Fault diagnosis method and system for photovoltaic power generation equipment

Through the combination of α-CEEMD signal decomposition and MF-CNN convolutional neural network, the problems of low efficiency and insufficient accuracy of photovoltaic power generation equipment are solved, and more efficient and accurate fault identification and predictive maintenance are achieved.

CN118940119APending Publication Date: 2024-11-12韩小雅
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202411065812.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-05
Publication Date
2024-11-12

AI Technical Summary

Technical Problem

The existing fault diagnosis methods for photovoltaic power generation equipment are inefficient, the traditional methods are time-consuming and inflexible, and the rules or threshold-based methods are easily affected by environmental changes and system status, resulting in false alarms or missed alarms, and the complex fault mode cannot be accurately captured.

Method used

The multi-dimensional historical original data signal is decomposed by α-CEEMD signal decomposition method, and the primary data signal feature matrix is ​​generated, and feature extraction and fault diagnosis are performed through the MF-CNN convolutional neural network, combining the multi-scale feature extraction module and feature fusion module. Enhance the accuracy and efficiency of diagnosis.

Benefits of technology

It improves the accuracy and efficiency of fault diagnosis, can more effectively capture signal changes at different scales, reduce noise interference, and realize real-time fault identification and predictive maintenance.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118940119A_ABST
    Figure CN118940119A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of photovoltaic power generation, in particular to a photovoltaic power generation equipment fault diagnosis method and system. The method comprises the following steps: acquiring multi-dimensional historical original data signals of photovoltaic power generation equipment; a complex multi-dimensional historical original data signal is decomposed into an intrinsic mode function (IMF) by using an alpha-CEEMD signal decomposition method, and a primary data signal feature matrix is extracted; performing further feature extraction on the primary data signal feature matrix by using an MF-CNN convolutional neural network to obtain feature data; and performing fault diagnosis on the photovoltaic power generation equipment according to the feature data. According to the invention, the efficiency and accuracy of fault diagnosis of the photovoltaic power generation equipment can be improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic power generation, and in particular to a photovoltaic power generation equipment fault diagnosis method and system. Background Art

[0002] With the rapid growth of photovoltaic power generation installed capacity in countries around the world, the fault diagnosis of photovoltaic power generation systems has attracted the attention of more and more scholars and related institutions at home and abroad in recent years. Since photovoltaic module arrays are installed and work in complex outdoor environments, and are affected by various environmental factors such as humidity and ultraviolet rays, they are prone to various faults such as local short circuits, performance degradation, cracks or open circuits. The occurrence of faults will reduce the power generation efficiency of photovoltaic power stations, and even cause fires in serious cases, endangering social property safety. In order to extend the life of photovoltaic power generation systems as much as possible, make photovoltaic power stations generate electricity as efficiently as possible, and improve the safety of power station operations, timely, effective and efficient maintenance is very critical.

[0003] In the prior art, traditional fault detection methods usually rely on basic measuring instruments and data recording systems. This diagnostic method is time-consuming and labor-intensive, especially in large-scale photovoltaic fields, which results in low efficiency in fault diagnosis of photovoltaic power generation equipment. In addition, some systems use rule-based fault detection methods, but these methods require detailed rule sets to be defined in advance, and lack flexibility and adaptability to new fault modes. There are also many existing technologies that use threshold-based methods to detect anomalies, but these methods are easily affected by changes in environmental conditions and system operating conditions, resulting in false alarms or missed alarms. New intelligent algorithms applied to the field of fault diagnosis of photovoltaic power generation equipment are currently one of the technologies that many experts and scholars are focusing on, including signal processing technology or artificial intelligence algorithms and the fusion of the two. For example, the document (“A New Convolutional Neural Network-Based Data-Driven Fault Diagnosis Method”, IEEE Transactions on Industrial Electronics, Wen Long et al., 2018-07-01) first converts the original data from a 1D signal to a 2D image signal, directly preprocesses the obtained image signal, and then uses the LetNet-5 framework of CNN for deep feature extraction. It consists of a total of seven layers of networks, and the accuracy is the highest among all CNN frameworks. First, the acquired two-dimensional image is used as the input layer of CNN. After three layers of convolutional layers and pooling layers, the output feature map size is 1x1. Finally, Softmax classification is used as the output layer to directly classify the fault. In the prior art, a Chinese patent with a publication number of CN115587642B discloses a photovoltaic system fault alarm method based on a BP neural network. The patent establishes a distributed photovoltaic system BP neural network, screens out training data, and then alarms for abnormal output characteristics during the operation of distributed photovoltaic systems in the region. However, the feature extraction technology used in this prior art may not be able to capture subtle changes in complex failure modes, which may lead to inaccurate diagnosis results.

[0004] In order to solve the above problems, the present invention proposes a photovoltaic power generation equipment fault diagnosis method and system. Summary of the invention

[0005] The purpose of the present invention is to provide a photovoltaic power generation equipment fault diagnosis method and system, firstly obtaining a multi-dimensional historical original data signal of a photovoltaic power generation component during a monitoring period; then using a signal decomposition method to decompose the multi-dimensional historical original data signal into a primary data signal feature matrix; then, using the primary data signal feature matrix as the input of a convolutional neural network; finally, performing fault diagnosis on the photovoltaic power generation equipment according to the primary data signal feature matrix.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A photovoltaic power generation equipment fault diagnosis method, specifically:

[0008] Acquire a multi-dimensional historical raw data signal output by a solar photovoltaic cell assembly during monitoring time, including temperature, voltage, and ambient irradiance, and add a time attribute to the multi-dimensional historical raw data signal of the solar photovoltaic cell assembly;

[0009] Decomposing the multidimensional historical original data signal by using an α-CEEMD signal decomposition method, adding paired weighted positive noise and negative noise to each of the multidimensional historical original data signals, performing EEMD decomposition on each of the multidimensional historical original data signals after adding the positive noise and the negative noise to obtain an intrinsic mode function IMF, and after obtaining all the intrinsic mode functions IMF, averaging all the intrinsic mode functions IMF to obtain a final IMF; the IMF is used as a primary data signal feature matrix;

[0010] The primary data signal feature matrix is ​​used as input and sent to the MF-CNN convolutional neural network for further feature extraction; the MF-CNN includes an input layer, a hidden layer, a multi-scale feature extraction module, a feature fusion module, a fully connected layer and an output layer; wherein the multi-scale feature extraction module includes a parallel convolution layer combination and a pooling layer; the feature fusion module includes multi-resolution feature fusion and an attention mechanism;

[0011] The trained MF-CNN is used to perform fault diagnosis on the photovoltaic power generation equipment.

