A method and system for identifying low efficiency types of photovoltaic modules

Abstract

The application discloses a photovoltaic module low-efficiency type identification method and system, and belongs to the technical field of intelligent operation and maintenance management, and comprises the following steps: S1, feature extraction is performed on multidimensional data respectively through a neural network to obtain a multidimensional feature vector, if there is a feature vector meeting a preset type judgment condition in the multidimensional feature vector, an ultimate low-efficiency type is obtained through the feature vector meeting the preset type judgment condition, otherwise, S2 is executed; S2, an influence weight of the multidimensional feature vector on a preset low-efficiency type is obtained respectively based on a random forest, and an environmental feature vector, an image feature vector and an operation feature vector in the multidimensional feature vector are fused to obtain a fusion feature vector based on the influence weight; S3, a classification model constructed by using a full connection neural network is used to obtain an ultimate low-efficiency type based on the fusion feature vector and a preset low-efficiency type corresponding to the influence weight in the fusion feature vector. The technical problem that the prior art is difficult to improve the low-efficiency type identification accuracy is solved.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance management technology, specifically to a method and system for identifying inefficient types of photovoltaic modules. Background Technology

[0002] With the explosive growth of the photovoltaic industry and the continuous expansion of installed photovoltaic capacity, photovoltaic modules, as the core power generation unit of photovoltaic power plants, directly determine the power generation efficiency of the plant. However, during long-term service, photovoltaic modules are prone to inefficiencies such as power degradation, appearance defects, and electrical faults due to multiple factors including environmental factors, equipment characteristics, and maintenance levels. These inefficiencies result in significant power generation losses, severely restricting the sustainable development of photovoltaic power plants. Therefore, identifying inefficiency types and taking targeted measures is crucial for ensuring the sustainable development of photovoltaic power plants. Traditional photovoltaic module inspection relies on manual inspection and simple instrument measurements, which are no longer adequate for the operation and maintenance needs of large-scale photovoltaic power plants. With the development of artificial intelligence technology, AI-based inspection methods have gradually become a research hotspot. Automatic identification of inefficiencies through computer vision, machine learning, and other technologies provides a new technological path for intelligent operation and maintenance of photovoltaic power plants. However, most studies rely on a single type of data for analysis, such as detecting appearance defects based solely on image data or judging performance status solely through operational parameter data. While a single type of data can accurately identify inefficient types with relatively prominent characteristics, it is prone to producing multiple incorrect identification results for inefficient types with hidden features and inconspicuous manifestations, making accurate identification difficult. Therefore, how to improve the accuracy of inefficient type identification is a technical challenge that current technologies struggle to solve. Summary of the Invention

[0003] To address the technical problem of existing technologies' difficulty in improving the accuracy of inefficiency type identification, this invention provides a method and system for identifying inefficiency types of photovoltaic modules. When an environmental feature vector, image feature vector, or operational feature vector contains a feature vector that meets preset type determination conditions, the final inefficiency type can be accurately determined based on this feature vector. If none of the three feature vectors meet the preset determination conditions, the influence weights of the environmental feature vector, image feature vector, and operational feature vector on the preset inefficiency type are calculated using a random forest algorithm. Based on these weights, the three feature vectors are weighted and fused to obtain a fused feature vector. This fused feature vector is then input into a classification model built on a fully connected neural network to complete the inefficiency type identification. By combining the preset inefficiency type corresponding to the influence weights of the fused feature vector with the output of the classification model for mutual verification, the accuracy of photovoltaic module inefficiency type identification is significantly improved. This solves the technical problem of existing technologies' difficulty in improving the accuracy of inefficiency type identification.

[0004] To address the aforementioned technical problems, this invention provides a method for identifying the inefficiency type of photovoltaic modules, comprising the following steps: S1: Use a neural network to extract features from environmental data, image data, and runtime data to obtain environmental feature vectors, image feature vectors, and runtime feature vectors respectively. If there are feature vectors among the environmental feature vectors, image feature vectors, and runtime feature vectors that meet the preset type determination conditions, then obtain the final inefficient type through the feature vectors that meet the preset type determination conditions; otherwise, execute S2. S2: Based on random forest, obtain the influence weights of environmental feature vector, image feature vector and running feature vector on the preset inefficient type respectively, and fuse the environmental feature vector, image feature vector and running feature vector based on the influence weights to obtain the fused feature vector; S3: Based on the fused feature vector and the preset inefficient type corresponding to the influencing weight in the fused feature vector, the classification model constructed using a fully connected neural network obtains the final inefficient type.

[0005] Preferably, in S1, before extracting environmental feature vectors, image feature vectors, and operational feature vectors from environmental data, image data, and operational data using a neural network, the method further includes: Image data were preprocessed using median filtering and histogram equalization, respectively. Outliers in environmental and operational data were identified and removed using the Laida criterion, and missing values ​​in environmental and operational data were filled using mean imputation.

