Photovoltaic module current mismatch fault detection method based on wavelet analysis and deep learning

Abstract

The invention discloses a photovoltaic module current mismatch fault detection method based on wavelet analysis and deep learning. High-precision distinguishing of multiple types of current mismatch faults is achieved. The method comprises the following steps: firstly, carrying out IV data detection on a to-be-diagnosed photovoltaic module by utilizing a photovoltaic IV detector, and removing abnormal values in a detection result; secondly, wavelet analysis processing is carried out on the preprocessed data, and effective signals and high-frequency noise are separated; and reconstructing the effective frequency band signal to obtain a smooth current signal. And then, constructing a CNN-LSTM hybrid neural network, automatically capturing local shape features and a time domain dependency relationship of the IV curve, and finally outputting depth features. And fusing the depth feature with the concave-convex detection feature value of the IV curve, generating a comprehensive fault index, and if a current mismatch fault occurs, positioning a curve inflection point and extracting and detecting the fault feature of IV data, thereby distinguishing three types of faults including local shadow, hot spot effect and cellular rupture.

Description

Technical Field

[0001] This invention relates to a method for detecting current mismatch faults in photovoltaic modules based on wavelet analysis and deep learning, belonging to the field of photovoltaic module fault detection technology. Background Technology

[0002] With the large-scale application of photovoltaic (PV) power generation technology, PV modules, as the core unit of the system, directly affect the overall power generation efficiency and lifespan. However, PV modules are exposed to harsh outdoor environments such as extreme temperatures, humidity, and ultraviolet radiation for extended periods, leading to frequent failures. Among these, current mismatch faults, such as localized shading, hotspot effects, and cell cracks, are particularly prominent. These faults cause a decrease in the module's output current, forming a stepped characteristic on the IV curve. This not only significantly reduces output power but may also accelerate module aging due to localized temperature increases, and even induce safety hazards such as fires.

[0003] Current photovoltaic module fault diagnosis methods can be mainly divided into the following categories:

[0004] 1. Visual inspection: Relies on computer vision algorithms to identify surface defects, but can only detect visible faults and is powerless against hidden problems (such as internal circuit abnormalities), and has low detection efficiency.

[0005] 2. Electrical parameter analysis method: This method detects anomalies by monitoring parameters such as voltage and current, but it requires the deployment of a large number of sensors, resulting in high hardware costs and susceptibility to environmental noise interference.

[0006] 3. Diagnostic techniques based on mathematical models: These rely on precise component parameters, but parameter drift in real-world environments can lead to misdiagnosis.

[0007] 4. Infrared thermal imaging: It locates fault areas by measuring temperature field distribution. Although it has high accuracy, the equipment is expensive and difficult to popularize in residential and small distributed systems.

[0008] 5. Intelligent detection method: It uses data-driven methods to classify faults, but it relies on large-scale labeled data and expert knowledge, and its generalization ability is limited.

[0009] Furthermore, the aforementioned methods primarily focus on system-level faults, making it difficult to achieve refined component-level diagnosis. In the field of current mismatch fault detection, existing research is generally based on IV curve feature analysis, but is mostly limited to the identification of single fault types. For example, shadow faults are detected through the stepped features of the IV curve, hot spot faults rely on current change rate analysis, and mature diagnostic methods exist for cell ruptured batteries. Since all three types of faults will cause stepped features in the IV curve, existing methods are unable to effectively distinguish specific fault types and cannot support precise operation and maintenance.

[0010] Therefore, there is an urgent need to develop an online fault diagnosis method suitable for distributed photovoltaic systems to address the shortcomings of existing technologies in terms of cost, accuracy, and ability to distinguish multiple fault types, and to provide technical support for the intelligent operation and maintenance of photovoltaic systems. Summary of the Invention

[0011] Technical problem: To address the aforementioned existing technologies, a photovoltaic module current mismatch fault detection method based on wavelet analysis and deep learning is proposed, which can achieve high-precision differentiation of multiple types of current mismatch faults.

[0012] Technical solution: A photovoltaic module current mismatch fault detection method based on wavelet analysis and deep learning, comprising the following steps:

[0013] Step 1: Use a photovoltaic IV tester to collect IV data of the photovoltaic module to be diagnosed, and preprocess the collected data to remove outliers.

[0014] Step 2: Perform wavelet analysis on the preprocessed data to separate the effective signal from high-frequency noise; then reconstruct the effective frequency band signal to obtain a smoothed current signal;

[0015] Step 3: Construct a CNN-LSTM hybrid neural network to automatically capture the local shape features and temporal dependencies of the IV curve, and finally output the deep feature F. deep ;

[0016] Step 4: Fuse the depth features F deep Concavity / convexity detection feature value ΔI of IV curve k+1 Generate a comprehensive fault index F fusion ;

[0017] Step 5: If a current mismatch fault occurs, locate the inflection point of the curve and extract the fault feature ΔK from the IV detection data;

[0018] Step 6: Compare the detected fault feature ΔK with the fault judgment threshold to distinguish between three types of faults: local shadow, hot spot effect, and cell breakage.

