected analysis window that cannot be explained by the baseline model; performing structural feature analysis on the residual signal from two dimensions, namely, energy distribution form and time domain oscillation form; and determining whether there is a DC arcing fault in the newly collected analysis window based on the results of the structural feature analysis, so as to realize fault identification of the photovoltaic power generation system.

[0007] By constructing a baseline model and comparing the real-time signal with it, the present invention shifts the diagnostic focus from the original signal, which is mixed with background noise and operating condition variations, to the residual signal, which contains only abnormal components. This significantly eliminates the interference of normal operating condition variations, provides a more ideal analysis target for subsequent precise diagnosis, and significantly improves the accuracy of fault diagnosis.

[0008] Preferably, the baseline model is a principal component analysis model, and the extraction of the residual signal characterizing the abnormal disturbance includes: reconstructing the newly acquired analysis window using the principal component analysis model to obtain a reconstructed signal; calculating the difference between the newly acquired analysis window and the reconstructed signal to obtain the residual signal characterizing the abnormal disturbance.

[0009] Preferably, the principal component analysis model is constructed in the following manner: principal component analysis is performed on a historical data matrix composed of multiple analysis windows under historical normal operating conditions to obtain multiple eigenvalues ​​and corresponding eigenvectors; according to a preset variance contribution rate threshold, the eigenvectors corresponding to the first several largest eigenvalues ​​are selected to form a projection matrix for signal reconstruction.

[0010] The present invention selects the principal components by setting the variance contribution rate threshold, which can achieve effective data dimensionality reduction while ensuring information fidelity, thereby ensuring the compactness and efficiency of the baseline model.

[0011] Preferably, the detecting of abnormal disturbances in the newly acquired analysis window comprises: calculating a square error between the newly acquired analysis window and the reconstructed signal as an abnormal deviation, wherein the abnormal deviation satisfies the relationship: ;in is the abnormal deviation, The first The value of the sampling points, is the first The value of the sampling points, is the total number of sampling points in the analysis window; when the abnormal deviation exceeds the preset primary screening threshold, it is determined that there is abnormal disturbance in the newly collected analysis window.

[0012] The present invention achieves preliminary screening of abnormal disturbances by defining the abnormal deviation and setting a preset threshold for judgment, effectively avoiding the limitations of fixed thresholds under dynamic working conditions.

[0013] Preferably, the structural feature analysis of the residual signal includes: calculating the spectral flatness of the residual signal to characterize the distribution form of its energy in the frequency domain; calculating the zero-crossing rate of the residual signal to characterize its oscillation form in the time domain.

[0014] The present invention analyzes two physically orthogonal features, namely spectrum flatness and zero-crossing rate, and dissects the intrinsic structure of abnormal signals from both frequency and time domain dimensions, thereby forming a comprehensive characterization of the fault morphology.

[0015] Preferably, the zero-crossing rate is calculated by counting the number of times the signs of adjacent sampling points in the residual signal sequence change, dividing the number by the total number of adjacent point pairs in the residual signal sequence, and using the obtained ratio as the zero-crossing rate.

[0016] Preferably, determining whether there is a DC arcing fault in the newly acquired analysis window based on the result of the structural feature analysis includes: comparing the calculated spectrum flatness with the spectrum flatness decision boundary; comparing the calculated zero-crossing rate with the zero-crossing rate decision boundary; when the spectrum flatness is greater than the spectrum flatness decision boundary and the zero-crossing rate is greater than the zero-crossing rate decision boundary, determining that a DC arcing fault exists.

[0017] The present invention establishes a decision boundary in a two-dimensional feature space and requires that two orthogonal features simultaneously meet the conditions before determining it as a fault, which greatly reduces the possibility of false alarms and ensures the reliability of fault diagnosis.

[0018] Preferably, the baseline model constructed to characterize normal operating conditions is dynamically constructed, including: real-time maintenance of a normal historical data pool that stores historical analysis windows that are judged to be normal within a preset time period; based on the data in the normal historical data pool, periodically updating the baseline model using principal component analysis.

[0019] In the present invention, the periodic dynamic update of the baseline model enables the normal standard of fault diagnosis to continuously evolve, thereby adapting to the changes in the working conditions of the photovoltaic system and ensuring the robustness of the diagnostic performance.

[0020] In a second aspect, the present invention provides a fault identification system for a photovoltaic power generation system, comprising a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned fault identification method for a photovoltaic power generation system is implemented.

