Photovoltaic panel detection method based on combination of meteorological data and image processing

Abstract

The invention provides a photovoltaic panel detection method based on combination of meteorological data and image processing, and the method comprises the steps: carrying out the data collection and analysis, obtaining an RGB color image through a standard camera, capturing an infrared image emitted by a solar cell panel through an infrared thermal imaging device, and carrying out the detection of a photovoltaic panel. Predicting the condition of the photovoltaic cell panel in combination with meteorological data and the thermal image and the RGB image of each solar cell panel; data processing is carried out, unnecessary areas in the image are eliminated, a focus is selected as the center of the thermal image, the size of the image is adjusted, and normalization and missing value preprocessing are carried out on meteorological data; the method comprises the following steps: carrying out model description, detecting and classifying thermal leakage of a solar cell panel by using a thermal image and an RGB image, carrying out anomaly prediction according to weather data, loading a pre-training model specially used for classifying the thermal image, loading a pre-training model designed for classifying the RGB image, and carrying out anomaly prediction on the weather data by using the model. And finally combining prediction results generated by the three models.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic technology, and in particular to a photovoltaic panel detection method based on a combination of meteorological data and image processing. Background Technology

[0002] Traditional photovoltaic (PV) panel inspection methods are time-consuming, labor-intensive, and experience-dependent. While automated inspection equipment has seen some development, it remains costly and lacks flexibility. Key inspection technologies include image processing-based visual inspection, infrared thermal imaging, electroluminescence (EL) detection, and ultraviolet fluorescence (UVF) detection. Visual inspection identifies physical defects on the module surface, such as scratches and cracks, through high-resolution imaging. Infrared thermal imaging utilizes temperature distribution characteristics to quickly locate abnormal areas, making it suitable for online inspection and large-scale module inspection. EL detection, with its high sensitivity, effectively reveals microcracks, hidden cracks, and internal electrical defects, making it an important tool for both laboratory and field applications. UVF detection is used to assess the aging and light-induced degradation of PV module materials. Furthermore, existing research largely focuses on single data sources (such as images or meteorological data), lacking comprehensive analysis of multimodal data. Single data sources (such as thermal imaging, RGB images, or meteorological data) have limitations in detection range and accuracy, failing to effectively identify multiple defects in complex environments. Meanwhile, meteorological data is closely related to the performance of photovoltaic panels, but traditional analysis methods have failed to fully explore its potential, and there is even less research on cross-modal fusion. The differences in the characteristics of multimodal data and the requirements for computational efficiency make efficient detection by fusing images and meteorological data a pressing technical challenge that needs to be addressed. Summary of the Invention

[0003] To address the problems existing in the prior art, this invention proposes a photovoltaic panel detection method based on the combination of meteorological data and image processing.

[0004] To achieve the above objectives, the present invention adopts the following technical solution: A photovoltaic panel detection method based on a combination of meteorological data and image processing includes the following steps: S1. Conduct data collection and analysis, use a standard camera to acquire RGB color images, use an infrared thermal imaging device to capture infrared images emitted by solar panels, and combine meteorological data with thermal images and RGB images of each solar panel to predict the condition of the photovoltaic panels. S2. Perform data processing, eliminate any unwanted areas in the image, select the focal point as the center of the thermal image, adjust the image size, normalize the meteorological data, and preprocess missing values. S3. Perform model description, use thermal images and RGB images to detect and classify heat leakage of solar panels, and make anomaly predictions based on weather data. Load a pre-trained model specifically for classifying thermal images and load a pre-trained model designed for classifying RGB images. Use the model to make anomaly predictions based on weather data. Finally, combine the prediction results generated by the above three models.

[0005] Preferably, in step S1, the solar panels are classified into the following five categories using thermal images and RGB images: standard panels, dirty panels, broken panels, shaded panels, and cracked panels.

[0006] Preferably, in step S1, the meteorological data includes temperature, relative humidity, wind speed, wind direction, precipitation, and cleanliness. The data is generated by Xlsat. Using the meteorological data as input, it is cleaned, and then the correlation between different parameters is calculated. The values ​​on the diagonal represent the correlation between each variable and itself, which is always equal to 1, while the values ​​outside the diagonal represent the correlation between two different variables. Multiple linear regression is used to calculate the regression matrix of one output variable and multiple input variables. Specifically, the data is prepared into the input variable matrix by adding a bias column, and then the regression coefficients are calculated as follows: Where X is the input variable matrix with an additional bias column, and y is the output vector.

[0007] Preferably, the cleanliness parameter represents the reciprocal of the fouling ratio, which is determined by the ratio of the maximum current of the dirty column to the maximum current of the clean column.

[0008] In the formula SRi,PV The pollution ratio of solar panels; Imax_s This represents the daily average of the maximum current in the contaminated string; Imax_c This represents the daily average value of the maximum current in a clean string.