[0019] Furthermore, in step 1, the photovoltaic IV testing instrument samples and obtains IV data of the series photovoltaic modules, including several sets of data points with different voltages and currents, and arranges the data points in descending order of voltage; the sorted data points are then judged, and when an I... n+1 ≤I n At that time, I n+1 If a value is identified as an outlier, the data point with that outlier value is removed from the sorted data set.

[0020] Furthermore, step 2 specifically includes: performing wavelet decomposition on the data after outlier removal using wavelet basis functions; setting the scale parameter j=4 for wavelet decomposition based on the 32 data samplings per second of the photovoltaic IV detector; and decomposing the signal into low-frequency approximate components and high-frequency detail components through j-level decomposition; then reconstructing the effective frequency band signal; the frequency band selection strategy for reconstruction is: discarding the high-frequency noise band and only reconstructing the mid-to-low frequency effective frequency band, retaining the step characteristics and inflection point information of the IV curve, thereby obtaining the denoised smooth current signal I. smooth .

[0021] Furthermore, step 3 specifically includes:

[0022] Smooth the current signal I smooth The data points consisting of the corresponding voltage signals are converted into a two-dimensional input matrix X:

[0023] Extracting local spatial features using one-dimensional convolutional layers:

[0024] F conv =ReLU(W conv *X+b conv )

[0025] In the formula, F conv The convolution output features are represented by ReLU, and W is the activation function. conv b represents the kernel weights. conv The bias vector of the convolutional layer;

[0026] Preserve salient features of the data through max pooling:

[0027] F pool =MaxPool(F conv )

[0028] In the formula, F pool Maxpool is the maximum pooling function for pooling output features.

[0029] The output of the pooling layer is flattened and then input into the LSTM network:

[0030]

[0031] H = [h1, h2, ..., h 16 ]

[0032] In the formula, Let h be the eigenvector at position t, where t = 1 to 16. t H represents the hidden state at time t, containing 32 units, and H is the LSTM output.

[0033] Deep features F are generated through a fully connected layer. deep :

[0034] F deep =W s tanh(W f flatten(H)+b f )

[0035] In the formula, W f Let b be the weight coefficient matrix of the fully connected layer. f W is the bias vector. s This is the weight vector.

[0036] Furthermore, step 4 specifically includes:

[0037] For the (k+1)th data point (U) on the IV curve k+1 ,I k+1 ), and two adjacent data points (U k ,I k ) and (U k+2 ,I k+2 Define variable λ k+1 :

[0038]

[0039] Define the feature value for concavity / convexity detection as ΔI k+1 :

[0040] ΔI k+1 =I k+1 -[λ k+1 I k +(1-λ k+1 )I k+2 ]

[0041] The concave-convexity detection feature value ΔI k+1 With depth features F deep Combining these factors, the comprehensive failure index F is finally obtained. fusion :

[0042] F fusion =αΔI k+1 +βF deep

[0043] In the formula, α and β are the weights corresponding to the two eigenvalues ​​respectively;

[0044] Based on the calculated F fusion The specific numerical value and the threshold relationship, when F fusion >When this threshold is reached, the photovoltaic module is considered to have a current mismatch fault.

[0045] Furthermore, step 5 specifically includes:

[0046] The concavity / convexity detection feature value ΔI calculated in step 4 k+1 The positive peak value corresponds to the upward convex inflection point of the IV curve, that is, the boundary point of the step-up edge (U). k+1 ,I k+1 ); Traverse ΔI k+1 The sequence is used to extract local maxima, and the corresponding voltage and current coordinates are the inflection points.

[0047] If the inflection point is (U) k+1 ,I k+1 If the inflection point is selected, then the three data points P1(U) after the inflection point are chosen. k ,I k P2(U) k-1 ,I k-1 ), P3(U k-2 ,I k-2 ), calculate the absolute values ​​of the slopes at points P1 and P2, and denote them as K1 and K2 respectively:

[0048]

[0049] The slope ratio is used as a fault characteristic ΔK, which is a sensitivity indicator of faults.

[0050]

[0051] Furthermore, step 6 specifically includes: if K1 and K2 tend to 0, it indicates that the photovoltaic module has a local shading fault; if ΔK approaches 1, it indicates that the photovoltaic module has a hot spot fault; if ΔK>1, it indicates that the photovoltaic module has a cell breakage fault.