[0012] The multi-dimensional historical raw data signal includes: temperature, including the internal temperature of the junction box, the temperature of the solar photovoltaic cell assembly and the temperature of the photovoltaic inverter; the temperature includes the temperatures of the junction box, the solar photovoltaic cell assembly and the photovoltaic inverter respectively when a fault occurs; voltage, including the three-phase voltage of the inverter; the voltage includes the voltage of the inverter when a fault occurs; environmental irradiance, the environmental irradiance refers to the radiation flux per unit area of ​​the illuminated surface of the solar photovoltaic cell assembly.

[0013] The time attribute refers to establishing a time axis for the multi-dimensional historical original data signal, and each time point on the time axis has a corresponding multi-dimensional historical original data signal.

[0014] The signal decomposition method is an α-CEEMD signal decomposition method. In the α-CEEMD signal decomposition method, the stability and accuracy of the decomposition are further improved by using positive and negative paired weighted noises. In the α-CEEMD, each multi-dimensional historical original data signal after adding positive noise and negative noise is subjected to EEMD decomposition to obtain an intrinsic mode function IMF. After all the intrinsic mode functions IMF are obtained, all the intrinsic mode functions IMF are averaged to obtain the final IMF. The IMF is used as a primary data signal feature matrix. The process of the α-CEEMD signal decomposition method is:

[0015] Set the overall average number to I;

[0016] Add noise: construct a complementary set to generate multiple sets of random weighted noise αε(t), and add the noise αε(t) to the original signal x(t) to generate multiple sets of random test signals x i (t); wherein the random noise αε(t) is divided into positive noise and negative noise Among them, the positive noise and the negative noise is a complementary relationship; the test signal x i (t) and the noise formula ε(t) are as follows:

[0017] x i (t) = x(t) + αε i (t)

[0018]

[0019] Wherein, i represents the i-th addition of the noise signal; α is the weight of the α-CEEMD signal decomposition method for noise adaptive adjustment, α is related to the input multidimensional historical original data signal, and the size of α is determined by calculating the oscillation amplitude of the multidimensional historical original data signal. The specific calculation formula is as follows:

[0020]

[0021] Among them, f max is the maximum oscillation amplitude of the historical raw data signal of the dimension, f min is the minimum oscillation amplitude of the historical raw data signal of the dimension, that is, f max With f min The greater the absolute value of the difference, the smaller α is, and f max With f min The smaller the absolute value of the difference, the larger α is;

[0022] Signal decomposition: for the multiple groups of test signals x i(t) Perform EMD decomposition respectively to obtain their respective intrinsic mode functions IMF:

[0023]

[0024] Among them, c i,j (t) is the jth IMF obtained by decomposing after adding white noise for the i-th time; r i,j (t) is the residual function, the average trend of the signal; J is the number of IMFs;

[0025] Repeat the two steps of adding noise and decomposing the signal once, adding one of the complementary positive noise and the negative noise each time, and obtain the set of IMFs as follows:

[0026] c 1,j (t)c 2,j (t)...c I,j (t),j=1,2,...J

[0027] Using the principle that the statistical mean of unrelated sequences is zero, the IMF is subjected to ensemble average calculation to obtain the final IMF after EEMD decomposition:

[0028]

[0029] Among them, c j (t) is the jth IMF of EEMD decomposition; c j (t) is the primary data signal characteristic.

[0030] The input layer takes the primary data signal feature matrix as input, uses a convolution kernel size of 3x3 and a step size of 2;

[0031] The hidden layer includes a convolution layer and a pooling layer. The convolution layer uses a convolution kernel size of 3x3 and a step size of 2. The pooling layer uses a pooling size of 2x2 and a step size of 2, and the pooling type is maximum pooling.

[0032] The multi-scale feature extraction module includes a parallel convolutional layer combination and a pooling layer. The parallel convolutional layer combination adopts multiple parallel convolutional layers, each of the convolutional layers has a convolution kernel of a different size, the pooling size used by the pooling layer is 2x2, the step size is 2, and the pooling type is maximum pooling.

[0033] The feature fusion module includes multi-resolution feature fusion and an attention mechanism. The multi-resolution feature fusion fuses feature maps from the input layer, hidden layer and multi-scale feature extraction module. The attention mechanism enhances the focus on and utilization of key features.

[0034] The fully connected layer has 256 neural units.

[0035] The output layer classifies the faults and outputs the final result.

[0036] A photovoltaic power generation equipment fault diagnosis system, the photovoltaic power generation equipment fault diagnosis system includes a data signal acquisition module, which acquires a multidimensional historical original data signal output by a solar photovoltaic cell assembly during a monitoring period, including temperature, voltage and environmental irradiance, and adds a time attribute to the multidimensional historical original data signal of the solar photovoltaic cell assembly; a data signal decomposition module, which decomposes the multidimensional historical original data signal using an α-CEEMD signal decomposition method, adds paired weighted positive noise and negative noise to each of the multidimensional historical original data signals, performs EEMD decomposition on each of the multidimensional historical original data signals after adding positive noise and negative noise to obtain an intrinsic mode function IMF, and obtains the After extracting some intrinsic mode functions IMF, all the intrinsic mode functions IMF are averaged to obtain the final IMF; the IMF is used as the primary data signal feature matrix; the fault diagnosis module takes the primary data signal feature matrix as input and sends it to the MF-CNN convolutional neural network for further feature extraction; the MF-CNN includes an input layer, a hidden layer, a multi-scale feature extraction module, a feature fusion module, a fully connected layer and an output layer; wherein the multi-scale feature extraction module includes a combination of parallel convolutional layers and a pooling layer; the feature fusion module includes multi-resolution feature fusion and an attention mechanism; and then the trained MF-CNN is used to perform fault diagnosis on the photovoltaic power generation equipment.