[0006] Preferably, in S1, the step of extracting environmental feature vectors, image feature vectors, and operational feature vectors from environmental data, image data, and operational data using a neural network includes: Image feature vectors are obtained by extracting features from image data using a convolutional neural network. Environmental feature vectors and operational feature vectors are obtained by extracting features from environmental data and operational data using Long Short-Term Memory Network and Transformer, respectively.

[0007] Preferably, in S2, the step of obtaining the influence weights of the environmental feature vector, image feature vector, and running feature vector on the preset inefficient type based on random forest includes: Obtain historical multidimensional feature vectors, and train the random forest to obtain the final random forest model based on the correlation between the historical multidimensional feature vectors and the preset inefficient type. Obtain the influence weights of the environmental feature vector, image feature vector, and running feature vector on the preset inefficient type based on the split gain of the historical environmental feature vector, historical image feature vector, and historical running feature vector in the final random forest model.

[0008] Preferably, the step of training the random forest to obtain the final random forest model based on the correlation between historical multidimensional feature vectors and preset inefficient types includes: In sequence, any one of the preset inefficient types is taken as the first category, and the remaining preset inefficient types are taken as the second category to construct an initial binary classification tree. A random forest is then constructed based on the initial binary classification tree. The final random forest model is obtained by training the random forest using the association between historical multidimensional feature vectors and preset inefficient types.

[0009] Preferably, the step of obtaining the influence weights of the environment feature vector, image feature vector, and running feature vector on the preset inefficient type based on the splitting gain of the historical environment feature vector, historical image feature vector, and historical running feature vector in the final random forest model includes: Obtain the sub-split gains of historical environment feature vectors, historical image feature vectors, and historical running feature vectors in the final binary classification tree of the final random forest model. Use the sum of the sub-split gains of historical environment feature vectors as the split gain of environment feature vectors in the final random forest model, the sum of the sub-split gains of historical image feature vectors as the split gain of image feature vectors in the final random forest model, and the sum of the sub-split gains of historical running feature vectors as the split gain of running feature vectors in the final random forest model. The splitting gain of the environmental feature vector, image feature vector, and running feature vector in the final random forest model is normalized to obtain the influence weights of the environmental feature vector, image feature vector, and running feature vector on the preset inefficient type.

[0010] Preferably, step S3: obtaining the final inefficient type using a classification model constructed with a fully connected neural network based on the fused feature vector and the preset inefficient types corresponding to the influencing weights in the fused feature vector, includes: S31: Train the classification model based on the acquired historical multidimensional feature vectors to obtain the confidence of the first inefficient type. Based on the confidence of the first inefficient type and the actual inefficient type of the historical multidimensional feature vectors, correct the model parameters in the classification model to obtain the final classification model. S32: Based on the fused feature vector and the preset inefficient type corresponding to the influence weight in the fused feature vector, the final inefficient type is obtained using the final classification model.

[0011] Preferably, step S32: based on the fused feature vector and the preset inefficient types corresponding to the influencing weights in the fused feature vector, the final inefficient type is obtained using the final classification model, including: The fused feature vector is input into the final classification model to obtain the confidence level of the second inefficient type. If the confidence levels of the second inefficient type are all less than or equal to the first preset value, it indicates that the photovoltaic module is normal. If there is a second inefficient type with a confidence level greater than the first preset value, and the confidence levels of the second inefficient type are all less than or equal to the second preset value, an error is reported. If there is a second inefficient type with a confidence level greater than the second preset value, and the inefficient type corresponding to the largest second inefficient type confidence level matches the preset inefficient type corresponding to the influence weight in the fused feature vector, then the inefficient type corresponding to the largest second inefficient type confidence level is the final inefficient type.

[0012] By adopting the above technical solution, the present invention has the following advantages: If any of the environmental feature vectors, image feature vectors, and operational feature vectors satisfy the preset type determination conditions, the final inefficient type can be accurately determined based on this feature vector. If none of the three feature vectors satisfy the preset determination conditions, the influence weights of the environmental feature vector, image feature vector, and operational feature vector on the preset inefficient type are calculated using the random forest algorithm. Based on these weights, the three feature vectors are weighted and fused to obtain a fused feature vector. The fused feature vector is then input into a classification model built on a fully connected neural network to complete the inefficient type identification. By combining the preset inefficient type corresponding to the influence weight of the fused feature vector with the output of the classification model for mutual verification, the accuracy of photovoltaic module inefficient type identification is greatly improved, solving the technical problem that existing technologies cannot improve the accuracy of inefficient type identification. By constructing a binary classification tree with a single preset inefficient type as an independent category and other types as another category, each tree focuses on the feature differentiation logic of a single inefficient type. This avoids the feature importance quantification bias caused by multi-category classification and ensures that the split gain truly reflects the actual judgment value of each type of feature vector for a specific preset inefficient type. By accumulating the sub-split gains of the three types of feature vectors in all binary classification trees of the final random forest, the randomness and overfitting risk of a single decision tree are offset, improving the accuracy of the weights. In addition, the fused feature vector is input into a classification model with nonlinear fitting capabilities to obtain the output. The preset inefficient type corresponding to the influence weight of the fused feature vector is used for bidirectional verification. This not only fully explores the nonlinear correlation between features but also eliminates the random error of algorithm fitting through physical law verification, thereby significantly improving the accuracy of photovoltaic module inefficient type identification.