[0021] By adopting the above technical solution, the above-mentioned fault identification method for photovoltaic power generation system is generated into a computer program and stored in the memory to be loaded and executed by the processor, so that a terminal device is made based on the memory and the processor for easy use.

[0022] Through layer-by-layer information purification and deep feature coupling, the present invention first achieves adaptive adaptation to dynamic operating conditions using a dynamic PCA model and extracts a pure residual signal. Subsequently, instead of analyzing the original signal, the residual signal is directly analyzed, and a joint decision is made based on two orthogonal features: spectral flatness and zero-crossing rate, fundamentally ensuring diagnostic reliability. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The above and other objects, features and advantages of the exemplary embodiments of the present invention will become readily understood by reading the following detailed description with reference to the accompanying drawings. In the accompanying drawings, several embodiments of the present invention are shown in an illustrative and non-limiting manner, and the same or corresponding reference numerals represent the same or corresponding parts, wherein:

[0024] Figure 1 is a flow chart schematically illustrating a fault identification method for a photovoltaic power generation system in the present invention;

[0025] Figure 2 is a schematic diagram schematically showing high-frequency current ripple signals under different working conditions;

[0026] Figure 3 is a schematic diagram schematically showing the change of the abnormal deviation GP value with the analysis window;

[0027] Figure 4 Schematic diagram of a decision plane based on a two-dimensional feature space. DETAILED DESCRIPTION

[0028] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work shall fall within the scope of protection of the present invention.

[0029] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0030] The embodiment of the present invention discloses a fault identification method for a photovoltaic power generation system, referring to Figure 1 , including steps S1 to S5:

[0031] S1. Collect the real-time current signal of the photovoltaic power generation system and divide it into multiple analysis windows.

[0032] In one optional embodiment, monitoring equipment is first installed at key nodes on the DC side of the photovoltaic power generation system. Specifically, a high-frequency current sensor is installed at the DC combiner box outlet of the photovoltaic array or the DC input of the inverter. This sensor has a high sampling capacity, exemplarily using a 1MHz sampling rate, to ensure that it can capture high-frequency details generated by arcing faults.

[0033] In this optional embodiment, the real-time current signal collected contains two parts: one part is a stable DC component with a larger value, and the other part is various high-frequency disturbances superimposed on it. Therefore, the original current signal can be processed by a digital high-pass filter to filter out the DC component and retain only the high-frequency current ripple signal. Furthermore, the present invention divides the continuous ripple signal stream into multiple non-overlapping analysis windows of fixed length. For example, the time length of each window can be set to 100 milliseconds. According to the sampling rate of 1MHz, each analysis window contains sampling points.

[0034] like Figure 2 The figure below schematically illustrates the waveforms of high-frequency current ripple signals under three different operating conditions. The dashed line represents normal background noise, which has a small amplitude and fluctuates smoothly along the horizontal axis. The thick solid line represents strong interference signals, such as the switching noise generated by an inverter operating under a specific load, which exhibits strong periodicity and large amplitude. The thin solid line represents a DC arc fault, whose waveform is significantly random and full of severe glitches.

[0035] In this way, through high-frequency sampling and window processing, the continuous current signal can be converted into discrete data units suitable for subsequent model analysis, laying the foundation for accurate extraction of fault characteristics.

[0036] S2. Based on multiple analysis windows under historical normal operating conditions, a baseline model for characterizing normal operating conditions is constructed.

[0037] In an optional embodiment, a normal historical data pool can be maintained in real time in the photovoltaic power generation system. This normal historical data pool stores all the analysis windows that are judged to be normal within a fixed period of time in the past, such as 10 minutes. Assume that there are m normal analysis windows in this 10 minutes, each analysis window includes n sampling points, and these windows can be arranged to form a dimensional historical data matrix.

[0038] Next, principal component analysis is performed on the historical data matrix X. Specifically, the first step is to calculate the mean vector of the historical data matrix, which represents the average shape of the historical normal signal. The second step is to calculate the covariance matrix of the historical data matrix X based on the obtained mean vector and perform eigenvalue decomposition on the covariance matrix to obtain n eigenvalues ​​and their corresponding eigenvectors. These eigenvectors constitute a set of orthogonal bases in the signal space. The third step is to select the principal components. According to a preset variance contribution rate threshold, such as 95%, eigenvalues ​​are selected from large to small and accumulated until the proportion of the accumulated result to the sum of the total eigenvalues ​​exceeds the variance contribution rate threshold for the first time. The eigenvectors corresponding to these eigenvalues ​​that meet the conditions are combined to form a projection matrix. Finally, this dynamically calculated projection matrix and the corresponding mean vector together constitute a baseline model that can characterize the current normal operating conditions.