[0052] Beneficial effects: This invention achieves "high-precision detection, type differentiation, and accurate location" of photovoltaic module current mismatch faults, demonstrating significant advantages in improving diagnostic efficiency, reducing operation and maintenance costs, and enhancing system security. It provides core technical support for the intelligent operation and maintenance of photovoltaic power plants and is of great significance for promoting the large-scale application of photovoltaic power generation technology. Details are as follows:

[0053] I. High-precision differentiation capability for multiple types of faults

[0054] 1. Overcoming the limitations of single fault diagnosis

[0055] Traditional methods often focus on a single fault type (such as identifying only shadows or hot spots), while this invention achieves unified detection and accurate differentiation of three types of current mismatch faults—local shadows, hot spot effects, and cell breakage—by integrating wavelet analysis, CNN-LSTM networks, and slope feature analysis. Experimental results show that the fault diagnosis accuracy exceeds 98%, with zero false positives in 40 sets of cell breakage data, and only 2 missed cases in 60 sets of hot spot faults and 100 sets of shadow faults, solving the classification problem of "same shape, different faults" in traditional techniques.

[0056] 2. Feature fusion enhances discriminative ability

[0057] By locating the inflection point of the IV curve through concavity / convexity detection and combining it with the deep features extracted by CNN-LSTM, a dual verification mechanism of "local geometric features + global temporal features" is formed, which effectively amplifies the differences in slope features among different fault types. For example, ΔK≈1 for hotspot faults and ΔK>1 for cell breakage, thus achieving quantitative differentiation of fault types.

[0058] II. Technical Advantages of Data Preprocessing and Feature Extraction

[0059] 1. Wavelet analysis improves data quality

[0060] Wavelet analysis is employed to perform multi-scale decomposition and reconstruction of IV data, effectively separating high-frequency noise from effective signals while preserving key fault features such as step inflection points. Compared to traditional filtering methods, this technique avoids blurring of curve shapes, providing high-quality input for subsequent deep learning modeling and enabling CNN-LSTM networks to accurately capture the temporal patterns and local shape features of current changes.

[0061] 2. Automated Feature Engineering for CNN-LSTM Hybrid Networks

[0062] One-dimensional convolutional layers extract spatial features such as local slopes and inflection points from the IV curve, while LSTM layers capture the temporal dependence of current on voltage. This replaces the limitations of manually designed features in traditional methods and enables adaptive learning of fault features. For example, for the convex function shape of cell breakage faults, the network can automatically extract the slope-increasing feature without the need for manually preset rules.

[0063] III. Improvement of Engineering Application and Operation and Maintenance Efficiency

[0064] 1. Online monitoring applicable to distributed photovoltaic systems

[0065] This method acquires data based on an IV detector, eliminating the need to deploy a large number of sensors or expensive infrared equipment. It has low hardware costs, and the detection process can complete the acquisition of 32 sets of data within 1 second, making it suitable for online diagnostics of residential and small distributed systems.

[0066] 2. Promote intelligent operation and maintenance and reduce security risks

[0067] Automated fault detection replaces manual inspection, improving efficiency while avoiding reliance on experience. It can promptly detect hidden faults (such as early cell breakage), reducing component aging and fire hazards caused by localized temperature rises. Experimental verification shows that this method can effectively identify current mismatch problems that lead to a decrease in output power, providing technical support for the intelligent management of photovoltaic systems.

[0068] IV. Environmental Adaptability and Robust Design

[0069] 1. Noise resistance and cross-condition capability

[0070] Wavelet denoising and the slope ratio (ΔK) design reduce the interference of environmental factors on characteristic parameters. For example, ΔK eliminates the influence of operating condition fluctuations on the absolute slope through relative slope changes, making the fault discrimination threshold universal and eliminating the need for dynamic adjustment based on real-time environment.

[0071] 2. Compatibility with multiple fault scenarios

[0072] This method can be used for single-step fault detection as well as for cases where multiple substrings are faulty. By extracting features from multiple inflection points and calculating statistics, it improves the diagnostic reliability of complex fault scenarios and avoids misjudgment based on a single inflection point.

[0073] V. Technological Innovation and Industry Value

[0074] 1. Innovative architecture integrating multiple technologies

[0075] For the first time, wavelet analysis, CNN-LSTM network and IV curve geometric feature analysis are combined to form a complete technology chain of "data preprocessing-deep feature extraction-fault classification", which breaks through the limitations of traditional electrical parameter analysis, infrared imaging and other methods, and provides a new technical path for photovoltaic module fault detection.

[0076] 2. Promote the improvement of photovoltaic energy utilization rate

[0077] By accurately identifying current mismatch faults, maintenance personnel can be guided to perform targeted operations such as component cleaning and cell replacement, reducing the degradation of power generation efficiency caused by the fault. Experimental data suggests that this method can increase the average annual power generation of photovoltaic systems by 3%-5%, demonstrating significant practical value in promoting the efficient utilization of photovoltaic energy. Attached Figure Description

[0078] Figure 1 This is a flowchart of the method of the present invention;

[0079] Figure 2 This example shows a comparison of the results before and after data preprocessing.

[0080] Figure 3 The CNN-LSTM network architecture used in the method of this invention;

[0081] Figure 4 The photovoltaic IV detector used in the method of this invention;

[0082] Figure 5 This is a schematic diagram of the fault diagnosis results for an example. Detailed Implementation

[0083] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. The following embodiments are explanations of the present invention, but the present invention is not limited to the following embodiments.