[0013] This invention also provides a photovoltaic module inefficiency type identification system, applicable to the aforementioned photovoltaic module inefficiency type identification method, comprising: The explicit inefficient type acquisition module is used to extract features from environmental data, image data, and runtime data through a neural network to obtain environmental feature vectors, image feature vectors, and runtime feature vectors, respectively. It then determines whether there are feature vectors among these vectors that meet preset type determination conditions. If a feature vector that meets the preset type determination conditions is found, the final inefficient type is obtained through the feature vector that meets the preset type determination conditions. Otherwise, the feature vector fusion module is activated. The feature vector fusion module is used to obtain the influence weights of environmental feature vectors, image feature vectors, and running feature vectors on a preset inefficient type based on random forest, and then fuse the environmental feature vectors, image feature vectors, and running feature vectors based on the influence weights to obtain a fused feature vector. The implicit inefficient type acquisition module is used to obtain the final inefficient type based on the fused feature vector and the preset inefficient type corresponding to the influencing weight in the fused feature vector, using a classification model built with a fully connected neural network.

[0014] By adopting the above technical solution, the present invention has the following advantages: If any of the environmental feature vectors, image feature vectors, and operational feature vectors meets the preset type determination criteria, the final inefficiency type can be accurately determined based on this feature vector. If none of the three feature vectors meet the preset criteria, the influence weights of the environmental feature vector, image feature vector, and operational feature vector on the preset inefficiency type are calculated using a random forest algorithm. Based on these weights, the three feature vectors are weighted and fused to obtain a fused feature vector. This fused feature vector is then input into a classification model built on a fully connected neural network to complete the inefficiency type identification. By combining the preset inefficiency type corresponding to the influence weights of the fused feature vector with the output of the classification model for mutual verification, the accuracy of photovoltaic module inefficiency type identification is significantly improved. This solves the technical problem that existing technologies struggle to improve the accuracy of inefficiency type identification.

[0015] The present invention also provides a computer device, including: a processor, a memory, and a bus, wherein the memory stores machine-readable instructions executable by the processor, and when the computer device is running, the processor communicates with the memory via the bus, and the processor executes the machine-readable instructions to perform the steps of the photovoltaic module inefficiency type identification method. Attached Figure Description

[0016] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings. The drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings.

[0017] Figure 1 This is a flowchart illustrating a method for identifying inefficient photovoltaic modules according to the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only one preferred embodiment of this invention and are only used to explain this invention. They do not limit the scope of protection of this invention. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0019] Before discussing the exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe the operations (or steps) as sequential processes, many of the operations (or steps) can be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the operations can be rearranged. The process can be terminated when its operation is completed, but it may also have additional steps not included in the figures; the process may correspond to a method, function, procedure, subroutine, subroutine, etc.

[0020] Example 1: like Figure 1 As shown, a method for identifying the inefficiency type of photovoltaic modules includes the following steps: S1: Use a neural network to extract features from environmental data, image data, and runtime data to obtain environmental feature vectors, image feature vectors, and runtime feature vectors. If there are feature vectors among the environmental feature vectors, image feature vectors, and runtime feature vectors that meet the preset type determination conditions, then obtain the final inefficient type through the feature vectors that meet the preset type determination conditions; otherwise, execute S2.

[0021] The preset type determination criteria are quantitative judgment rules set for three types of feature vectors: environment, image, and operation. The core is to transform the physical, visual, and operational patterns of various inefficiency types of photovoltaic modules into numerical thresholds / feature distribution rules for feature vectors. When the value of a certain type of feature vector meets the corresponding rule, it can be directly determined as a matching inefficiency type. For example, if the normal operating temperature of a photovoltaic module is less than or equal to 60 degrees Celsius, and the operating temperature is more than 50% higher than the normal operating temperature, it is considered an appearance defect. In this case, if the temperature in the image feature vector is more than 50% higher than the normal operating temperature, it indicates the existence of a feature vector that meets the preset type determination criteria, and the final inefficiency type is an appearance defect. It is understandable that there are multiple determination criteria in the preset type determination criteria, specifically set according to the number of features in the normal operating condition benchmark, environmental feature vector, image feature vector, and operational feature vector.

[0022] In some embodiments, before S1, the step of extracting environmental feature vectors, image feature vectors, and operational feature vectors from environmental data, image data, and operational data respectively using a neural network further includes: Image data were preprocessed using median filtering and histogram equalization, respectively. Outliers in environmental and operational data were identified and removed using the Laida criterion, and missing values ​​in environmental and operational data were filled using mean imputation.