[0039] Furthermore, the baseline model defines a normal subspace with fewer dimensions than the original. Any normal signal window projected into this space and then reconstructed back should experience minimal information loss. Because the data pool is continuously updated, the model is also periodically updated, enabling adaptation to dynamic operating conditions.

[0040] In this way, by constructing a dynamic and adaptive baseline model, a standard can be obtained that can reflect the current normal state at any time, providing a reliable benchmark for subsequent distinction between normal and abnormal conditions.

[0041] S3. Based on the baseline model, the newly acquired analysis window is processed to detect abnormal disturbances in the newly acquired analysis window, and a residual signal representing the abnormal disturbances is extracted, wherein the residual signal is the portion of the newly acquired analysis window that cannot be explained by the baseline model.

[0042] In an optional embodiment, for each newly acquired analysis window , it can be reconstructed by the obtained baseline model. The role of reconstruction is to predict the normal data rules summarized based on historical data. The theoretical normal form that should be grown.

[0043] For example, the reconstruction process is: let the projection matrix corresponding to the baseline model be P, and the mean vector be , then first Centralized processing: Then Projection to the principal subspace: ,in yes Representation in low-dimensional space; then reconstruct the original space from the principal subspace: ,in is the reconstructed signal The result after centering, so by adding the corresponding mean vector, the final reconstructed signal can be obtained .

[0044] Get the reconstructed signal After that, the square error between the newly acquired analysis window and the reconstructed signal can be calculated as the abnormal deviation. Satisfies the relationship:

[0045]

[0046] in The first The value of the sampling points, To reconstruct the signal The value of the sampling points, is the total number of sampling points in the analysis window.

[0047] Furthermore, a preset initial screening threshold Th1 can be set, wherein the initial screening threshold can be set dynamically based on historical statistics. For example, the value that 90% of the abnormal deviations in the past 100 normal analysis windows are less than can be taken as the initial screening threshold. When the abnormal deviation exceeds the preset initial screening threshold, it is determined that there is abnormal disturbance in the newly collected analysis window. Figure 3 As shown in the figure, it is a schematic diagram showing the change of the abnormal deviation GP value with the analysis window. Assuming the initial screening threshold is 1.14, it can be seen that analysis windows 1 to 5 are judged to be normal, while the abnormal deviation values ​​caused by the strong interference of windows 6 to 10 and the arcing fault of windows 16 to 20 are higher than the threshold and are judged to be abnormal.

[0048] In this optional embodiment, for all windows determined to be abnormal, the difference between the newly acquired analysis window and the reconstructed signal can be calculated as a residual signal. This residual signal is actually the abnormal signal in the original signal that cannot be understood and explained by the normal baseline model. Therefore, the subsequent analysis process does not need to analyze the original signal, but only needs to analyze this abnormal signal.

[0049] In this way, by calculating the square error between the original signal and the reconstructed signal and extracting the residual signal, abnormal disturbances can be preliminarily screened out, and the corresponding abnormal signal can be provided for subsequent accurate diagnosis.

[0050] S4. Analyze the structural characteristics of the residual signal from two dimensions: energy distribution and time domain oscillation.

[0051] In an optional embodiment, spectral flatness measure (SFM) can be used to describe the distribution of signal energy in the frequency domain because it measures the uniformity of energy distribution. A signal with an uneven spectrum, where energy is concentrated in a few frequencies, such as periodic interference, has a low flatness; a signal with a flat spectrum, where energy is evenly distributed across all frequencies, such as white noise, has a high flatness.

[0052] Even if a strong interference signal is filtered out, its residual signal may still retain periodicity, resulting in some weak energy concentration points in the spectrum, resulting in low spectral flatness. However, DC arcing is essentially a random discharge process, and its residual signal is close to white noise. The energy is evenly distributed across the frequency band, resulting in high spectral flatness.