[0084] A method for detecting current mismatch faults in photovoltaic modules based on wavelet analysis and deep learning is illustrated in the flowchart below. Figure 1 As shown, it includes the following steps:

[0085] Step 1: Use a photovoltaic IV tester to collect IV data from the photovoltaic module to be diagnosed, and preprocess the collected data to remove outliers, providing high-quality basic data for subsequent fault analysis. Specifically, this includes:

[0086] In this embodiment, the photovoltaic IV testing instrument can obtain IV data of the series photovoltaic module within 1 second, including 32 sets of data points with different voltages and currents, denoted as (U i I i ), where i = 0 to 31. Each set of data corresponds to the output current under different operating voltages, covering the entire operating state of the component from open circuit to short circuit.

[0087] Will (U i I i The data is sorted in descending order of voltage, resulting in 32 pairs of sorted voltage and current data, denoted as (U... n I n ), where n = 0 to 31. Theoretically, in the IV curve of a photovoltaic module, the current monotonically decreases as the voltage increases, approaching 0 in the open-circuit state (highest voltage) and reaching its maximum value in the short-circuit state (voltage approaching 0). Arranging the data points in descending order ensures that the data points conform to the physical law of "higher voltage - lower current", facilitating subsequent outlier detection.

[0088] However, due to factors such as sensor noise and environmental interference during data acquisition, data fluctuations can be significant, resulting in outliers in the collected data. These outliers can severely affect the accuracy of subsequent fault feature extraction. Therefore, data preprocessing is necessary to remove the small number of outliers present in the data.

[0089] Based on the monotonicity of theoretical current decreasing with increasing voltage, if the current value at a certain data point does not conform to this rule, it is judged as an outlier. Specifically, for the sorted data points (U... n I n ) Make a judgment, and when the case of equation (1) exists:

[0090] I n+1 ≤I n (1)

[0091] If the current does not increase as the voltage decreases, it indicates that the data does not conform to the monotonically increasing characteristic of the I-V curve. n+1 If a value is identified as an outlier, it should be removed from the sorted data set, and then interpolation should be performed to keep the total number of data points unchanged.

[0092] The comparison results before and after data preprocessing in this embodiment are as follows: Figure 2 As shown.

[0093] Step 2: Perform wavelet analysis on the preprocessed data to separate the effective signal from high-frequency noise; then reconstruct the effective frequency band signal to obtain a smooth current signal, providing a clean signal basis for subsequent feature extraction.

[0094] Photovoltaic IV data is susceptible to environmental interference (such as electromagnetic noise and temperature fluctuations) and errors in acquisition equipment, containing a large amount of high-frequency noise. Wavelet analysis, as a time-frequency domain analysis tool, can decompose the signal into different frequency sub-bands, accurately separating noise from effective features, and is particularly suitable for processing non-stationary electrical signals.

[0095] The theoretical form of the continuous wavelet transform is shown in equation (2):

[0096] W j,k =∫I(t)ψ j,k (t)dt (2)

[0097] In the formula, W j,k Here, is the wavelet packet coefficient, j is the scale parameter (j = 4 in this embodiment), k is the position index, representing the position of the wavelet function on the time axis, I(t) is the current data, and ψ is the wavelet packet coefficient. j,k The wavelet basis function is obtained by scaling and translating the mother wavelet ψ(t), as shown in equation (3).

[0098] ψ j,k (t)=2 -j / 2 ψ(2 -j tk) (3)

[0099] Through j-level decomposition, the signal is decomposed into low-frequency approximate components and high-frequency detail components. The high-frequency components mainly contain noise, while the low-frequency components retain the main characteristics of the IV curve. The optimal wavelet basis, such as DB4, DB6, or Sym4, can be dynamically selected based on the IV curve characteristics of different photovoltaic modules to improve denoising accuracy.

[0100] In the formula, t is a continuous variable. In actual discrete signal processing, the discrete wavelet transform is used, and the integral of equation (2) is converted into a summation in the discrete case, as shown in equation (4):

[0101]

[0102] After wavelet decomposition, the effective frequency band signal is reconstructed to obtain effective current data. The reconstruction method is shown in equation (5).

[0103]

[0104] In the formula, I smooth The signal is a smoothed current signal after denoising, and Ω represents the effective frequency band set. The frequency band selection strategy is as follows: discard high-frequency noise bands, such as the highest frequency band corresponding to j=4, and only reconstruct the effective mid-to-low frequency bands, retaining the step characteristics and inflection point information of the IV curve.

[0105] The theoretical maximum decomposition level for 32 sampling points is 5. In this embodiment, j = 0 to 4 is selected, i.e., 4-level decomposition is performed, which can avoid excessive decomposition leading to feature loss. The noise in photovoltaic IV data is mostly concentrated in the high-frequency band, such as above 100Hz. 4-level decomposition can decompose the signal into different frequency bands and accurately locate the noise frequency band.

