A deep learning-based photovoltaic module fault automatic identification method and system

Abstract

The application discloses a kind of photovoltaic module fault automatic identification method and system based on deep learning, belong to photovoltaic module fault identification field, including: first, the material characteristics, structural parameters and environmental conditions of photovoltaic module are acquired, mathematical model is established, and the theoretical output characteristic curve of photovoltaic module under different working conditions is obtained by solving the model.Then, by collecting the current, voltage, power electrical parameters and surface temperature, color, texture image data of photovoltaic module, the measured data is preprocessed, and the key feature data set reflecting the health state is extracted.Next, based on YOLO algorithm and space pyramid pooling module, an optimized YOLO model is constructed, and the theoretical output characteristic curve is used as a constraint condition, the model is trained through the feature data set, and a fault prediction model is obtained, the image data of photovoltaic module is obtained in real time, and it is input into the fault prediction model for calculation, so as to realize the automatic identification and prediction of photovoltaic module fault.

Description

Technical Field

[0001] This invention belongs to the field of photovoltaic module fault identification, and particularly relates to a method and system for automatic fault identification of photovoltaic modules based on deep learning. Background Technology

[0002] As the photovoltaic (PV) power generation installation boom sweeps the globe, the resulting market developments have also raised concerns. The core component of PV modules is the series-connected photovoltaic cells made of crystalline silicon. These cells, constantly exposed to the natural environment, are inevitably susceptible to damage from leaves, dust, bird droppings, and other foreign objects, as well as other mechanical damage, leading to problems such as diode short circuits and internal component failures. During operation, these damaged areas can cause a decrease in power generation or become a load, resulting in abnormal temperature rise and hotspot failures.

[0003] To achieve automatic fault identification in photovoltaic (PV) modules, a comprehensive fault diagnosis model needs to be established. First, various data from the PV module's operation must be collected, including electrical parameters such as current, voltage, and power, as well as image features such as surface temperature, color, and texture. However, this data often contains noise, missing data, and anomalies, requiring preprocessing and feature extraction. Second, a mathematical model of the PV module needs to be constructed to describe its internal structure and working principle, and corresponding diagnostic algorithms should be established based on the fault type and mechanism. However, the complex structure, diverse materials, and variable environmental factors of PV modules make it difficult for the mathematical model to accurately describe their behavior, resulting in insufficient generalization ability of the diagnostic algorithms. Finally, multiple diagnostic algorithms need to be integrated to comprehensively assess the health status of the PV module. However, different algorithms have different applicable conditions, judgment criteria, and decision-making mechanisms. How to coordinate their relationships and balance diagnostic speed and accuracy is a problem that urgently needs to be solved. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a method for automatic fault identification of photovoltaic modules based on deep learning, comprising:

[0005] The material properties, structural parameters and environmental conditions of photovoltaic modules are obtained, a multi-physics coupling mathematical model of photovoltaic modules is established, and the theoretical output characteristic curves of photovoltaic modules under different operating conditions are obtained by solving the mathematical model.

[0006] Acquire measured data and image data during the operation of photovoltaic modules, preprocess the measured data and image data, and extract key feature datasets reflecting the health status of photovoltaic modules; the measured data includes electrical parameters such as current, voltage, and power, as well as image features such as surface temperature, color, and texture;

[0007] An optimized YOLO model is constructed based on the YOLO algorithm and the spatial pyramid pooling module. The theoretical output characteristic curve is used as a constraint condition. The optimized YOLO model is trained using the key feature parameter dataset to obtain a fault prediction model.

[0008] Image data of the photovoltaic module's real-time operation is acquired, and the image data is input into the fault prediction model for calculation to obtain fault identification results.

[0009] Preferably, the process of obtaining the theoretical output characteristic curves of the photovoltaic module under different operating conditions includes:

[0010] Obtain input data on the material properties, structural parameters, and environmental conditions of photovoltaic modules, and establish a mathematical model that considers the multi-physics coupling effect based on the material properties, structural parameters, and environmental conditions.