[0037] The multi-dimensional historical raw data signal includes: temperature, including the internal temperature of the junction box, the temperature of the solar photovoltaic cell assembly and the temperature of the photovoltaic inverter; the temperature includes the temperatures of the junction box, the solar photovoltaic cell assembly and the photovoltaic inverter respectively when a fault occurs; voltage, including the three-phase voltage of the inverter; the voltage includes the voltage of the inverter when a fault occurs; environmental irradiance, the environmental irradiance refers to the radiation flux per unit area of ​​the illuminated surface of the solar photovoltaic cell assembly.

[0038] The additional time attribute refers to establishing a time axis for the multi-dimensional historical original data signal, and each time point on the time axis has a corresponding multi-dimensional historical original data signal.

[0039] The process of the α-CEEMD signal decomposition method is:

[0040] Set the overall average number to I;

[0041] Add noise: construct a complementary set to generate multiple sets of random weighted noise αε(t), and add the noise αε(t) to the original signal x(t) to generate multiple sets of random test signals x i (t); wherein the random noise αε(t) is divided into positive noise and negative noise Among them, the positive noise and the negative noise is a complementary relationship; the test signal x i (t) and the noise formula ε(t) are as follows:

[0042] x i (t) = x(t) + αε i (t)

[0043]

[0044] Wherein, i represents the i-th addition of the noise signal; α is the weight of the α-CEEMD signal decomposition method for noise adaptive adjustment, α is related to the input multidimensional historical original data signal, and the size of α is determined by calculating the oscillation amplitude of the multidimensional historical original data signal. The specific calculation formula is as follows:

[0045]

[0046] Among them, f max is the maximum oscillation amplitude of the historical raw data signal of the dimension, f min is the minimum oscillation amplitude of the historical raw data signal of the dimension, that is, f max With f min The greater the absolute value of the difference, the smaller α is, and f max With f min The smaller the absolute value of the difference, the larger α is;

[0047] Signal decomposition: for the multiple groups of test signals x i (t) Perform EMD decomposition respectively to obtain their respective intrinsic mode functions IMF:

[0048]

[0049] Among them, c i,j (t) is the jth IMF obtained by decomposing after adding white noise for the i-th time; r i,j (t) is the residual function, the average trend of the signal; J is the number of IMFs;

[0050] Repeat the two steps of adding noise and decomposing the signal once, adding one of the complementary positive noise and the negative noise each time, and obtain the set of IMFs as follows:

[0051] c 1,j (t)c 2,j (t)...c I,j (t),j=1,2,...J

[0052] Using the principle that the statistical mean of unrelated sequences is zero, the IMF is subjected to ensemble average calculation to obtain the final IMF after EEMD decomposition:

[0053]

[0054] Among them, c j (t) is the jth IMF of EEMD decomposition; c j (t) is the primary data signal characteristic.

[0055] The input layer takes the primary data signal feature matrix as input, uses a convolution kernel size of 3x3 and a step size of 2;

[0056] The hidden layer includes a convolution layer and a pooling layer. The convolution layer uses a convolution kernel size of 3x3 and a step size of 2. The pooling layer uses a pooling size of 2x2 and a step size of 2, and the pooling type is maximum pooling.

[0057] The multi-scale feature extraction module includes a parallel convolutional layer combination and a pooling layer. The parallel convolutional layer combination adopts multiple parallel convolutional layers, each of the convolutional layers has a convolution kernel of a different size, the pooling size used by the pooling layer is 2x2, the step size is 2, and the pooling type is maximum pooling.

[0058] The feature fusion module includes multi-resolution feature fusion and an attention mechanism. The multi-resolution feature fusion fuses feature maps from the input layer, hidden layer and multi-scale feature extraction module. The attention mechanism enhances the focus on and utilization of key features.

[0059] The fully connected layer has 256 neural units.

[0060] The output layer classifies the faults and outputs the final result.

[0061] Compared with the prior art, the present invention has the following beneficial effects:

[0062] 1. The multi-dimensional historical raw data sample signal obtained by the present invention can cover more data details, including the impact of different light, temperature, voltage and other factors on the performance of solar photovoltaic cell modules, which helps to establish a more accurate prediction model, predict the problems and performance degradation trends of solar photovoltaic cell modules, improve the accuracy of analysis, and thus achieve more effective predictive maintenance.

[0063] 2. The present invention adopts an α-CEEMD signal decomposition method, which introduces an adaptive weighted noise adjustment mechanism, adds a pair of mutually opposite weighted positive and negative white noises as auxiliary noises to the multi-dimensional historical original data sample signal, and calculates differentiated noise for different data with different oscillation amplitudes. This method can eliminate the excess auxiliary white noise remaining in the reconstructed signal after decomposition, and reduce the number of iterations required for decomposition; decomposes multiple groups of data of solar photovoltaic power generation components, and extracts reliable and useful information and trends from the multi-dimensional historical original data sample signals. The α-CEEMD signal decomposition method helps to reduce noise interference in the data and improve the quality of subsequent data feature extraction. At the same time, combining the time-frequency features extracted by α-CEEMD with the spatial feature representation of the MF-CNN convolutional neural network can improve the accuracy of fault diagnosis and prediction.