[0106] The smoothed signal I output in this step smooth As input to the CNN-LSTM network in step 3, it ensures that: the CNN convolutional layer can accurately extract the local shape features of the IV curve, such as the step slope; and the LSTM layer can capture the temporal dependence of current changes, avoiding misjudgment of temporal features caused by noise, such as false inflection points; in addition, the smoothed signal ensures that the LSTM can accurately capture the true trend of current changes.

[0107] Wavelet analysis preserves abrupt change points in the IV curve, such as step inflection points, while denoising. These points correspond to key features of current mismatch faults, providing a basis for concavity / convexity detection in step 4 and fault classification in step 6.

[0108] Furthermore, the scale parameter j can be used to estimate the noise through power spectral density, and then adaptively determine the high-frequency bands that need to be discarded, thus determining the specific value of j.

[0109] Step 3: Construct a CNN-LSTM hybrid neural network to perform deep feature extraction on the IV data processed in Step 2, automatically extract the temporal fault features of the IV curve, and automatically capture the local shape features and temporal dependencies of the IV curve, providing multi-dimensional feature representation for subsequent fault classification.

[0110] The current mismatch fault characteristics of the IV curve include both spatial features such as local step inflection points and temporal patterns of current variation with voltage. A single CNN or LSTM network is insufficient to effectively extract both types of features simultaneously. A CNN-LSTM hybrid network architecture can achieve complementary advantages and improve the comprehensiveness of fault feature extraction.

[0111] The CNN-LSTM hybrid network architecture used is as follows: Figure 3As shown, smooth the data {(U n ,I smooth,n Convert |n=0~31} into a two-dimensional input matrix:

[0112]

[0113] That is, voltage U n As a one-dimensional input, voltage U n The corresponding current I smooth,n As another dimension, it forms a two-dimensional matrix structure similar to an image, thereby adapting to the two-dimensional convolution operation of CNN and transforming the "voltage-current" mapping relationship of the IV curve into spatial features that can be captured by the convolution kernel.

[0114] Next, a one-dimensional convolutional layer is used to extract local spatial features:

[0115] F conv =ReLU(W conv *X+b conv (7)

[0116] In the formula, F conv The convolution output features are represented by ReLU, and W is the activation function. conv b represents the kernel weights. conv This is the bias vector for the convolutional layer. Local shape features of the IV curve, such as the slope of the stepped segments and the location of inflection points, are extracted using a sliding window.

[0117] Then, the data enters the pooling layer, preserving its salient features:

[0118] F pool =MaxPool(F conv (8)

[0119] In the formula, F pool Maxpool is the maximum pooling function used to pool the output features. It selects the maximum value within a local window, retaining the most significant features, such as the maximum slope change at the inflection point of a staircase. Pooling reduces feature dimensionality, decreases computation, and enhances the translation invariance of features, avoiding feature recognition errors caused by small data shifts.

[0120] The output of the tiled pooling layer is then fed into an LSTM network to capture temporal features:

[0121]

[0122] In the formula, Let h be the eigenvector at position t, where t = 1 to 16. tThe hidden state at time t contains 32 cells, and H is the LSTM output. The LSTM is used to capture the dynamic law of current changing with voltage, such as the continuity of the step segment, the change in the rate of current decrease, and the non-stationary timing characteristics of the IV curve caused by faults, such as the step-like abrupt change of the shadow fault.

[0123] Finally, deep features are generated through fully connected layers:

[0124] F deep =W s tanh(W f flatten(H)+b f (10)

[0125] In the formula, W f Let b be the weight coefficient matrix of the fully connected layer. f W is the bias vector. s The weight vector is used. The temporal features H output by the LSTM network are mapped to a one-dimensional deep feature scalar through a fully connected layer. This scalar integrates the local features extracted by the CNN with the temporal features captured by the LSTM, forming a comprehensive fault feature representation. This is achieved through the weight matrix W. f and W s Learn the nonlinear relationships between features to improve fault differentiation capabilities.

[0126] The depth feature F generated in this step deep In step 4, the features are fused with the concavity and convexity characteristics to generate a comprehensive fault index. The local shape features extracted by CNN (such as the slope of a step) are mutually verified with the inflection point localization results of concavity and convexity detection; the temporal features captured by LSTM provide a dynamic basis for fault type discrimination (such as the current change rate of hot spot faults).

[0127] Step 4: Fuse the depth features extracted by CNN-LSTM with the concavity and convexity geometric features of the IV curve to generate a comprehensive fault index F. fusion This enables accurate identification of current mismatch faults.

[0128] Due to the deep learning feature F deep While capable of capturing complex fault modes, it may be susceptible to noise interference; the concavity and convexity features depend on the geometry of the IV curve, are sensitive to inflection point location but lack global feature correlation. This invention integrates the two to form a complementary system, improving the robustness and accuracy of fault detection.