[0011] Based on the established mathematical model, the finite element analysis method is used to solve the model, and the theoretical output characteristic curves of the photovoltaic module under different operating conditions are obtained.

[0012] The Long Short-Term Memory network is trained using data on material properties, structural parameters, and environmental conditions to establish a prediction model. Based on the input material properties, structural parameters, and environmental conditions, the model predicts the output characteristic curve of the photovoltaic module.

[0013] The theoretical output characteristic curve and the output characteristic curve of the photovoltaic module are fused together to obtain the theoretical output characteristic curve of the photovoltaic module under different operating conditions.

[0014] Preferably, the process of extracting the key feature dataset reflecting the health status of photovoltaic modules includes:

[0015] Acquire measured data and image data during the operation of photovoltaic modules;

[0016] The measured data is labeled as data tags in the image data, and the labeled image data is augmented by an adversarial neural network to obtain the key feature dataset. The adversarial neural network is a conditional adversarial neural network, and the loss function is the cross-entropy loss function.

[0017] Preferably, the process of obtaining the key feature dataset further includes:

[0018] Acquire measured data and image data of photovoltaic modules during operation, including but not limited to current, voltage, and power;

[0019] After removing noise and outliers from the measured data and image data, normalization is performed to obtain a normalized dataset.

[0020] The features of the normalized dataset are reduced in dimensionality using principal component analysis, and key feature parameters are extracted from the dimensionality-reduced dataset using a similarity metric to obtain the key feature dataset.

[0021] Preferably, the process of constructing an optimized YOLO model based on the YOLO algorithm and the spatial pyramid pooling module includes:

[0022] A YOLO algorithm model is constructed by adding the path aggregation module of PANet to the YOLO algorithm model and optimizing the detection head of the YOLO algorithm model to obtain the optimized YOLO model. The model for optimizing the detection head of the YOLO algorithm model is a Bounding Box regression optimization model.

[0023] Preferably, the process of obtaining the fault prediction model includes:

[0024] The key feature parameter dataset is divided into training, validation, and test sets using the k-fold cross-validation method.

[0025] The YOLO model is trained using a training dataset, and the model parameters are optimized using the backpropagation algorithm. The theoretical output characteristic curve is used as a constraint, and the constraint is introduced by modifying the loss function to obtain the trained model.

[0026] The trained model is validated using a validation set, and the model is fine-tuned based on the validation results to obtain a fine-tuned model.

[0027] The generalization ability of the fine-tuned model is tested using a test set to obtain the fault prediction model.

[0028] Preferably, the process of obtaining the fault identification result includes:

[0029] Image data of the real-time operation of photovoltaic modules is acquired, the image data is input into the fault prediction model for calculation, the calculation results are obtained, and an evidence body representation is constructed based on the calculation results.

[0030] The evidence body representation is calculated according to the preset evidence fusion rules to obtain the probability distribution of fault causes;

[0031] Based on the probability distribution of fault causes, the cause of the fault is determined as the final fault diagnosis result. If the maximum probability value in the fault cause probability distribution exceeds a preset threshold, then the fault cause is taken as the final diagnosis result; otherwise, it is determined that the current evidence is insufficient and a definitive diagnosis result cannot be given.

[0032] On the other hand, the present invention also provides an automatic fault identification system for photovoltaic modules based on deep learning, comprising:

[0033] The data acquisition module is used to acquire the material properties, structural parameters and environmental conditions of the photovoltaic module, establish a multi-physics coupling mathematical model of the photovoltaic module, and obtain the theoretical output characteristic curves of the photovoltaic module under different operating conditions by solving the mathematical model.

[0034] The preprocessing module is used to acquire measured data and image data during the operation of photovoltaic modules, and to preprocess the measured data and image data to extract key feature datasets reflecting the health status of photovoltaic modules; the measured data includes electrical parameters such as current, voltage, and power, as well as image features such as surface temperature, color, and texture;

[0035] The training module is used to construct an optimized YOLO model based on the YOLO algorithm and the spatial pyramid pooling module. The theoretical output characteristic curve is used as a constraint condition, and the optimized YOLO model is trained through the key feature parameter dataset to obtain a fault prediction model.