[0023] In this embodiment, the images include visible light images and infrared thermal images. Specifically, visible light images (for detecting surface defects such as cracks, delamination, and corrosion) and infrared thermal images (for detecting temperature anomalies such as hot spots) of the photovoltaic modules are collected through drone inspections and ground-based high-definition cameras. Operating data such as module voltage, current, output power, and module temperature are collected through the power station's SCADA system and sensors to reflect the electrical performance status. Environmental data such as light intensity, ambient temperature, humidity, and wind speed are collected through weather stations to analyze the impact of environmental factors on module performance. Image data is collected 1-2 times per month (adjustable according to maintenance needs), operating data is collected once per minute, and environmental data is collected once every 5 minutes. Median filtering is used to remove image noise, histogram equalization is used to enhance image contrast, and the image is normalized (uniform size and pixel range) to lay the foundation for feature extraction. The Laida criterion is used to identify and remove outliers in the operating parameters and environmental data, the mean imputation method is used to fill missing values, and the data is standardized (converted to the [0,1] interval) to ensure data consistency and accuracy.

[0024] In some embodiments, in S1, the step of extracting environmental feature vectors, image feature vectors, and operational feature vectors from environmental data, image data, and operational data using a neural network includes: Image feature vectors are obtained by extracting features from image data using a convolutional neural network. Environmental feature vectors and operational feature vectors are obtained by extracting features from environmental data and operational data using Long Short-Term Memory Network and Transformer, respectively.

[0025] In this embodiment, the convolutional neural network specifically adopts a CNN architecture of three convolutional layers + two pooling layers + one fully connected layer. The detailed steps for extracting image features using a convolutional neural network are as follows: feature extraction is performed separately for visible light images (appearance defect detection) and infrared thermal imaging images (hot spot detection). The specific steps are as follows: 1. Image input preprocessing and adaptation: Input data standardization: The preprocessed 256×256 pixel image is converted into tensor format. The visible light image is 3-channel (RGB), and the infrared thermal imaging image is a single channel (temperature gray value). The pixel values ​​are uniformly normalized to the [0,1] interval. Data augmentation and adaptation: In order to improve the generalization ability of the model, the input image is subjected to augmentation operations such as random flipping (horizontal / vertical), random cropping (224×224 pixels), and Gaussian noise addition (variance 0.01) to generate an expanded dataset. 2. Convolutional Layer Feature Extraction: First Convolutional Layer (Low-Level Feature Extraction): Convolutional Kernel Settings: 32 convolutional kernels of size 3×3, stride = 1, padding = 1 (ensuring the output feature map size is consistent with the input); Convolution Calculation: Each convolutional kernel performs sliding window calculation on the input image, capturing low-level features such as edges, textures, and gray-level abrupt changes through local receptive fields; Output: Generates 32 256×256 feature maps, each corresponding to a low-level edge / texture feature. 3. First Pooling Layer (Downsampling and Feature Preservation): Pooling Method: Max Pooling, kernel size 2×2, stride = 2; Calculation Process: Non-overlapping sliding window sampling is performed on the feature map output from the first convolutional layer, and the maximum value within each window is taken as the output pixel; Output: The feature map size is compressed to 128×128, retaining key features while reducing computation and suppressing overfitting. 4. Second convolution (intermediate feature extraction): Convolution kernel settings: 64 3×3 convolution kernels, stride=1, padding=1; Convolution calculation: Based on the feature map after the first pooling layer, low-level features are combined through convolution kernels to capture intermediate features such as local shape of the image (e.g., linear features of cracks, regional features of delamination), gray-level distribution patterns (e.g., mottled features of corrosion), etc. The calculation logic is the same as the first convolution layer, and the activation function is still ReLU; Output: 64 128×128 feature maps are generated, realizing the aggregation of low-level features into intermediate features. 5. Second pooling (further downsampling): Pooling parameters: max pooling, 2×2 pooling kernel, stride=2; Output: The feature map size is compressed to 64×64, strengthening the spatial invariance of features (i.e., the positional shift of features in the image does not affect the recognition result). 6. Third layer convolution (advanced feature extraction): Convolution kernel settings: 128 3×3 convolution kernels, stride=1, padding=1; Convolution calculation: fuse the intermediate features after the second layer pooling to capture high-level features with semantic information (such as "crack continuity area" and "delamination boundary" in visible light images, and "hot spot core area" and "temperature gradient distribution" in infrared thermal imaging images). The activation function is LeakyReLU to avoid gradient vanishing; Output result: generate 128 64×64 high-level feature maps, comprehensively covering the core features of component appearance defects and hot spots. 7. Feature Map Flattening and Fully Connected Mapping: Flattening: The 128×64×64 feature map output from the third convolutional layer is converted into a one-dimensional vector with a length of 128×64×64=524,288. Fully Connected Layer Mapping: The flattened vector is processed by a fully connected layer (with 1024 hidden nodes) to integrate features and compress dimensions. Output Image Feature Vector: The fully connected layer outputs a 256-dimensional feature vector, corresponding to the visible light image feature vector and the infrared thermal imaging image feature vector, respectively. Finally, a 512-dimensional comprehensive image feature vector is obtained by concatenating the vectors.