[0053] It is worth noting that spectral flatness is calculated by first performing a Fourier transform on the residual signal to obtain its power spectrum P(f). Spectral flatness SFM is then calculated by taking the ratio of the geometric mean to the arithmetic mean of the power spectrum.

[0054] For example, assuming that after Fourier transform of a residual signal, the energy values ​​of its power spectrum at three discrete frequency points are [10, 10, 10], then the arithmetic mean of its power spectrum is: (10+10+10) / 3=10, and the geometric mean of its power spectrum is: , so the spectrum flatness SFM = geometric mean / arithmetic mean = 10 / 10 = 1. The result is 1, which means that the energy is completely flat in the frequency domain.

[0055] Furthermore, because the zero-crossing rate (ZCR) measures the intensity of a time-domain waveform's oscillations, it can be used to measure how often a signal waveform crosses the zero-value horizontal line. A smooth, low-frequency waveform will have a low zero-crossing rate; whereas a waveform characterized by high-frequency oscillations and intense jitter will have a very high zero-crossing rate.

[0056] Since the residual signal of strong interference is usually composed of smooth harmonics remaining after filtering out the fundamental wave, its waveform is relatively flat and its zero-crossing rate is low. However, the residual signal of DC arcing is full of random, intense, and high-frequency glitches, frequently crossing the zero axis, resulting in an extremely high zero-crossing rate. It is worth noting that the zero-crossing rate is calculated by directly counting the number of times the sign of two adjacent sampling points in the residual sequence changes, and then dividing it by the total number of adjacent point pairs in the residual sequence to obtain the normalized zero-crossing rate.

[0057] For example, suppose the statistical residual sequence is [0.5, -0.2, 0.8, -0.1, -0.3, 0.6], the corresponding symbol sequence is [+, -, +, -, -, +], the symbol change occurs at the 1st, 2nd, 3rd, and 5th positions, a total of 4 times, the residual sequence length is 6, there are 5 adjacent pairs, so the zero crossing rate ZCR is .

[0058] Finally, the calculated features that can describe the signal structure from both frequency domain and time domain perspectives are combined into a two-dimensional structural fingerprint feature vector [SFM, ZCR].

[0059] In this way, by calculating the spectral flatness and zero-crossing rate of the residual signal, a two-dimensional feature vector is obtained that can comprehensively and quantitatively describe the intrinsic structure of the abnormal signal.

[0060] S5. Based on the results of the structural feature analysis, determine whether there is a DC arc fault in the newly acquired analysis window, so as to realize fault identification of the photovoltaic power generation system.

[0061] In an optional embodiment, if Figure 4 As shown, a two-dimensional decision plane can be established with spectrum flatness as the X-axis and zero-crossing rate as the Y-axis. Figure 3 The residual signal of the strong interference signal corresponding to the analysis window 6 to 10 still has a certain regularity, and the calculated SFM and ZCR are both low. Therefore, the corresponding feature points will stably gather in the lower left corner of the decision plane, as shown in Figure 4 As shown by the hollow circles in .

[0062] for Figure 3 The residual signal of the DC arc fault signal corresponding to the analysis window 16 to 20 has a high degree of randomness and broadband characteristics, so the calculated SFM and ZCR are extremely high. Therefore, the corresponding feature points will stably gather in the upper right corner of the decision plane, such as Figure 4 as indicated by the cross mark in the .

[0063] In this optional embodiment, the spectrum flatness decision boundary and the zero-crossing rate decision boundary can be set separately. These two decision boundaries are not fixed, but can be dynamically generated based on long-term statistical analysis of the residual signals of all normal windows in the normal historical data pool. For example, the 99% quantile or mean of the SFM and ZCR values ​​of the historical normal residual signal plus three times the standard deviation can be used as the dynamic decision boundary. For example, Figure 4 As shown, the corresponding spectrum flatness decision boundary and zero-crossing rate decision boundary can be set to 0.5 ( Figure 4 The numerical values ​​in dashed lines) and 0.3 ( Figure 4 horizontal dashed line in the .

[0064] Furthermore, when the spectrum flatness is greater than the spectrum flatness decision boundary and the zero-crossing rate is greater than the zero-crossing rate decision boundary, it is determined that a DC arc fault exists; in other cases, no alarm is required, thereby effectively avoiding false alarms.

[0065] In this way, by performing region division and decision-making in the two-dimensional orthogonal feature space and making joint judgments on two unrelated features, the possibility of disguise of interference signals is fundamentally eliminated, and reliable identification of DC arc faults is achieved.