[0064] 3. The present invention adopts MF-CNN convolutional neural network. The multi-scale feature extraction module in the MF-CNN can effectively capture the features of different scales in the signal, including subtle and overall signal changes, through parallel convolution layers and convolution kernels of different scales. In addition, MF-CNN can better integrate information from different scales through the feature fusion module, and enhance the expression and recognition capabilities of key features. The trained complete MF-CNN model has the ability of real-time fault diagnosis. Its parallel computing and effective feature learning capabilities significantly accelerate the fault diagnosis rate, and can respond and identify complex fault modes more quickly. It can monitor and identify fault signs in real time during the operation of photovoltaic power generation components, thereby improving the efficiency of fault diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 It is a flow chart of the photovoltaic power generation equipment fault diagnosis method of the present invention;

[0066] Figure 2 It is a structural diagram of the photovoltaic power generation equipment fault diagnosis system of the present invention;

[0067] Figure 3 It is a network framework diagram of the MF-CNN convolutional neural network in the photovoltaic power generation equipment fault diagnosis method of the present invention. DETAILED DESCRIPTION

[0068] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0069] See also Figure 1 to Figure 2The present invention provides a photovoltaic power generation equipment fault diagnosis method and system. In order to illustrate the function of the present invention, the effectiveness of the present invention will be described from two embodiments below.

[0070] Embodiment 1

[0071] As an embodiment of the present invention, refer to Figure 1 , showing a flow chart of the photovoltaic power generation equipment fault diagnosis method of the present invention, the specific steps include: S101. Acquire the multi-dimensional historical original data signal of the solar photovoltaic cell assembly during the monitoring time; S102. Use the α-CEEMD signal decomposition method to decompose the multi-dimensional historical original data signal to generate a primary data signal feature matrix; S103. Send it to the MF-CNN convolutional neural network for further feature extraction, and use the trained MF-CNN to diagnose the fault of the photovoltaic power generation equipment.

[0072] Further, Figure 1 The steps are Figure 2 The system unit is implemented, and the specific process is as follows; according to Figure 2 The data signal acquisition module in the system acquires various operating data of photovoltaic power generation equipment; based on the acquired data, the α-CEEMD data signal decomposition module is used to decompose it into a primary data signal feature matrix; the primary signal feature matrix is ​​used as the input of the MF-CNN convolutional neural network for training to predict the operating failures of photovoltaic power generation equipment.

[0073] Among them, in the above step S101, the operation data of the solar photovoltaic cell assembly obtained include: temperature, including the temperature inside the junction box, the temperature of the solar photovoltaic cell assembly and the temperature of the photovoltaic inverter; the temperature includes the temperature of the solar photovoltaic cell assembly, the junction box and the photovoltaic inverter respectively when a fault occurs; voltage, including the three-phase voltage of the inverter; the voltage includes the voltage of the inverter when a fault occurs; environmental irradiance, the environmental irradiance refers to the radiation flux per unit area of ​​the solar photovoltaic cell assembly. The operation data of the solar photovoltaic cell assembly obtained are diverse, providing data support for subsequent fault diagnosis, ensuring the accuracy of diagnosis and the diversity of faults.

[0074] Furthermore, a time attribute is added to the multidimensional historical raw data signal of the solar photovoltaic cell assembly, and the additional time attribute means establishing a time axis for the multidimensional historical raw data signal, and each time point on the time axis has its own corresponding multidimensional historical raw data signal. The time axis is established to convert the multidimensional historical raw data signal into a time domain signal, which is convenient for subsequent signal decomposition and at the same time, the data of different dimensions correspond one to one.

[0075] Furthermore, the multi-dimensional historical original data signal is decomposed using an α-CEEMD signal decomposition method to generate a primary data signal feature matrix, corresponding to the above step S102.

[0076] In the α-CEEMD signal decomposition method, the stability and accuracy of the decomposition are further improved by using positive and negative paired weighted noise; in the α-CEEMD, each multidimensional historical original data signal after adding positive noise and negative noise is subjected to EEMD decomposition to obtain an intrinsic mode function IMF, and after all the intrinsic mode functions IMF are obtained, all the intrinsic mode functions IMF are averaged to obtain the final IMF; the IMF is used as the primary data signal feature matrix; the process of the α-CEEMD signal decomposition method is:

[0077] Set the overall average number to I;

[0078] Add noise: construct a complementary set to generate multiple sets of random weighted noise αε(t), and add the noise αε(t) to the original signal x(t) to generate multiple sets of random test signals x i (t); wherein the random noise αε(t) is divided into positive noise and negative noise Among them, the positive noise and the negative noise is a complementary relationship; the test signal x i (t) and the noise formula ε(t) are as follows:

[0079] x i (t) = x(t) + αε i (t)

[0080]

[0081] Wherein, i represents the i-th addition of the noise signal; α is the weight of the α-CEEMD signal decomposition method for noise adaptive adjustment, α is related to the input multidimensional historical original data signal, and the size of α is determined by calculating the oscillation amplitude of the multidimensional historical original data signal. The specific calculation formula is as follows:

[0082]

[0083] Among them, f max is the maximum oscillation amplitude of the historical raw data signal of the dimension, f min is the minimum oscillation amplitude of the historical raw data signal of the dimension, that is, f max With f min The greater the absolute value of the difference, the smaller α is, and f max With f minThe smaller the absolute value of the difference, the larger α is;

[0084] Signal decomposition: for the multiple groups of test signals x i (t) Perform EMD decomposition respectively to obtain their respective intrinsic mode functions IMF:

[0085]

[0086] Among them, c i,j (t) is the jth IMF obtained by decomposing after adding white noise for the i-th time; r i,j (t) is the residual function, the average trend of the signal; J is the number of IMFs;

[0087] Repeat the two steps of adding noise and decomposing the signal once, adding one of the complementary positive noise and the negative noise each time, and obtain the set of IMFs as follows:

[0088] c 1,j (t)c 2,j (t)...c I,j (t),j=1,2,...J

[0089] Using the principle that the statistical mean of unrelated sequences is zero, the IMF is subjected to ensemble average calculation to obtain the final IMF after EEMD decomposition:

[0090]

[0091] Among them, c j (t) is the jth IMF of EEMD decomposition; c j (t) is the primary data signal feature matrix.