[0026] In this embodiment, the first environmental feature vector and the second environmental feature vector, which are respectively extracted from the environmental data by the Long Short-Term Memory Network and the Transformer, are weighted and fused to obtain the environmental feature vector. The first running feature vector and the second running feature vector, which are respectively extracted from the running data by the Long Short-Term Memory Network and the Transformer, are weighted and fused to obtain the running feature vector.

[0027] S2: Based on random forest, obtain the influence weights of environmental feature vector, image feature vector and running feature vector on the preset inefficient type. Based on the influence weights, fuse the environmental feature vector, image feature vector and running feature vector to obtain the fused feature vector.

[0028] In some preferred embodiments, in S2, the step of obtaining the influence weights of the environmental feature vector, image feature vector, and running feature vector on the preset inefficient type based on random forest includes: Obtain historical multidimensional feature vectors, and train the random forest to obtain the final random forest model based on the correlation between the historical multidimensional feature vectors and the preset inefficient type. Obtain the influence weights of the environmental feature vector, image feature vector, and running feature vector on the preset inefficient type based on the split gain of the historical environmental feature vector, historical image feature vector, and historical running feature vector in the final random forest model.

[0029] Specifically, the step of training the random forest to obtain the final random forest model based on the correlation between historical multidimensional feature vectors and preset inefficient types includes: In sequence, any one of the preset inefficient types is taken as the first category, and the remaining preset inefficient types are taken as the second category to construct an initial binary classification tree. A random forest is then constructed based on the initial binary classification tree. The final random forest model is obtained by training the random forest using the association between historical multidimensional feature vectors and preset inefficient types.

[0030] Understandably, the preset inefficiency types include minor power attenuation, severe power attenuation, appearance defects, minor electrical faults, and severe electrical faults. In this embodiment, historical multi-dimensional feature vectors are integrated to match the corresponding preset inefficiency type labels, forming the original multi-classification training dataset. For each preset inefficiency type... The labels are recoded into binary labels, with a label value of 1 indicating that the historical multidimensional feature vector belongs to a preset inefficient type. A label value of 0 indicates that the historical multidimensional feature vector belongs to the category excluding... All other preset inefficient types besides; based on the encoded binary labels, for each Generate a custom training dataset consisting of "historical multidimensional feature vectors + binary classification labels". (like Corresponding to "slight power attenuation vs. others" Corresponding to "Appearance Defects vs. Others"). Based on each Dedicated historical multidimensional feature vectors - binary classification label training dataset The final random forest model is obtained by training the initial binary classification tree in the random forest.

[0031] Specifically, the step of obtaining the influence weights of the environment feature vector, image feature vector, and running feature vector on the preset inefficient type based on the splitting gain of the historical environment feature vector, historical image feature vector, and historical running feature vector in the final random forest model includes: Obtain the sub-split gains of historical environment feature vectors, historical image feature vectors, and historical running feature vectors in the final binary classification tree of the final random forest model. Use the sum of the sub-split gains of historical environment feature vectors as the split gain of environment feature vectors in the final random forest model, the sum of the sub-split gains of historical image feature vectors as the split gain of image feature vectors in the final random forest model, and the sum of the sub-split gains of historical running feature vectors as the split gain of running feature vectors in the final random forest model. The splitting gain of the environmental feature vector, image feature vector, and running feature vector in the final random forest model is normalized to obtain the influence weights of the environmental feature vector, image feature vector, and running feature vector on the preset inefficient type.

[0032] Understandably, split gain can directly quantify the contribution of feature vectors to the improvement of class purity when splitting at a binary classification tree node. Therefore, by obtaining split gain, the influence weights of environmental feature vectors, image feature vectors, and running feature vectors on a preset inefficient type can be accurately obtained. In this embodiment, ; Represents historical feature vectors In the final binary classification tree t, the pre-defined inefficient type is... Sub-split gain, This represents the historical feature vector in the final binary classification tree t. The total number of nodes participating in the split. Represents historical feature vectors In the final binary classification tree t, node n pairs with the preset inefficient type. Splitting gain, , This indicates that the parent node (not split) in the final binary classification tree t represents the inefficient type. The Gini coefficient represents the degree of classification disorder of the parent node before splitting. The smaller the Gini coefficient, the purer the samples within the parent node. This represents the weighted Gini coefficient of all child nodes after a parent node splits in the final binary classification tree t. The weights are the proportion of each child node's sample size to the parent node's sample size, representing the average degree of classification confusion among these child nodes after the split. When obtaining the sub-split gain of the historical environment feature vector in the final binary classification tree t, the historical feature vector... The historical feature vector represents the sub-split gain of the historical image feature vector in the final binary classification tree t. The historical image feature vector is used to obtain the sub-split gain of the historical feature vector in the final binary classification tree t. This represents the historical operational feature vector. Understandably, it's necessary to match the environmental feature vector, image feature vector, and operational feature vector with the historical environmental feature vector, historical image feature vector, and historical operational feature vector from the historical multi-dimensional feature vector to obtain the influence weights of the environmental feature vector, image feature vector, and operational feature vector. If the historical feature vector... With feature vectors If they match, then the feature vectors Split gain in the final random forest model When obtaining the splitting gain of the environmental feature vector in the final random forest model, the feature vector This represents the environmental feature vector. When obtaining the splitting gain of the image feature vector in the final random forest model, the feature vector... The feature vector represents the image feature vector. When obtaining the split gain of the running feature vector in the final random forest model, the feature vector... denoted as the running feature vector, and T represents the total number of final binary classification trees.