[0036] The identification module is used to acquire image data of the photovoltaic module in real time, input the image data into the fault prediction model for calculation, and obtain the fault identification result.

[0037] On the other hand, the present invention also provides an electronic device, including a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the deep learning-based automatic fault identification method for photovoltaic modules.

[0038] On the other hand, the present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements a method for automatic fault identification of photovoltaic modules based on deep learning.

[0039] Compared with the prior art, the present invention has the following advantages and technical effects:

[0040] This invention provides a photovoltaic module health status prediction and fault diagnosis system based on multiphysics models and image data. It can efficiently and accurately identify and predict module faults, reducing downtime, improving the power generation efficiency of photovoltaic power plants, and lowering operation and maintenance costs. Furthermore, it enhances the accuracy and reliability of fault detection through advanced machine learning algorithms and multidimensional data fusion technology. This technology has significant application value in the photovoltaic power generation field and can provide effective support for intelligent operation and maintenance in the photovoltaic industry. Attached Figure Description

[0041] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0042] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention. Detailed Implementation

[0043] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0044] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0045] Example 1

[0046] like Figure 1 As shown, this embodiment provides a method for automatic fault identification of photovoltaic modules based on deep learning, including:

[0047] The material properties, structural parameters and environmental conditions of photovoltaic modules are obtained, a multi-physics coupling mathematical model of photovoltaic modules is established, and the theoretical output characteristic curves of photovoltaic modules under different operating conditions are obtained by solving the mathematical model.

[0048] Acquire measured data and image data during the operation of photovoltaic modules, preprocess the measured data and image data, and extract key feature datasets reflecting the health status of photovoltaic modules; the measured data includes electrical parameters such as current, voltage, and power, as well as image features such as surface temperature, color, and texture;

[0049] An optimized YOLO model is constructed based on the YOLO algorithm and the spatial pyramid pooling module. The theoretical output characteristic curve is used as a constraint condition. The optimized YOLO model is trained using the key feature parameter dataset to obtain a fault prediction model.

[0050] Image data of the photovoltaic module's real-time operation is acquired, and the image data is input into the fault prediction model for calculation to obtain fault identification results.

[0051] Input data on the material properties, structural parameters, and environmental conditions of photovoltaic (PV) modules were acquired. A mathematical model considering multiphysics coupling effects was established, taking into account these factors. Based on the established mathematical model, the finite element method was used to solve the model, obtaining the theoretical output characteristics of the PV modules under different operating conditions. A prediction model was established by training a large amount of material property, structural parameter, and environmental condition data using a machine learning algorithm. Based on the input material properties, structural parameters, and environmental conditions, the model predicted the output characteristic curves of the PV modules. Output characteristic data of the PV modules under actual operating conditions were acquired and compared with the theoretical output characteristics predicted by the mathematical and machine learning models. The two models were evaluated and optimized. For the optimized mathematical and machine learning models, the output characteristics of the PV modules under different combinations of operating conditions were predicted, generating a series of characteristic curves. Cluster analysis was performed on the generated characteristic curves. Based on the shape and characteristics of the curves, the output characteristics under different operating conditions were divided into several categories, obtaining the classification results of the PV module output characteristics.

[0052] Furthermore, the material properties of photovoltaic modules include the resistivity, carrier concentration, and diffusion length of the silicon wafer; structural parameters include cell size, encapsulation material thickness, and electrode layout; and environmental conditions include light intensity, temperature, and humidity. Taking monocrystalline silicon cells as an example, their resistivity is generally between 0.5 and 5 ohm-cm, their carrier concentration is approximately 10 to the power of 15 per cubic centimeter, and their diffusion length is between 100 and 300 micrometers. Multiphysics coupling models need to consider the interactions of physical processes such as photoelectric conversion, heat conduction, and mechanical stress. For example, increased temperature reduces the bandgap width of the silicon wafer, thus affecting photoelectric conversion efficiency; mechanical stress can lead to microcrack propagation, affecting module lifespan. Finite element analysis can obtain the internal temperature distribution and stress distribution of the module. For instance, under standard test conditions, the surface temperature of the module can reach 45 degrees Celsius, and the stress in the central region can reach 50 MPa.