[0129] Step 4 specifically includes:

[0130] The current-voltage relationship on the IV curve can be expressed as a function:

[0131] I k =f(U k (11)

[0132] Among them, I k For current, U k Where is the voltage, k is an integer between 0 and 31, and f represents the nonlinear mapping relationship. The normal IV curve is a monotonically decreasing function, while a step-like undulating change occurs during a fault.

[0133] For the (k+1)th data point (U) on the IV curve k+1 ,I k+1 ), and its two adjacent data points (U k ,I k ) and (U k+2 ,I k+2 Define the proportionality coefficient λ that reflects the voltage interval between adjacent voltages. k+1 As shown in equation (12):

[0134]

[0135] Therefore, the following expression can be obtained through transformation:

[0136] U k+1 =λ k+1 U k +(1-λ k+1 )U k+2 (13)

[0137] I k+1 =f(U k+1 )=f(λ k+1 U k +(1-λ k+1 )U k+2 (14)

[0138] Ideally, the current should satisfy the following linear interpolation relationship:

[0139] I k+1,idea =λ k+1 I k +(1-λ k+1 )I k+2 (15)

[0140] In the formula, I k,idea I represents the linear interpolation. k Ideal value.

[0141] The deviation between the actual value and the ideal value is defined as the concavity / convexity feature, and the concavity / convexity detection feature value ΔI is then defined as follows. k+1 for:

[0142] ΔI k+1 =I k+1 -[λ k+1 I k +(1-λ k+1)I k+2 (16)

[0143] That is:

[0144] ΔI k+1 =f(U k+1 )-[λ k+1 f(U k )+(1-λ k+1 )f(U k+2 (17)

[0145] The deviation ΔI between the actual current value and the linear interpolation estimate is used to... k+1 To detect abrupt changes in the concavity / convexity of the IV curve. ΔI k+1 >0 indicates that the curve is convex at that point, corresponding to the rising edge of the step; ΔI k+1 <0 indicates that the curve dips at that point, corresponding to a stepped descending edge. Under normal, fault-free conditions, the IV curve is a smooth, monotonically decreasing function, ΔI k+1 Approaching 0, with no significant concavity / convexity mutations.

[0146] The concave-convexity detection feature value ΔI k+1 With depth features F deep Combining these factors, the comprehensive failure index F is finally obtained. fusion :

[0147] F fusion =αΔI k+1 +βF deep (18)

[0148] In the formula, α and β are the weights of the two features respectively, and α + β = 1. The optimal ratio is determined through experiments, such as α = 0.6 and β = 0.4; or through an adaptive weighting mechanism, such as dynamically adjusting α and β according to the data signal-to-noise ratio, or automatically optimizing the weight ratio through reinforcement learning.

[0149] By capturing global fault patterns through deep learning features, such as the number of steps and the current decline trend, and locating local fault inflection points through concavity and convexity features, a dual verification of "global pattern + local location" is achieved. When there is slight noise in the V curve, ΔI may show false peaks, but F... deep Because deep learning is robust enough to suppress the effects of noise; when fault characteristics are not obvious (such as early cell breakage), F deep Potential faults can be identified through global feature correlation, thus compensating for the insufficient sensitivity of ΔI.

[0150] Relying on F fusion The specific value can be used to determine whether a photovoltaic module has experienced a current mismatch fault. After comparing it with actual data, it can be confirmed that when F... fusion When the value is greater than 0.41, the photovoltaic module experiences a current mismatch fault.

[0151] Step 5: If a current mismatch fault occurs, locate the inflection point of the curve and extract the fault feature ΔK from the IV detection data to provide key parameters for accurate identification of subsequent fault types.

[0152] All three types of current mismatch faults (local shading, hot spots, and cell breakage) will cause a stepped characteristic in the IV curve, but the geometric shape of the steps (such as slope changes and inflection point distribution) differs. This step quantifies the slope characteristics near the inflection point, transforming the curve differences that are difficult to distinguish with the naked eye into calculable numerical indicators, thus achieving fine-grained differentiation of fault types.

[0153] The concavity / convexity detection feature value ΔI calculated in step 4 k+1 The positive peak value corresponds to the upward convex inflection point of the IV curve, that is, the boundary point of the step-up edge (U). k+1 ,I k+1 Traversing ΔI k+1 The sequence is used to extract local maxima, and the corresponding voltage and current coordinates are the inflection points.

[0154] Data before the inflection point may belong to the previous normal segment or another step, which is prone to interference; if the inflection point is (U k+1 ,I k+1 If ), then select the three data points P1(U) that follow it. k ,I k P2(U) k-1 ,I k-1 ), P3(U k-2 ,I k-2 These three points are located on the step section after the inflection point, and the changes in their slopes can reflect the steepness and morphological characteristics of the steps.