[0066] An embodiment of the present invention further discloses a fault identification system for a photovoltaic power generation system, comprising a processor and a memory, wherein the memory stores computer program instructions. When the computer program instructions are executed by the processor, a fault identification method for a photovoltaic power generation system according to the present invention is implemented.

[0067] The above system also includes other components well known to those skilled in the art, such as a communication bus and a communication interface. The configuration and functions of these components are known in the art and will not be described in detail here.

[0068] In the description of this specification, “a plurality” or “several” means at least two, for example, two, three or more, etc., unless otherwise clearly and specifically defined.

[0069] While this specification has shown and described several embodiments of the present invention, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Numerous modifications, variations, and alternatives will occur to those skilled in the art without departing from the concept and spirit of the present invention. It should be understood that in practicing the present invention, alternatives to the embodiments of the present invention described herein may be employed.

Claims

1. A fault identification method for a photovoltaic power generation system, characterized in that: include: Collect the real-time current signal of the photovoltaic power generation system and divide it into multiple analysis windows; Constructing a baseline model for characterizing normal operating conditions based on a plurality of analysis windows under historical normal operating conditions; Based on the baseline model, a newly acquired analysis window is processed to detect abnormal disturbances present in the newly acquired analysis window, and a residual signal representing the abnormal disturbance is extracted, wherein the residual signal is a portion of the newly acquired analysis window that cannot be explained by the baseline model; the baseline model is a principal component analysis model, and extracting the residual signal representing the abnormal disturbance includes: reconstructing the newly acquired analysis window using the principal component analysis model to obtain a reconstructed signal; and calculating a difference between the newly acquired analysis window and the reconstructed signal to obtain a residual signal representing the abnormal disturbance; The residual signal is subjected to structural feature analysis from two dimensions, namely, energy distribution form and time domain oscillation form, including: calculating the spectral flatness of the residual signal to characterize the energy distribution form in the frequency domain; calculating the zero-crossing rate of the residual signal to characterize the oscillation form in the time domain; Determining whether a DC arc fault exists in the newly acquired analysis window based on the results of the structural feature analysis to achieve fault identification of the photovoltaic power generation system includes: comparing the calculated spectrum flatness with a spectrum flatness decision boundary; comparing the calculated zero-crossing rate with a zero-crossing rate decision boundary; and determining that a DC arc fault exists when the spectrum flatness is greater than the spectrum flatness decision boundary and the zero-crossing rate is greater than the zero-crossing rate decision boundary.

2. A fault identification method for a photovoltaic power generation system according to claim 1, characterized in that: The principal component analysis model is constructed as follows: Perform principal component analysis on the historical data matrix composed of multiple analysis windows under historical normal operating conditions to obtain multiple eigenvalues ​​and corresponding eigenvectors; According to the preset variance contribution rate threshold, the eigenvectors corresponding to the first several largest eigenvalues ​​are selected to form a projection matrix for signal reconstruction.

3. A fault identification method for a photovoltaic power generation system according to claim 1, characterized in that: The detecting of abnormal disturbances existing in the newly acquired analysis window includes: The square error between the newly acquired analysis window and the reconstructed signal is calculated as the abnormal deviation, and the abnormal deviation satisfies the relationship: in is the abnormal deviation, The first The value of the sampling points, is the first The value of the sampling points, is the total number of sampling points in the analysis window; When the abnormal deviation exceeds a preset primary screening threshold, it is determined that abnormal disturbance exists in the newly collected analysis window.

4. A fault identification method for a photovoltaic power generation system according to claim 1, characterized in that: The zero-crossing rate is calculated by counting the number of times the signs of adjacent sampling points in the residual signal sequence change, dividing the number by the total number of adjacent point pairs in the residual signal sequence, and using the obtained ratio as the zero-crossing rate.

5. A fault identification method for a photovoltaic power generation system according to claim 1, characterized in that: The baseline model used to characterize normal operating conditions is constructed dynamically and includes: Maintain in real time a normal historical data pool that stores historical analysis windows that are determined to be normal within a preset time period; The baseline model is periodically updated using principal component analysis based on the data in the normal historical data pool.

6. A fault identification system for a photovoltaic power generation system, characterized in that: include: A processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, a fault identification method for a photovoltaic power generation system according to any one of claims 1 to 5 is implemented.

Images