[0092] Furthermore, the primary data signal feature matrix is ​​used as input and sent to a convolutional neural network for further feature extraction to perform fault diagnosis on the photovoltaic power generation equipment, corresponding to the above step S103.

[0093] The data generated in photovoltaic power generation systems usually have complex nonlinear and non-stationary characteristics. The α-CEEMD signal decomposition method can effectively decompose these complex signals into intrinsic mode functions (IMFs) of different frequencies and time series of different scales. In the α-CEEMD signal decomposition method, the technology of calculating weighted noise allows differentiated noise evaluation for different data sets with different oscillation amplitudes. This method can accurately separate and quantify the noise components in the data by considering the changing characteristics of the data in the frequency domain and time domain. Specifically, it can effectively distinguish signal components that behave differently in frequency and amplitude, thereby providing more accurate noise analysis and processing for complex data sets. The fine signal after α-CEEMD decomposition makes it possible to more clearly observe the changing pattern and frequency components of the signal during a fault, which can more clearly reflect the true fault characteristics, help to more accurately locate and diagnose the fault, and improve the accuracy and reliability of fault diagnosis.

[0094] Reference Figure 3 , the convolutional neural network MF-CNN has an input layer, a hidden layer, a multi-scale feature extraction module, a feature fusion module, a fully connected layer and an output layer;

[0095] The input layer uses the primary data signal feature matrix as input, uses a convolution kernel size of 3x3 and a step size of 2; then uses the activation function ReLU to extract nonlinear features;

[0096] The hidden layer performs feature extraction and feature mapping, and has two parts: a convolution layer and a pooling layer; the convolution layer uses a convolution kernel size of 3x3 and a step size of 2; the pooling layer helps reduce the data dimension and retain the most significant features, and uses a pooling size of 2x2, a step size of 2, and a pooling type of maximum pooling to reduce the spatial size of the feature map;

[0097] The multi-scale feature extraction module includes a combination of parallel convolutional layers and pooling layers. Through the convolution kernels of different scales in the parallel convolutional layers, MF-CNN can effectively capture features of different scales in the signal, including subtle and overall signal changes;

[0098] Among them, the parallel convolution layer combination adopts multiple parallel convolution layers, each of the convolution layers has a convolution kernel of different sizes, which is used to capture features at different scales; the pooling size used by the pooling layer is 2x2, the step size is 2, and the pooling type is maximum pooling; the spatial distribution of the features is adjusted.

[0099] The feature fusion module introduces multi-resolution feature fusion and attention mechanism, which can better integrate information from different scales and enhance the expression and recognition capabilities of key features.

[0100] Among them, the multi-resolution feature fusion fuses the feature maps from the input layer, hidden layer and multi-scale feature extraction module to obtain richer and more comprehensive feature representation; the attention mechanism dynamically adjusts the importance of each scale feature to enhance the focus on and utilization of key features.

[0101] The fully connected layer has 256 neural units for learning more advanced feature representations, followed by an activation function ReLU to extract nonlinear features.

[0102] The output layer classifies the faults and outputs the final result.

[0103] In the example of the present application, the MF-CNN also includes the setting of parameters during training: a Dropout mechanism is added between layers to prevent the model from overfitting, and the parameter value of the Dropout mechanism is set to 0.5, which is a value commonly used in deep learning models; the number of training rounds is set to 500, the model is optimized using stochastic gradient descent SGD, the learning rate is set to 0.01, and a training strategy of decreasing the learning rate round by round is used;

[0104] The MF-CNN convolutional neural network can distinguish the characteristics of the data to the greatest extent through continuous iteration. When a photovoltaic power generation component and its components fail, such as a hot spot failure in a battery cell, the temperature of the corresponding component will rise. The MF-CNN convolutional neural network will capture the relationship characteristics between the above failure and temperature. In the feature matrix generated by MF-CNN, the weight between the two will be correspondingly larger. Through MF-CNN's powerful feature capture capability and the ability to analyze high-order features, the fault diagnosis of photovoltaic power generation equipment can be more accurate and achieve real-time results.

[0105] In photovoltaic power generation systems, faults usually manifest as specific patterns or anomalies in the data, and MF-CNN can learn the most effective feature representation directly from the data through layer-by-layer convolution. This not only saves time for manual feature engineering, but also captures the key features in the data more comprehensively, improving the efficiency of diagnosis. In addition, since the MF-CNN model can quickly process new input data after training, it can achieve almost real-time fault diagnosis without human intervention or complex intermediate processing steps. This is especially important for photovoltaic power generation systems that require rapid response and processing, reducing system downtime and losses, simplifying the entire diagnostic process, and improving efficiency.

[0106] Embodiment 2

[0107] The inventive method in the first embodiment is combined with the system to form a photovoltaic power generation equipment fault diagnosis system. The inventive method is described again in this embodiment.

[0108] Reference Figure 2 , showing the module diagram of the photovoltaic power generation equipment fault diagnosis system of the present invention, the specific modules include: 1. a data signal acquisition module, which acquires the multi-dimensional historical original data signal of the solar photovoltaic cell assembly during the monitoring time; 2. a data signal decomposition module, which generates a primary data signal feature matrix; 3. a fault diagnosis module, which takes the primary data signal feature matrix as input, uses the MF-CNN convolutional neural network to extract features, and uses the trained MF-CNN to perform fault diagnosis on the photovoltaic power generation equipment.

[0109] Wherein, the multi-dimensional historical raw data signal includes: temperature, including the temperature inside the junction box, the temperature of the solar photovoltaic cell assembly and the temperature of the photovoltaic inverter; the temperature includes the temperatures of the solar photovoltaic cell assembly, the junction box and the photovoltaic inverter respectively when a fault occurs; voltage, including the three-phase voltage of the inverter; the voltage includes the voltage of the inverter when a fault occurs; environmental irradiance, the environmental irradiance refers to the radiation flux per unit area of ​​the illuminated surface of the solar photovoltaic cell assembly.