[0053] The process involves acquiring measured data of photovoltaic modules during operation, including electrical parameters such as current, voltage, and power, as well as image features such as surface temperature, color, and texture. The acquired measured data is preprocessed to remove noise and outliers, and then normalized to ensure comparability across different dimensions. Key feature parameters are extracted from the preprocessed measured data, and principal component analysis is used to reduce the dimensionality of the electrical parameters and image features, yielding key feature parameters that reflect the health status of the photovoltaic modules.

[0054] The acquisition of measured data during photovoltaic module operation needs to consider multiple dimensions. Regarding electrical parameters, current and voltage data can be collected using high-precision current transformers and voltage sensors, and output power can be calculated using a power analyzer. Image feature acquisition can use an infrared thermal imager to obtain surface temperature distribution, and an industrial camera to capture the surface color and texture features of the module. In the data preprocessing stage, median filtering can be used to remove sudden noise for electrical parameters, and Gaussian filtering can be used for smoothing temperature data. Outlier handling uses the three-standard-deviation method to remove data exceeding the normal range. During normalization, current values ​​are typically mapped linearly from zero to the rated current, and voltage values ​​are mapped linearly from zero to the open-circuit voltage. In the feature extraction stage, electrical parameters can extract features such as the maximum power point voltage, current ratio, and power curve slope. Image features can extract parameters such as temperature distribution uniformity, hot spot area ratio, and surface color uniformity. Principal component analysis reduces the multidimensional features to three to five principal components, retaining more than 90% of the information.

[0055] Combining PANet with the YOLO model aims to enhance the expressive power of features, especially in object detection tasks. PANet primarily improves the efficiency of multi-scale feature utilization and information flow, while YOLO is a real-time object detection framework with high speed and accuracy.

[0056] The health status assessment of photovoltaic modules first requires the extraction of key characteristic parameters. Regarding electrical parameters, the deviation rate between the maximum power point voltage and the nominal value can be extracted. Under normal operating conditions, the deviation should not exceed 5%; exceeding this range indicates a potential fault. The slope of the power curve reflects the module's power generation efficiency. By calculating the difference between the slopes of the actual and theoretical power curves, the degree of performance degradation can be determined. During image feature extraction, temperature distribution uniformity can be characterized by the standard deviation; the surface temperature difference of a normal module is typically within three degrees Celsius. When the hot spot area exceeds 10%, it indicates a serious hot spot fault in the module. Feature dimensionality reduction uses principal component analysis (PCA) to transform multiple original features into a small number of principal components. For example, the original features may include ten parameters such as the maximum power point voltage deviation rate, the power curve slope, and temperature uniformity. PCA can reduce this to three to four principal components while retaining over 90% of the information. The first principal component primarily reflects electrical performance, the second reflects temperature characteristics, and the third is related to surface morphology.

[0057] PANet is primarily used to enhance the capabilities of the Feature Pyramid Network (FPN), improving feature flow and fusion, thereby increasing the detection accuracy of small objects. The basic steps for combining PANet with the YOLO model are as follows:

[0058] YOLO uses CNNs as its feature extraction network (e.g., Darknet). The first step in integrating PANet into YOLO is to replace or enhance its existing feature extraction network. A PANet module can be added after the YOLO backbone (such as Darknet-53 or other deep networks) to perform multi-scale feature aggregation. PANet's path aggregation module is inserted between the high-level and low-level features output from the YOLO backbone. These features at different scales are then fused through PANet's feature fusion layer to better capture fine-grained object features.

[0059] Optimizing the detection head of the YOLO (You Only Look Once) algorithm typically involves improvements and adjustments in multiple aspects, aiming to enhance detection accuracy, speed, or adaptability to specific tasks. The YOLO detection head usually refers to the last few layers of the network, used to generate the final object detection results (including bounding boxes, class predictions, etc.) from feature maps. The optimization process includes the following main aspects: optimizing the loss function for bounding box regression, employing more suitable regression losses (such as CIoU (Complete Intersection over Union) or GIoU (Generalized Intersection over Union)). Compared to traditional IoU, CIoU and GIoU can better handle shape and positional deviations of bounding boxes. Simultaneously, class-balanced losses (such as Focal Loss) are used to optimize the accuracy of class predictions. Focal Loss can reduce the focus on easily classified samples and increase the focus on difficult-to-classify samples, thereby improving the overall performance of the model.