[0033] In this embodiment, the influence weights of environmental feature vectors, image feature vectors, and running feature vectors on preset inefficient types can also be used as learnable parameters and embedded into the training process of a fully connected neural network. Through backpropagation, the influence weights can be dynamically adjusted to update them. When the validation set loss no longer decreases after 10 consecutive training rounds, or when the change in influence weights is less than a preset change, the iteration stops. In addition, a scene-aware factor can be introduced to dynamically fine-tune the influence weights, thereby ensuring that the weights are adaptively optimized in different scenarios, further improving the adaptability of the influence weights, and thus improving the recognition accuracy.

[0034] S3: Based on the fused feature vector and the preset inefficient type corresponding to the influencing weight in the fused feature vector, the classification model constructed using a fully connected neural network obtains the final inefficient type.

[0035] In one embodiment, step S3: obtaining the final inefficient type using a classification model constructed with a fully connected neural network based on the fused feature vector and the preset inefficient type corresponding to the influencing weights in the fused feature vector, includes: S31: Train the classification model based on the acquired historical multidimensional feature vectors to obtain the confidence of the first inefficient type. Based on the confidence of the first inefficient type and the actual inefficient type of the historical multidimensional feature vectors, correct the model parameters in the classification model to obtain the final classification model. S32: Based on the fused feature vector and the preset inefficient type corresponding to the influence weight in the fused feature vector, the final inefficient type is obtained using the final classification model.

[0036] In this embodiment, when the inefficient type corresponding to the first confidence level does not match the actual inefficient type of the historical multidimensional feature vector, the model parameters in the classification model are corrected based on the first confidence level. The acquired historical multidimensional feature vectors are divided into training, validation, and test sets in a 7:2:1 ratio. Cross-validation is used to adjust the model parameters (learning rate, number of hidden layer nodes, number of iterations, etc.) with accuracy, precision, recall, and F1 score as optimization objectives to improve the model's generalization ability and recognition performance.

[0037] Specifically, S32: Based on the fused feature vector and the preset inefficient types corresponding to the influencing weights in the fused feature vector, the final inefficient type is obtained using the final classification model, including: The fused feature vector is input into the final classification model to obtain the confidence level of the second inefficient type. If the confidence levels of the second inefficient type are all less than or equal to the first preset value, it indicates that the photovoltaic module is normal. If there is a second inefficient type with a confidence level greater than the first preset value, and the confidence levels of the second inefficient type are all less than or equal to the second preset value, an error is reported. If there is a second inefficient type with a confidence level greater than the second preset value, and the inefficient type corresponding to the largest second inefficient type confidence level matches the preset inefficient type corresponding to the influence weight in the fused feature vector, then the inefficient type corresponding to the largest second inefficient type confidence level is the final inefficient type.

[0038] In this embodiment, users can flexibly set the first and second preset values ​​according to actual needs. By combining the preset inefficiency types corresponding to the influence weights of the fused feature vectors with the output results of the classification model for mutual verification, the accuracy of photovoltaic module inefficiency type identification is greatly improved. When severe inefficiency characteristics (such as severe power degradation or severe electrical faults) are identified, an early warning mechanism is triggered, and early warning information is pushed to staff through the operation and maintenance platform to guide timely handling.