[0155] Calculate the absolute values ​​of the slopes at points P1 and P2, and denote them as K1 and K2, respectively:

[0156]

[0157] Since the current decreases as the voltage increases, the slope of the IV curve is theoretically negative. Taking the absolute value converts the slope into a positive scalar, which facilitates subsequent ratio calculations and threshold comparisons. Two-point slopes are sensitive to local abrupt changes and are suitable for capturing steep changes at step boundaries. If a multi-point moving average is used, it will smooth out the step features, leading to a decrease in the discriminative power of ΔK. Therefore, using two adjacent points to calculate the slope instead of more points balances feature resolution and noise sensitivity.

[0158] The slope ratio is used as a fault characteristic ΔK, which is a sensitivity indicator of faults.

[0159]

[0160] ΔK reflects the changing trend of the slope of the stepped segment, and different fault types correspond to different ΔK distribution characteristics.

[0161] Step 6: Compare the detected fault feature ΔK with the fault judgment threshold to determine the actual fault type.

[0162] All three types of faults exhibit a stepped characteristic on the IV curve, but the geometric shape of the steps (slope changes, inflection point distribution) differs fundamentally. This step quantifies the relative relationship of slope characteristics, transforming the visually imperceptible curve differences into computable decision rules, thus solving the classification problem of "same shape, different faults" in traditional methods.

[0163] This method identifies three types of photovoltaic module faults: localized shading, hot spots, and cell breakage.

[0164] 1) If K1 and K2 approach 0, it indicates that the photovoltaic module has a local shading fault.

[0165] Local shading causes multiple small steps in the IV curve, with the slope of each step approaching 0, meaning the current changes slowly with voltage. Typical scenarios include when some cells are partially obscured by leaves or bird droppings, creating multiple small shaded areas. This is because the cells in the shaded areas have increased equivalent series resistance, causing the current to decrease in segments. However, the slope of each step is small, so K1 and K2 approach 0.

[0166] 2) If ΔK≈1, it indicates that the photovoltaic module has experienced a hot spot failure.

[0167] The hotspot effect causes a localized increase in cell temperature, resulting in a single, steep step on the IV curve, which approximates a sloping straight line. The slopes K1 and K2 are almost equal, and their ratio ΔK approaches 1. This is because the cells in the hotspot region are reverse-biased, forming a localized high-resistance path. The current changes abruptly, but the slope changes uniformly, hence K1≈K2.

[0168] 3) If ΔK>1, it indicates that the photovoltaic module has experienced a cell breakage fault.

[0169] Cell breakage leads to PN junction failure, resulting in a distinct convex function shape in the IV curve, with an increasing slope in the stepped section, where K1 > K2. Typical scenarios include cell cracking due to transportation vibration or thermal stress, leading to a nonlinear current decrease. This is because the equivalent parallel resistance of the broken cell decreases, enhancing the nonlinearity of current-voltage changes, causing the slope K2 in the later section to be less than K1 in the earlier section, resulting in ΔK > 1.

[0170] For specific fault threshold settings, please refer to Figure 1As shown, the threshold values ​​were determined after analyzing and processing a large amount of actual fault data. In this embodiment, if K1 does not exceed 0.025, the photovoltaic module is considered to have a local shading fault; if ΔK does not exceed 1.13, the photovoltaic module is considered to have a local shading fault; and if ΔK is greater than 1.13, the photovoltaic module is considered to have a cell breakage fault.

[0171] To verify the effectiveness of this method, current mismatch fault detection was performed on 300 sets of measured IV data, including 100 sets of fault-free data, 100 sets of local shading fault data, 60 sets of hot spot effect fault data, and 40 sets of cell breakage fault data. The detection instrument used was a photovoltaic IV detector, and the device is as follows: Figure 4 As shown, the corresponding fault diagnosis results are as follows: Figure 5 As shown, the results indicate that the fault diagnosis accuracy of the method proposed in this invention is over 98%. Only two groups of local shadow faults failed to be diagnosed successfully, and there were no diagnostic errors for the other fault types, which verifies the feasibility and effectiveness of the method.

[0172] This method can be used for single-step detection as well as for cases where faults exist in multiple substrings. For a single faulty substring, the fault can be identified by the ΔK of a single inflection point; if multiple substrings are faulty, the ΔK of all inflection points is extracted and statistics (such as mean and mode) are calculated, and the fault type is determined by majority rule.

[0173] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for detecting current mismatch faults in photovoltaic modules based on wavelet analysis and deep learning, characterized in that, Includes the following steps: Step 1: Use a photovoltaic IV tester to collect IV data of the photovoltaic module to be diagnosed, and preprocess the collected data to remove outliers. Step 2: Perform wavelet analysis on the preprocessed data to separate the effective signal from high-frequency noise; then reconstruct the effective frequency band signal to obtain a smoothed current signal; Step 3: Construct a CNN-LSTM hybrid neural network to automatically capture the local shape features and temporal dependencies of the IV curve, and finally output the deep feature F. deep ; Step 4: Fuse the depth features F deep Concavity / convexity detection feature value ΔI of IV curve k+1 Generate a comprehensive fault index F fusion ; Step 5: If a current mismatch fault occurs, locate the inflection point of the curve and extract the fault feature ΔK from the IV detection data; Step 6: Compare the detected fault feature ΔK with the fault judgment threshold to distinguish between three types of faults: local shadow, hot spot effect, and cell breakage.