[0110] Furthermore, a time attribute is added to the multidimensional historical original data signal of the solar photovoltaic cell assembly, and the added time attribute refers to establishing a time axis for the multidimensional historical original data signal, and each time point on the time axis has its own corresponding multidimensional historical original data signal.

[0111] Furthermore, the data signal decomposition module uses an α-CEEMD signal decomposition method, in which the stability and accuracy of the decomposition are further improved by using positive and negative paired weighted noise; in the α-CEEMD, each multidimensional historical original data signal after adding positive noise and negative noise is subjected to EEMD decomposition to obtain an intrinsic mode function IMF, and after obtaining all the intrinsic mode functions IMF, all the intrinsic mode functions IMF are averaged to obtain the final IMF; the IMF is used as the primary data signal feature matrix; the process of the α-CEEMD signal decomposition method is:

[0112] Set the overall average number to I;

[0113] Add noise: construct a complementary set to generate multiple sets of random weighted noise αε(t), and add the noise αε(t) to the original signal x(t) to generate multiple sets of random test signals x i (t); wherein the random noise αε(t) is divided into positive noise and negative noise Among them, the positive noise and the negative noise is a complementary relationship; the test signal xi (t) and the noise formula ε(t) are as follows:

[0114] x i (t) = x(t) + αε i (t)

[0115]

[0116] Wherein, i represents the i-th addition of the noise signal; α is the weight of the α-CEEMD signal decomposition method for noise adaptive adjustment, α is related to the input multidimensional historical original data signal, and the size of α is determined by calculating the oscillation amplitude of the multidimensional historical original data signal. The specific calculation formula is as follows:

[0117]

[0118] Among them, f max is the maximum oscillation amplitude of the historical raw data signal of the dimension, f min is the minimum oscillation amplitude of the historical raw data signal of the dimension, that is, f max With f min The greater the absolute value of the difference, the smaller α is, and f max With f min The smaller the absolute value of the difference, the larger α is;

[0119] Signal decomposition: for the multiple groups of test signals x i (t) Perform EMD decomposition respectively to obtain their respective intrinsic mode functions IMF:

[0120]

[0121] Among them, c i,j (t) is the jth IMF obtained by decomposing after adding white noise for the i-th time; r i,j (t) is the residual function, the average trend of the signal; J is the number of IMFs;

[0122] Repeat the two steps of adding noise and decomposing the signal once, adding one of the complementary positive noise and the negative noise each time, and obtain the set of IMFs as follows:

[0123] c 1,j (t)c 2,j (t)...c I,j (t),j=1,2,...J

[0124] Using the principle that the statistical mean of unrelated sequences is zero, the IMF is subjected to ensemble average calculation to obtain the final IMF after EEMD decomposition:

[0125]

[0126] Among them, c j (t) is the jth IMF of EEMD decomposition; c j (t) is the primary data signal feature matrix.

[0127] Furthermore, the primary data signal feature matrix is ​​used as input and sent to a convolutional neural network for further feature extraction to perform fault diagnosis on the photovoltaic power generation equipment.

[0128] Further, see Figure 3 , the fault diagnosis module uses a convolutional neural network MF-CNN, which has an input layer, a hidden layer, a multi-scale feature extraction module, a feature fusion module, a fully connected layer and an output layer;

[0129] The input layer uses the primary data signal feature matrix as input, uses a convolution kernel size of 3x3 and a step size of 2; then uses the activation function ReLU to extract nonlinear features;

[0130] The hidden layer performs feature extraction and feature mapping, and has two parts: a convolution layer and a pooling layer; the convolution layer uses a convolution kernel size of 3x3 and a step size of 2; the pooling layer helps reduce the data dimension and retain the most significant features, and uses a pooling size of 2x2, a step size of 2, and a pooling type of maximum pooling to reduce the spatial size of the feature map;

[0131] The multi-scale feature extraction module includes a combination of parallel convolutional layers and pooling layers. Through the convolution kernels of different scales in the parallel convolutional layers, MF-CNN can effectively capture features of different scales in the signal, including subtle and overall signal changes;

[0132] Among them, the parallel convolution layer combination adopts multiple parallel convolution layers, each of the convolution layers has a convolution kernel of different sizes, which is used to capture features at different scales; the pooling size used by the pooling layer is 2x2, the step size is 2, and the pooling type is maximum pooling; the spatial distribution of the features is adjusted.

[0133] The feature fusion module introduces multi-resolution feature fusion and attention mechanism, which can better integrate information from different scales and enhance the expression and recognition capabilities of key features.

[0134] Among them, the multi-resolution feature fusion fuses the feature maps from the input layer, hidden layer and multi-scale feature extraction module to obtain richer and more comprehensive feature representation; the attention mechanism dynamically adjusts the importance of each scale feature to enhance the focus on and utilization of key features.

[0135] The fully connected layer has 256 neural units for learning more advanced feature representations, followed by an activation function ReLU to extract nonlinear features.

[0136] The output layer classifies the faults and outputs the final result.

[0137] The MF-CNN also includes the setting of necessary parameters during training: adding a Dropout mechanism between layers to prevent overfitting of the model, and the parameter value of the Dropout mechanism is set to 0.5, a value commonly used in deep learning models; the number of training rounds is set to 500, the model is optimized using stochastic gradient descent SGD, the learning rate is set to 0.01, and a training strategy of decreasing the learning rate round by round is used.

[0138] The present invention utilizes multi-dimensional historical raw data sample signals, covers richer data details, helps to establish a more accurate prediction model, predicts possible failures of solar photovoltaic cell modules, and achieves more accurate predictive maintenance. By using the α-CEEMD signal decomposition method to process these multi-dimensional historical raw data signals, its adaptive weighted noise adjustment mechanism can perform differentiated noise calculations for data with different oscillation amplitudes to achieve efficient signal decomposition. The MF-CNN convolutional neural network is used to further extract signal features, effectively capturing signal features of different scales, including small and overall signal changes. At the same time, the time-frequency features extracted by α-CEEMD are combined with the spatial feature representation of the MF-CNN convolutional neural network to improve the accuracy of fault diagnosis and prediction.