[0039] In this embodiment, three different types of photovoltaic power stations were selected as experimental subjects, namely: 1. Scenario A: Large-scale ground-mounted photovoltaic power station (a power station with an installed capacity of 100MW, using monocrystalline silicon modules, with an operating life of 5 years, and an environment characterized by strong winds and sandstorms and large temperature differences). 2. Scenario B: Distributed rooftop photovoltaic power station (an industrial park, with an installed capacity of 5MW, using polycrystalline silicon modules, and a service life of 2 years. The surrounding area is densely populated with buildings, resulting in localized shading). 3. Scenario C: Mountain photovoltaic power station (a mountainous area with an installed capacity of 20MW, thin-film modules, 3-year service life, complex terrain, and uneven distribution of sunlight and temperature). Data Acquisition and Preprocessing: 1. Data collection period: January 2024 - December 2024 (covering different seasons and weather conditions throughout the year); 2. Data acquisition equipment: Image data: A DJI M300 drone equipped with a high-definition visible light camera (4K resolution) and an infrared thermal imager (temperature measurement accuracy ±0.5℃) was used. Operating parameter data: collected through the power plant's SCADA system and component-level monitoring sensors; Environmental data: Deploy automatic weather stations (monitoring accuracy: light intensity ±5%, temperature ±0.2℃). 3. Data preprocessing: Image data: Median filtering removes image noise caused by wind and sand, histogram equalization enhances the recognition of features such as cracks and delamination, and the data is uniformly scaled to 256×256 pixels; Numerical data: The Raida criterion was used to remove about 2.5% of outlier data (such as instantaneous abnormal power under extreme weather conditions), the mean was used to fill about 1.8% of missing data, and after standardization, a regularized dataset (containing more than 100,000 samples) was formed. Model training and optimization: 1. Model parameter settings: CNN module: 3 convolutional layers (3×3 kernel size, number of kernels: 32, 64, 128 respectively), 2 pooling layers (max pooling, pooling kernel 2×2). LSTM module: 2 hidden layers, 128 nodes per layer; Transformer module: 4 encoder layers, 8 multi-head attention heads; Fully connected neural network: 2 hidden layers (256 and 128 nodes respectively), output layer with 6 categories (normal, mild power decay, severe power decay, appearance defect, mild electrical fault, severe electrical fault); Training parameters: learning rate 0.001, batch size 64, number of iterations 80, loss function is cross-entropy loss function. 2. Model optimization: The feature extraction performance of CNN+LSTM and CNN+Transformer was compared by cross-validation. Finally, the combination of CNN+Transformer was selected (which has higher accuracy in temporal feature extraction), and random forest weighted fusion weights were adopted. Operation and maintenance application effect: Scenario A: By identifying surface corrosion and power degradation of components caused by wind and sand, targeted cleaning and component replacement can be guided, resulting in an annual increase in power generation of approximately 7%. Scenario B: Accurately identify localized hot spots and poor contact at wiring terminals of roof components, address them promptly to prevent escalation of the fault, and reduce maintenance costs by 20%; Scenario C: Effectively identify regional power attenuation and thin-film module delamination defects caused by terrain shading, optimize module layout and operation and maintenance plans, and reduce power generation loss by 15%.

[0040] Example 2: This embodiment also provides a photovoltaic module inefficiency type identification system, applicable to the aforementioned photovoltaic module inefficiency type identification method, including: The explicit inefficient type acquisition module is used to extract features from environmental data, image data, and runtime data through a neural network to obtain environmental feature vectors, image feature vectors, and runtime feature vectors, respectively. It then determines whether there are feature vectors among these vectors that meet preset type determination conditions. If a feature vector that meets the preset type determination conditions is found, the final inefficient type is obtained through the feature vector that meets the preset type determination conditions. Otherwise, the feature vector fusion module is activated. The feature vector fusion module is used to obtain the influence weights of environmental feature vectors, image feature vectors, and running feature vectors on a preset inefficient type based on random forest, and then fuse the environmental feature vectors, image feature vectors, and running feature vectors based on the influence weights to obtain a fused feature vector. The implicit inefficient type acquisition module is used to obtain the final inefficient type based on the fused feature vector and the preset inefficient type corresponding to the influencing weight in the fused feature vector, using a classification model built with a fully connected neural network.

[0041] Example 3: This embodiment also provides a computer device, including: a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the computer device is running, the processor communicates with the memory via the bus, and the processor executes the machine-readable instructions to perform the steps of the photovoltaic module inefficiency type identification method.

[0042] The specific embodiments described above are preferred embodiments of the photovoltaic module inefficiency type identification method and system of the present invention, and are not intended to limit the specific scope of the present invention. The scope of the present invention includes but is not limited to the specific embodiments described above. All equivalent changes made in accordance with the shape and structure of the present invention are within the protection scope of the present invention.

Claims

1. A method for identifying the inefficiency type of photovoltaic modules, characterized in that, Includes the following steps: S1: Use a neural network to extract features from environmental data, image data, and runtime data to obtain environmental feature vectors, image feature vectors, and runtime feature vectors respectively. If there are feature vectors among the environmental feature vectors, image feature vectors, and runtime feature vectors that meet the preset type determination conditions, then obtain the final inefficient type through the feature vectors that meet the preset type determination conditions; otherwise, execute S2. S2: Based on random forest, obtain the influence weights of environmental feature vector, image feature vector and running feature vector on the preset inefficient type respectively, and fuse the environmental feature vector, image feature vector and running feature vector based on the influence weights to obtain the fused feature vector; S3: Based on the fused feature vector and the preset inefficient type corresponding to the influencing weight in the fused feature vector, the classification model constructed using a fully connected neural network obtains the final inefficient type.

2. The method for identifying the inefficiency type of photovoltaic modules according to claim 1, characterized in that, In S1, before obtaining the environmental feature vector, image feature vector, and operational feature vector by extracting features from the environmental data, image data, and operational data respectively through a neural network, the method further includes: Image data were preprocessed using median filtering and histogram equalization, respectively. Outliers in environmental and operational data were identified and removed using the Laida criterion, and missing values ​​in environmental and operational data were filled using mean imputation.

3. The method for identifying the inefficiency type of photovoltaic modules according to claim 1, characterized in that, In S1, the step of extracting environmental feature vectors, image feature vectors, and operational feature vectors from environmental data, image data, and operational data using a neural network includes: Image feature vectors are obtained by extracting features from image data using a convolutional neural network. Environmental feature vectors and operational feature vectors are obtained by extracting features from environmental data and operational data using Long Short-Term Memory Network and Transformer, respectively.