[0060] Design a loss function suitable for fault prediction tasks, typically combining regression and classification losses. The loss function should consider not only prediction accuracy but also the error between the theoretical output characteristic curve and the actual prediction results. The theoretical output characteristic curve is used as a constraint, introduced by modifying the loss function to ensure the model follows specific physical laws or expected characteristics during training. The key feature parameter dataset is divided into training, validation, and test sets. k-fold cross-validation can be used to improve the model's generalization ability. The YOLO model is trained using the training dataset, and its parameters are optimized using backpropagation. Hyperparameters such as learning rate, batch size, and optimizer are adjusted to achieve optimal training results. During training, the theoretical output characteristic curve constraint is introduced to ensure the model's output conforms to actual physical or system characteristics. The validation set is used to adjust the model and avoid overfitting. Evaluation metrics include precision, recall, and F1 score to ensure the model performs well in fault prediction. The trained model is then tested using the test set to verify its generalization ability.

[0061] Evaluation indicators:

[0062] Accuracy: The percentage of correct predictions of faults or normal states.

[0063] Recall: The proportion of actual failure instances that the model can correctly predict.

[0064] F1 score: The harmonic mean that takes into account both precision and recall.

[0065] Fault prediction time accuracy: Evaluate the model's accuracy in predicting the timing of fault occurrence.

[0066] Robustness testing: Testing the robustness of the model under different noise, interference, or data distribution variations.

[0067] Obtaining measured data from photovoltaic (PV) modules is fundamental to assessing their health status. This data typically includes electrical parameters such as current, voltage, and power, as well as visual features like surface temperature, color, and texture. For example, under standard test conditions, a PV module might have a measured current of 8.5 amps, a voltage of 35 volts, a power of 297.5 watts, and a surface temperature of 45 degrees Celsius. Color and texture are captured using a high-resolution camera and converted into digital features. Preprocessing the measured data is a crucial step in ensuring data quality. Preprocessing includes noise and outlier removal, as well as data normalization. For example, filtering algorithms remove high-frequency noise from the current data, resulting in a smooth current curve. Normalization converts parameters such as current and voltage into values ​​between 0 and 1, facilitating subsequent analysis. The preprocessed data is then input into a pre-built machine learning model, such as a Support Vector Machine (SVM) or Neural Network (NN). The model predicts the health status of the PV module based on the input features. For example, the model might output "normal," "mildly degraded," or "severely degraded." Finally, the theoretical output characteristic curve of the PV module is obtained as a reference standard for its health status. For example, under standard conditions, the theoretical output power of a certain type of photovoltaic module is 300 watts, while the actual output power is 290 watts. By comparing the actual and theoretical outputs, the module performance is evaluated. The deviation between the measured data and the theoretical output characteristic curve is calculated, i.e., the deviation value.

[0068] Based on historical fault data and expert experience of photovoltaic (PV) modules, a knowledge base for PV module fault diagnosis is constructed, containing various fault modes and their corresponding diagnostic rules. A decision tree algorithm is used to construct a fault tree model of the PV modules based on the fault modes and diagnostic rules in the knowledge base, with leaf nodes corresponding to specific fault types. Real-time monitoring data of the PV modules, including parameters such as voltage, current, and temperature, is acquired. Based on data preprocessing and feature extraction, the measured data feature vectors of the PV modules are obtained. These measured data feature vectors are input into the fault tree model. Through the fault tree's reasoning and diagnostic process, it is determined whether the measured data meets the diagnostic rules for a specific fault mode. If the measured data meets the diagnostic rules for a fault mode, the PV module is determined to have a corresponding fault based on the fault tree's diagnostic results, and the specific fault type and location are output. If the measured data does not meet the diagnostic rules for any fault mode, the PV module is considered to be operating normally and no fault repair or handling is required. Based on the PV module fault diagnosis results, a fault report and maintenance suggestions are generated to guide maintenance personnel in handling the faults, thereby improving the reliability and power generation efficiency of the PV power plant.