[0139] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A photovoltaic power generation equipment fault diagnosis method, characterized in that: include: Acquire a multi-dimensional historical raw data signal output by a solar photovoltaic cell assembly during monitoring time, including temperature, voltage, and ambient irradiance, and add a time attribute to the multi-dimensional historical raw data signal of the solar photovoltaic cell assembly; Decomposing the multidimensional historical original data signal by using an α-CEEMD signal decomposition method, adding paired weighted positive noise and negative noise to each of the multidimensional historical original data signals, performing EEMD decomposition on each of the multidimensional historical original data signals after adding the positive noise and the negative noise to obtain an intrinsic mode function IMF, and after obtaining all the intrinsic mode functions IMF, averaging all the intrinsic mode functions IMF to obtain a final IMF; the IMF is used as a primary data signal feature matrix; The primary data signal feature matrix is ​​used as input and sent to the MF-CNN convolutional neural network for further feature extraction; the MF-CNN includes an input layer, a hidden layer, a multi-scale feature extraction module, a feature fusion module, a fully connected layer and an output layer; wherein the multi-scale feature extraction module includes a parallel convolution layer combination and a pooling layer; the feature fusion module includes multi-resolution feature fusion and an attention mechanism; The trained MF-CNN is used to perform fault diagnosis on the photovoltaic power generation equipment.

2. A photovoltaic power generation equipment fault diagnosis method according to claim 1, characterized in that: The multi-dimensional historical raw data signal includes: Temperature, including the temperature inside the junction box, the temperature of the solar photovoltaic cell assembly and the temperature of the photovoltaic inverter; the temperature includes the temperature of the junction box, the solar photovoltaic cell assembly and the photovoltaic inverter respectively when a fault occurs; Voltage, including the three-phase voltage of the inverter; the voltage includes the voltage of the inverter when a fault occurs; Ambient irradiance, the ambient irradiance refers to the radiation flux per unit area of ​​the illuminated surface of the solar photovoltaic cell assembly; the additional time attribute refers to establishing a time axis for the multi-dimensional historical original data signal, and each time point on the time axis has its own corresponding multi-dimensional historical original data signal.

3. A photovoltaic power generation equipment fault diagnosis method according to claim 1, characterized in that: The additional time attribute refers to establishing a time axis for the multi-dimensional historical original data signal, and each time point on the time axis has a corresponding multi-dimensional historical original data signal.

4. A photovoltaic power generation equipment fault diagnosis method according to claim 1, characterized in that: In the α-CEEMD signal decomposition method, the stability and accuracy of the decomposition are further improved by using positive and negative paired weighted noise; in the α-CEEMD, each multidimensional historical original data signal after adding positive noise and negative noise is subjected to EEMD decomposition to obtain an intrinsic mode function IMF, and after all the intrinsic mode functions IMF are obtained, all the intrinsic mode functions IMF are averaged to obtain the final IMF; the IMF is used as the primary data signal feature matrix; the process of the α-CEEMD signal decomposition method is: Set the overall average number to I; Add noise: Construct a complementary set to generate multiple sets of random weighted noise αε(t), and add the noise αε(t) to the original signal x(t) to generate multiple sets of random test signals x i (t); wherein the random noise αε(t) is divided into positive noise and negative noise Among them, the positive noise and the negative noise is a complementary relationship; the test signal x i (t) and the noise formula ε(t) are as follows: x i (t)=x(t)+αε i (t) Wherein, i represents the i-th addition of the noise signal; α is the weight of the α-CEEMD signal decomposition method for noise adaptive adjustment, α is related to the input multidimensional historical original data signal, and the size of α is determined by calculating the oscillation amplitude of the multidimensional historical original data signal. The specific calculation formula is as follows: Among them, f max is the maximum oscillation amplitude of the historical raw data signal of the dimension, f min is the minimum oscillation amplitude of the historical raw data signal of the dimension, that is, f max With f min The greater the absolute value of the difference, the smaller α is, and f max With f min The smaller the absolute value of the difference, the larger α is; Signal decomposition: for the multiple groups of test signals x i (t) Perform EMD decomposition respectively to obtain their respective intrinsic mode functions IMF: Among them, c i,j (t) is the jth IMF obtained by decomposing after adding white noise for the i-th time; r i,j (t) is the residual function, the average trend of the signal; J is the number of IMFs; Repeat the two steps of adding noise and decomposing the signal once, adding one of the complementary positive noise and the negative noise each time, and obtain the set of IMFs as follows: c 1,j (t)c 2,j (t)...c I,j (t),j=1,2,...J Using the principle that the statistical mean of unrelated sequences is zero, the IMF is subjected to ensemble average calculation to obtain the final IMF after EEMD decomposition: i=1,2,...I,j=1,2,...J Among them, c j (t) is the jth IMF of EEMD decomposition; c j (t) is the primary data signal feature matrix.

5. A photovoltaic power generation equipment fault diagnosis method according to claim 1, characterized in that: The input layer takes the primary data signal feature matrix as input, uses a convolution kernel size of 3x3 and a step size of 2; The hidden layer includes a convolution layer and a pooling layer. The convolution layer uses a convolution kernel size of 3x3 and a step size of 2. The pooling layer uses a pooling size of 2x2 and a step size of 2, and the pooling type is maximum pooling. The multi-scale feature extraction module includes a parallel convolution layer combination and a pooling layer, wherein the parallel convolution layer combination uses a plurality of parallel convolution layers, each of the convolution layers has a convolution kernel of a different size, the pooling size used by the pooling layer is 2x2, the step size is 2, and the pooling type is maximum pooling; The feature fusion module includes multi-resolution feature fusion and attention mechanism, wherein the multi-resolution feature fusion fuses feature maps from the input layer, hidden layer and multi-scale feature extraction module, and the attention mechanism enhances the focus on and utilization of key features; The fully connected layer has 256 neural units; The output layer classifies the faults and outputs the final result.