4. The method for identifying the inefficiency type of photovoltaic modules according to claim 1, characterized in that, In S2, the step of obtaining the influence weights of environmental feature vectors, image feature vectors, and running feature vectors on a preset inefficient type based on random forest includes: Obtain historical multidimensional feature vectors, and train the random forest to obtain the final random forest model based on the correlation between the historical multidimensional feature vectors and the preset inefficient type. Obtain the influence weights of the environmental feature vector, image feature vector, and running feature vector on the preset inefficient type based on the split gain of the historical environmental feature vector, historical image feature vector, and historical running feature vector in the final random forest model.

5. The method for identifying the inefficiency type of photovoltaic modules according to claim 4, characterized in that, The step of training the random forest to obtain the final random forest model based on the correlation between historical multidimensional feature vectors and preset inefficient types includes: In sequence, any one of the preset inefficient types is taken as the first category, and the remaining preset inefficient types are taken as the second category to construct an initial binary classification tree. A random forest is then constructed based on the initial binary classification tree. The final random forest model is obtained by training the random forest using the association between historical multidimensional feature vectors and preset inefficient types.

6. The method for identifying the inefficiency type of a photovoltaic module according to claim 4, characterized in that, The step of obtaining the influence weights of the environment feature vector, image feature vector, and running feature vector on the preset inefficient type based on the splitting gain of the historical environment feature vector, historical image feature vector, and historical running feature vector in the final random forest model includes: Obtain the sub-split gains of historical environment feature vectors, historical image feature vectors, and historical running feature vectors in the final binary classification tree of the final random forest model. Use the sum of the sub-split gains of historical environment feature vectors as the split gain of environment feature vectors in the final random forest model, the sum of the sub-split gains of historical image feature vectors as the split gain of image feature vectors in the final random forest model, and the sum of the sub-split gains of historical running feature vectors as the split gain of running feature vectors in the final random forest model. The splitting gain of the environmental feature vector, image feature vector, and running feature vector in the final random forest model is normalized to obtain the influence weights of the environmental feature vector, image feature vector, and running feature vector on the preset inefficient type.

7. The method for identifying the inefficiency type of photovoltaic modules according to claim 1, characterized in that, S3: Based on the fused feature vector and the preset inefficient types corresponding to the influencing weights in the fused feature vector, a classification model constructed using a fully connected neural network is used to obtain the final inefficient type, including: S31: Train the classification model based on the acquired historical multidimensional feature vectors to obtain the confidence of the first inefficient type. Based on the confidence of the first inefficient type and the actual inefficient type of the historical multidimensional feature vectors, correct the model parameters in the classification model to obtain the final classification model. S32: Based on the fused feature vector and the preset inefficient type corresponding to the influence weight in the fused feature vector, the final inefficient type is obtained using the final classification model.

8. The method for identifying the inefficiency type of photovoltaic modules according to claim 7, characterized in that, S32: Based on the fused feature vector and the preset inefficient types corresponding to the influencing weights in the fused feature vector, the final inefficient type is obtained using the final classification model, including: The fused feature vector is input into the final classification model to obtain the confidence level of the second inefficient type. If the confidence levels of the second inefficient type are all less than or equal to the first preset value, it indicates that the photovoltaic module is normal. If there is a second inefficient type with a confidence level greater than the first preset value, and the confidence levels of the second inefficient type are all less than or equal to the second preset value, an error is reported. If there is a second inefficient type with a confidence level greater than the second preset value, and the inefficient type corresponding to the largest second inefficient type confidence level matches the preset inefficient type corresponding to the influence weight in the fused feature vector, then the inefficient type corresponding to the largest second inefficient type confidence level is the final inefficient type.

9. A photovoltaic module inefficiency type identification system, applicable to the photovoltaic module inefficiency type identification method according to any one of claims 1-8, characterized in that, include: The explicit inefficient type acquisition module is used to extract features from environmental data, image data, and runtime data through a neural network to obtain environmental feature vectors, image feature vectors, and runtime feature vectors, respectively. It then determines whether there are feature vectors among these vectors that meet preset type determination conditions. If a feature vector that meets the preset type determination conditions is found, the final inefficient type is obtained through the feature vector that meets the preset type determination conditions. Otherwise, the feature vector fusion module is activated. The feature vector fusion module is used to obtain the influence weights of environmental feature vectors, image feature vectors, and running feature vectors on a preset inefficient type based on random forest, and then fuse the environmental feature vectors, image feature vectors, and running feature vectors based on the influence weights to obtain a fused feature vector. The implicit inefficient type acquisition module is used to obtain the final inefficient type based on the fused feature vector and the preset inefficient type corresponding to the influencing weight in the fused feature vector, using a classification model built with a fully connected neural network.

10. A computer device, comprising: The computer device includes a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the computer device is running, the processor communicates with the memory via the bus, and the processor executes the machine-readable instructions to perform the steps of a photovoltaic module inefficiency type identification method as described in any one of claims 1-8.

Images