[0069] Based on historical fault data and expert experience of photovoltaic (PV) modules, a knowledge base for PV module fault diagnosis is constructed. This knowledge base contains various fault modes and their corresponding diagnostic rules. For example, the knowledge base may include common fault modes such as "hot spot effect," "short circuit fault," and "open circuit fault," each with a series of diagnostic rules. Diagnostic rules for the hot spot effect might include: abnormal temperature increase, local current decrease, etc. Rules for short circuit faults might include: voltage drop, current surge, etc. A decision tree algorithm is used to construct a fault tree model of the PV module based on the fault modes and diagnostic rules in the knowledge base. The leaf nodes of the fault tree correspond to specific fault types. For example, the root node of the decision tree might be "abnormal voltage," branch nodes might include "overvoltage" and "undervoltage," and the final leaf node might be "short circuit fault" or "open circuit fault." In this way, the fault tree model can systematically infer specific fault types. Real-time monitoring data of the PV modules is acquired, including parameters such as voltage, current, and temperature. For example, a monitoring data point might show a voltage of 220V, a current of 5A, and a temperature of 45℃. Based on data preprocessing and feature extraction, the measured data feature vector of the photovoltaic module is obtained. Preprocessing may include steps such as noise removal and normalization, while feature extraction may include calculating voltage fluctuation range and current change rate. The measured data feature vector of the photovoltaic module is input into a fault tree model. Through the fault tree's reasoning and diagnostic process, it is determined whether the measured data meets the diagnostic rules for a specific fault mode. For example, if the measured data shows a sudden voltage drop to 150V, a surge in current to 10A, and no significant temperature change, the fault tree model will infer step by step based on these characteristics, and may ultimately determine it as a "short circuit fault." If the measured data meets the diagnostic rules for a specific fault mode, the photovoltaic module is determined to have experienced a corresponding fault based on the fault tree's diagnostic results, and the specific fault type and location are output. For example, the system may output "short circuit fault, located in the 3rd row, 5th column photovoltaic module." Such diagnostic results can guide maintenance personnel to quickly locate and handle the fault. If the measured data does not meet the diagnostic rules for any fault mode, the photovoltaic module is determined to be operating normally, and no fault repair or handling is required. For example, if measured data shows that voltage, current, and temperature fluctuate within normal ranges, the system will determine that the photovoltaic module is healthy and requires no further intervention. Based on the fault diagnosis results of the photovoltaic module, a fault report and maintenance suggestions are generated to guide maintenance personnel in troubleshooting. For example, the fault report may include information such as fault type, fault location, and fault occurrence time, while the maintenance suggestions may include specific measures such as replacing the faulty module or checking the connection lines.