6. A photovoltaic power generation equipment fault diagnosis system, the photovoltaic power generation equipment fault diagnosis system is used to run a program, wherein: When the program is running, the photovoltaic power generation equipment fault diagnosis method according to any one of claims 1 to 5 is executed, characterized in that it includes: A data signal acquisition module, which acquires the multi-dimensional historical original data signal output by the solar photovoltaic cell assembly during the monitoring time, including temperature, voltage and environmental irradiance, and adds a time attribute to the multi-dimensional historical original data signal of the solar photovoltaic cell assembly; A data signal decomposition module, using an α-CEEMD signal decomposition method to decompose the multidimensional historical original data signal, adding paired weighted positive noise and negative noise to each of the multidimensional historical original data signals, performing EEMD decomposition on each of the multidimensional historical original data signals after adding the positive noise and the negative noise to obtain an intrinsic mode function IMF, after obtaining all the intrinsic mode functions IMF, averaging all the intrinsic mode functions IMF to obtain a final IMF; the IMF is used as a primary data signal feature matrix; The fault diagnosis module takes the primary data signal feature matrix as input and sends it to the MF-CNN convolutional neural network for further feature extraction; the MF-CNN includes an input layer, a hidden layer, a multi-scale feature extraction module, a feature fusion module, a fully connected layer and an output layer; wherein the multi-scale feature extraction module includes a parallel convolution layer combination and a pooling layer; the feature fusion module includes multi-resolution feature fusion and an attention mechanism; and then the trained MF-CNN is used to perform fault diagnosis on the photovoltaic power generation equipment.

7. A photovoltaic power generation equipment fault diagnosis system according to claim 6, characterized in that: The multi-dimensional historical raw data signal includes: Temperature, including the temperature inside the junction box, the temperature of the solar photovoltaic cell assembly and the temperature of the photovoltaic inverter; the temperature includes the temperature of the junction box, the solar photovoltaic cell assembly and the photovoltaic inverter respectively when a fault occurs; Voltage, including the three-phase voltage of the inverter; the voltage includes the voltage of the inverter when a fault occurs; Ambient irradiance refers to the radiation flux per unit area of ​​the illuminated surface of the solar photovoltaic cell assembly.

8. A photovoltaic power generation equipment fault diagnosis system according to claim 6, characterized in that: The additional time attribute refers to establishing a time axis for the multi-dimensional historical original data signal, and each time point on the time axis has a corresponding multi-dimensional historical original data signal.

9. A photovoltaic power generation equipment fault diagnosis system according to claim 6, characterized in that: The process of the α-CEEMD signal decomposition method is: Set the overall average number to I; Add noise: Construct a complementary set to generate multiple sets of random weighted noise αε(t), and add the noise αε(t) to the original signal x(t) to generate multiple sets of random test signals x i (t); wherein the random noise αε(t) is divided into positive noise and negative noise Among them, the positive noise and the negative noise is a complementary relationship; the test signal x i (t) and the noise formula ε(t) are as follows: x i (t)=x(t)+αε i (t) Wherein, i represents the i-th addition of the noise signal; α is the weight of the α-CEEMD signal decomposition method for noise adaptive adjustment, α is related to the input multidimensional historical original data signal, and the size of α is determined by calculating the oscillation amplitude of the multidimensional historical original data signal. The specific calculation formula is as follows: Among them, f max is the maximum oscillation amplitude of the historical raw data signal of the dimension, f min is the minimum oscillation amplitude of the historical raw data signal of the dimension, that is, f max With f min The greater the absolute value of the difference, the smaller α is, and f max With f min The smaller the absolute value of the difference, the larger α is; Signal decomposition: for the multiple groups of test signals x i (t) Perform EMD decomposition respectively to obtain their respective intrinsic mode functions IMF: Among them, c i,j (t) is the jth IMF obtained by decomposing after adding white noise for the i-th time; r i,j (t) is the residual function, the average trend of the signal; J is the number of IMFs; Repeat the two steps of adding noise and decomposing the signal once, adding one of the complementary positive noise and the negative noise each time, and obtain the set of IMFs as follows: c 1,j (t)c 2,j (t)...c I,j (t),j=1,2,...J Using the principle that the statistical mean of unrelated sequences is zero, the IMF is subjected to ensemble average calculation to obtain the final IMF after EEMD decomposition: i=1,2,...I,j=1,2,...J Among them, c j (t) is the jth IMF of EEMD decomposition; c j (t) is the primary data signal feature matrix.

10. A photovoltaic power generation equipment fault diagnosis system according to claim 6, characterized in that: The input layer takes the primary data signal feature matrix as input, uses a convolution kernel size of 3x3 and a step size of 2; The hidden layer includes a convolution layer and a pooling layer. The convolution layer uses a convolution kernel size of 3x3 and a step size of 2. The pooling layer uses a pooling size of 2x2 and a step size of 2, and the pooling type is maximum pooling. The multi-scale feature extraction module includes a parallel convolution layer combination and a pooling layer, wherein the parallel convolution layer combination uses a plurality of parallel convolution layers, each of the convolution layers has a convolution kernel of a different size, the pooling size used by the pooling layer is 2x2, the step size is 2, and the pooling type is maximum pooling; The feature fusion module includes multi-resolution feature fusion and attention mechanism, wherein the multi-resolution feature fusion fuses feature maps from the input layer, hidden layer and multi-scale feature extraction module, and the attention mechanism enhances the focus on and utilization of key features; The fully connected layer has 256 neural units; The output layer classifies the faults and outputs the final result.

Citation Information

Cited By

  • High tower wind turbine generator fault early warning and diagnosis method and system

    CN119982387A

  • Methods and systems for early warning and diagnosis of faults in high-tower wind turbines

    CN119982387B