2. The photovoltaic module current mismatch fault detection method based on wavelet analysis and deep learning according to claim 1, characterized in that: In step 1, the photovoltaic IV testing instrument samples and obtains IV data from the series-connected photovoltaic modules, including several sets of data points with different voltages and currents, and arranges the data points in descending order of voltage; the sorted data points are then judged, and when an I... n+1 ≤I n At that time, I n+1 If a value is identified as an outlier, the data point with that outlier value is removed from the sorted data set.

3. The photovoltaic module current mismatch fault detection method based on wavelet analysis and deep learning according to claim 1, characterized in that: Step 2 specifically includes: performing wavelet decomposition on the data after outlier removal using wavelet basis functions; setting the scale parameter j=4 for wavelet decomposition based on the 32 data samplings per second of the photovoltaic IV detector; and decomposing the signal into low-frequency approximate components and high-frequency detail components through j-level decomposition; then reconstructing the effective frequency band signal; the frequency band selection strategy for reconstruction is: discarding the high-frequency noise band and reconstructing only the mid-to-low-frequency effective frequency band, retaining the step characteristics and inflection point information of the IV curve, thereby obtaining the denoised smooth current signal I. smooth .

4. The photovoltaic module current mismatch fault detection method based on wavelet analysis and deep learning according to claim 1, characterized in that: Step 3 specifically includes: Smooth the current signal I smooth The data points consisting of the corresponding voltage signals are converted into a two-dimensional input matrix X: Extracting local spatial features using one-dimensional convolutional layers: F conv =ReLU(W conv *X+b conv ) In the formula, F conv The convolution output features are represented by ReLU, and W is the activation function. conv b represents the kernel weights. conv The bias vector of the convolutional layer; Preserve salient features of the data through max pooling: F pool =MaxPool(F conv ) In the formula, F pool Maxpool is the maximum pooling function for pooling output features. The output of the pooling layer is flattened and then input into the LSTM network: H=[h1,h2,...,h 16 ] In the formula, Let h be the eigenvector at position t, where t = 1 to 16. t H represents the hidden state at time t, containing 32 units, and H is the LSTM output. Deep features F are generated through a fully connected layer. deep : F deep =W s tanh(W f flatten(H)+b f ) In the formula, W f Let b be the weight coefficient matrix of the fully connected layer. f W is the bias vector. s This is the weight vector.

5. The photovoltaic module current mismatch fault detection method based on IV detection data according to claim 1, characterized in that: Step 4 specifically includes: For the (k+1)th data point (U) on the IV curve k+1 ,I k+1 ), and two adjacent data points (U k ,I k ) and (U k+2 ,I k+2 Define variable λ k+1 : Define the feature value for concavity / convexity detection as ΔI k+1 : ΔI k+1 =I k+1 -[λ k+1 I k +(1-λ k+1 )I k+2 ] The concave-convexity detection feature value ΔI k+1 With depth features F deep Combining these factors, the comprehensive failure index F is finally obtained. fusion : F fusion =αΔI k+1 +βF deep In the formula, α and β are the weights corresponding to the two eigenvalues ​​respectively; Based on the calculated F fusion The specific numerical value and the threshold relationship, when F fusion >When this threshold is reached, the photovoltaic module is considered to have a current mismatch fault.

6. The photovoltaic module current mismatch fault detection method based on IV detection data according to claim 1, characterized in that: Step 5 specifically includes: The concavity / convexity detection feature value ΔI calculated in step 4 k+1 The positive peak value corresponds to the upward convex inflection point of the IV curve, that is, the boundary point of the step-up edge (U). k+1 ,I k+1 ); Traverse ΔI k+1 The sequence is used to extract local maxima, and the corresponding voltage and current coordinates are the inflection points. If the inflection point is (U) k+1 ,I k+1 If the inflection point is selected, then the three data points P1(U) after the inflection point are chosen. k ,I k P2(U) k-1 ,I k-1 ), P3(U k-2 ,I k-2 ), calculate the absolute values ​​of the slopes at points P1 and P2, and denote them as K1 and K2 respectively: The slope ratio is used as a fault characteristic ΔK, which is a sensitivity indicator of faults.

7. The photovoltaic module current mismatch fault detection method based on IV detection data according to claim 1, characterized in that: Step 6 specifically includes: if K1 and K2 approach 0, it indicates that the photovoltaic module has a local shading fault; if ΔK approaches 1, it indicates that the photovoltaic module has a hot spot fault; if ΔK>1, it indicates that the photovoltaic module has a cell breakage fault.

Images