[0070] The fault diagnosis results of photovoltaic modules include several key parameters, such as current, voltage, and temperature. Taking monocrystalline silicon photovoltaic modules as an example, the output voltage during normal operation should be above 90% of the nominal value, the operating temperature should be controlled below 70 degrees Celsius, and the current value should be positively correlated with the light intensity. When these parameters are abnormal, such as the voltage being lower than 80% of the nominal value or the temperature exceeding 80 degrees Celsius, it may indicate that the module has a fault. Remaining lifetime prediction based on physical models needs to consider multiple influencing factors. Taking the glass encapsulation material of photovoltaic modules as an example, its aging rate is closely related to ambient temperature, humidity, and ultraviolet radiation intensity. For every 10-degree increase in ambient temperature, the material aging rate approximately doubles. In high-temperature and high-humidity environments, the lifespan of the encapsulation material may be shortened to two-thirds of its original design lifespan. The establishment of predictive models requires combining historical data and on-site monitoring data. For example, after five years of operation, monitoring revealed that the output power of modules in a photovoltaic power station had decreased by 10%. Considering local climate conditions and the usage environment, the remaining lifetime is predicted to be approximately 15 years. This prediction is based on the module's performance degradation curve and environmental stress analysis. Setting the remaining lifetime threshold requires consideration of multiple aspects. Typically, a component's output power dropping to 80% of its rated power is used as the replacement threshold. For example, a photovoltaic module with a rated power of 400 watts might be considered for replacement when its actual output power falls below 320 watts. Furthermore, the module's appearance must be considered; obvious yellowing or bubbles should also prompt immediate replacement. Generating maintenance or replacement tasks requires comprehensive evaluation. When the predicted remaining lifespan falls below the threshold, the system generates a task instruction. For instance, if a photovoltaic power station's monitoring system detects that a group of modules has a predicted remaining lifespan of only two years, while the preset threshold is three years, the system will generate a replacement task instruction. This instruction includes information such as the module's location, model, and fault type, facilitating accurate location and handling by maintenance personnel. Upon receiving the task instruction, the operation and maintenance management system will schedule maintenance based on task priority. For example, modules with severely degraded output power will be classified as emergency maintenance items, requiring replacement within 48 hours. Modules with minor abnormalities can be handled during routine maintenance. This tiered approach ensures system safety while optimizing the efficiency of maintenance resource utilization.

[0071] On the other hand, this embodiment also provides a photovoltaic module fault automatic identification system based on deep learning, including:

[0072] The data acquisition module is used to acquire the material properties, structural parameters and environmental conditions of the photovoltaic module, establish a multi-physics coupling mathematical model of the photovoltaic module, and obtain the theoretical output characteristic curves of the photovoltaic module under different operating conditions by solving the mathematical model.

[0073] The preprocessing module is used to acquire measured data and image data during the operation of photovoltaic modules, and to preprocess the measured data and image data to extract key feature datasets reflecting the health status of photovoltaic modules; the measured data includes electrical parameters such as current, voltage, and power, as well as image features such as surface temperature, color, and texture;

[0074] The training module is used to construct an optimized YOLO model based on the YOLO algorithm and the spatial pyramid pooling module. The theoretical output characteristic curve is used as a constraint condition, and the optimized YOLO model is trained through the key feature parameter dataset to obtain a fault prediction model.

[0075] The identification module is used to acquire image data of the photovoltaic module in real time, input the image data into the fault prediction model for calculation, and obtain the fault identification result.

[0076] On the other hand, this embodiment also provides an electronic device, including a memory, a processor, and a computing program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the deep learning-based automatic fault identification method for photovoltaic modules.

[0077] On the other hand, this embodiment also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements a method for automatic fault identification of photovoltaic modules based on deep learning.

[0078] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for automatic fault identification of photovoltaic modules based on deep learning, characterized in that, include: The material properties, structural parameters and environmental conditions of photovoltaic modules are obtained, a multi-physics coupling mathematical model of photovoltaic modules is established, and the theoretical output characteristic curves of photovoltaic modules under different operating conditions are obtained by solving the mathematical model. Acquire measured data and image data during the operation of photovoltaic modules, preprocess the measured data and image data, and extract key feature datasets reflecting the health status of photovoltaic modules; the measured data includes electrical parameters such as current, voltage, and power, as well as image features such as surface temperature, color, and texture; An optimized YOLO model is constructed based on the YOLO algorithm and the spatial pyramid pooling module. The theoretical output characteristic curve is used as a constraint condition. The optimized YOLO model is trained using the key feature parameter dataset to obtain a fault prediction model. Image data of the photovoltaic module's real-time operation is acquired, and the image data is input into the fault prediction model for calculation to obtain fault identification results.

2. The method according to claim 1, characterized in that, The process of obtaining the theoretical output characteristic curves of photovoltaic modules under different operating conditions includes: Obtain input data on the material properties, structural parameters, and environmental conditions of photovoltaic modules, and establish a mathematical model that considers the multi-physics coupling effect based on the material properties, structural parameters, and environmental conditions. Based on the established mathematical model, the finite element analysis method is used to solve the model, and the theoretical output characteristic curves of the photovoltaic module under different operating conditions are obtained. The Long Short-Term Memory network is trained using data on material properties, structural parameters, and environmental conditions to establish a prediction model. Based on the input material properties, structural parameters, and environmental conditions, the model predicts the output characteristic curve of the photovoltaic module. The theoretical output characteristic curve and the output characteristic curve of the photovoltaic module are fused together to obtain the theoretical output characteristic curve of the photovoltaic module under different operating conditions.

3. The method according to claim 1, characterized in that, The process of extracting the key feature dataset reflecting the health status of photovoltaic modules includes: Acquire measured data and image data during the operation of photovoltaic modules; The measured data is labeled as data tags in the image data, and the labeled image data is augmented by an adversarial neural network to obtain the key feature dataset. The adversarial neural network is a conditional adversarial neural network, and the loss function is the cross-entropy loss function.

4. The method according to claim 3, characterized in that, The process of obtaining the key feature dataset also includes: Acquire measured data and image data of photovoltaic modules during operation, including but not limited to current, voltage, and power; After removing noise and outliers from the measured data and image data, normalization is performed to obtain a normalized dataset. The features of the normalized dataset are reduced in dimensionality using principal component analysis, and key feature parameters are extracted from the dimensionality-reduced dataset using a similarity metric to obtain the key feature dataset.

5. The method according to claim 1, characterized in that, The process of building an optimized YOLO model based on the YOLO algorithm and spatial pyramid pooling module includes: A YOLO algorithm model is constructed by incorporating the path aggregation module of PANet into the YOLO algorithm model and optimizing the detection head of the YOLO algorithm model to obtain the optimized YOLO model. The model for optimizing the detection head of the YOLO algorithm model is a Bounding Box regression optimization model.

6. The method according to claim 1, characterized in that, The process of obtaining the fault prediction model includes: The key feature parameter dataset is divided into training, validation, and test sets using the k-fold cross-validation method. The YOLO model is trained using a training dataset, and the model parameters are optimized using the backpropagation algorithm. The theoretical output characteristic curve is used as a constraint, and the constraint is introduced by modifying the loss function to obtain the trained model. The trained model is validated using a validation set, and the model is fine-tuned based on the validation results to obtain a fine-tuned model. The generalization ability of the fine-tuned model is tested using a test set to obtain the fault prediction model.

7. The method according to claim 1, characterized in that, The process of obtaining fault identification results includes: Image data of the real-time operation of photovoltaic modules is acquired, the image data is input into the fault prediction model for calculation, the calculation results are obtained, and an evidence body representation is constructed based on the calculation results. The evidence body representation is calculated according to the preset evidence fusion rules to obtain the probability distribution of fault causes; Based on the probability distribution of fault causes, the cause of the fault is determined as the final fault diagnosis result. If the maximum probability value in the fault cause probability distribution exceeds a preset threshold, then the fault cause is taken as the final diagnosis result; otherwise, it is determined that the current evidence is insufficient and a definitive diagnosis result cannot be given.

8. A photovoltaic module fault automatic identification system based on deep learning, characterized in that, include: The data acquisition module is used to acquire the material properties, structural parameters and environmental conditions of the photovoltaic module, establish a multi-physics coupling mathematical model of the photovoltaic module, and obtain the theoretical output characteristic curves of the photovoltaic module under different operating conditions by solving the mathematical model. The preprocessing module is used to acquire measured data and image data during the operation of photovoltaic modules, and to preprocess the measured data and image data to extract key feature datasets reflecting the health status of photovoltaic modules; the measured data includes electrical parameters such as current, voltage, and power, as well as image features such as surface temperature, color, and texture; The training module is used to construct an optimized YOLO model based on the YOLO algorithm and the spatial pyramid pooling module. The theoretical output characteristic curve is used as a constraint condition, and the optimized YOLO model is trained through the key feature parameter dataset to obtain a fault prediction model. The identification module is used to acquire image data of the photovoltaic module in real time, input the image data into the fault prediction model for calculation, and obtain the fault identification result.

9. An electronic device comprising a memory, a processor, and a computing program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method according to any one of claims 1-7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1-7.

Images