[0009] Preferably, in step S2, a defect detection method is used to locate and identify affected areas in the image. After detecting defects, the surrounding environment is analyzed to understand the patterns and structures within the image. A mask is applied to the image, and pixel-by-pixel operations are used to hide or delete unwanted objects. If both the image and the mask are grayscale, each pixel of the image is multiplied by the corresponding pixel in the mask. If the mask pixel is white, the pixel value in the image is retained. If the mask pixel is black, the pixel value in the image is replaced with the selected value. If both the image and the mask are color, operations need to be applied to each color channel separately.

[0010] Preferably, in step S2, data normalization scales the data to values ​​between 0 and 1, standardizing the data and facilitating comparisons between different variables; the method for handling missing values ​​is to delete rows or columns with missing values, and to input missing values ​​using the mean, median, or regression method.

[0011] Preferably, in step S3, the pre-trained model for classifying thermal images is a ResNet50 model, which consists of 50 convolutional layers, pooling operations, and fully connected layers; the classification of RGB images is done using an AlexNet model; and the anomaly prediction of weather data is done using a decision tree model.

[0012] Preferably, the three different models are classified into multiple classes using a weighted voting method. The predictions of the three models are combined by assigning weights based on the performance and reliability of the three models. The performance of each model is evaluated on the validation set, and weights are assigned to each model. The model that achieves better results is given higher weights. After assigning weights, the predictions of each model are weighed by multiplying each prediction by its corresponding weight. The final prediction for each data instance is obtained by combining the weighted predictions from each model.

[0013] Compared with the prior art, the present invention has the following beneficial effects: This invention proposes a multimodal fusion model that combines thermal imaging, RGB image processing, and meteorological data analysis, significantly improving the accuracy (up to 96%) and comprehensiveness of photovoltaic panel defect detection. Portable device and lightweight model design enhance the flexibility of the detection method, adapting to different scenarios and environmental conditions. Efficient image processing algorithms and support vector machine classification models achieve real-time performance and rapid response capabilities, suitable for large-scale on-site inspections. The integration of thermal imaging to identify temperature anomalies, RGB image processing to detect surface defects, and meteorological data to provide supplementary environmental information ensures intelligent and multi-dimensional detection. Furthermore, the detection results are presented in a user-friendly format, facilitating rapid problem location and maintenance, thus improving the operational efficiency and durability of the solar energy system. Attached Figure Description

[0014] Figure 1 A flowchart for multimodal anomaly detection of photovoltaic systems combining thermal and RGB images based on meteorological data; Figure 2 A flowchart illustrating the workflow for anomaly detection in thermal images of photovoltaic systems using ResNet50. Figure 3 This is a flowchart of the RGB image anomaly detection workflow for photovoltaic systems based on AlexNet. Detailed Implementation

[0015] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0016] This embodiment proposes a photovoltaic panel detection method based on the combination of meteorological data and image processing. It uses machine learning classification models (such as logistic regression and support vector machines) to model the pollution rate, and performs image processing on RGB and thermal data (such as AlexNet and ResNet) with different structures using ML and CNN. It also incorporates environmental data such as temperature, solar irradiance, wind speed, and wind direction to enhance anomaly detection. The method includes the following steps: S1. Conduct data collection and analysis, specifically including: S11. Acquire thermal and RGB images. Temperature significantly affects the efficiency of solar panels; as temperature increases, the efficiency of solar panels decreases. This phenomenon is attributed to the "temperature coefficient," which affects the performance of the photovoltaic cells that make up the solar panel. Photovoltaic cells use semiconductor materials (usually silicon) to convert sunlight into electrical energy. When the temperature rises, the conductivity of the semiconductor material also increases, resulting in greater energy loss as heat.

[0017] Thermal images depict the temperature distribution on the surface of a solar panel, acquired using a thermal imager that measures the infrared radiation emitted by the solar panel. In contrast, RGB images are color images acquired using a standard camera and presented as regular photographs. In this embodiment, an infrared thermal imaging technique is used with a FLIR E6xt camera to measure the temperature of the object. Unlike traditional cameras, thermal imagers can capture not only visible light images but also infrared images emitted by the object based on its temperature. The focus is primarily on five different categories: dirty, broken, shadowed, cracked, and a category of panels that are completely clean and without any issues.

[0018] Dirt buildup reduces the efficiency and lifespan of photovoltaic (PV) panels. Dirt on solar panels acts as an insulating layer, increasing panel temperature and reducing conversion efficiency. Maintaining a high level of cleanliness is crucial to preventing degradation and deterioration of solar equipment and optimizing PV system efficiency. To ensure optimal performance, the goal is to keep the contamination level above 0.7. A threshold of 0.7 is typically chosen as the minimum threshold to exceed before initiating cleaning, representing approximately 30% of performance loss. Exceeding this threshold risks panel degradation, leading to a significant drop in efficiency. Furthermore, there is a risk of dust buildup, complicating subsequent cleaning processes. Therefore, maintaining a high level of cleanliness is essential for ensuring optimal performance and facilitating PV system maintenance. For disconnection issues, solar panels may disconnect due to various reasons, including extreme weather conditions, installation problems, or electrical issues. Shading issues involve shading and affect solar panel efficiency. Several factors can cause shading, including physical obstacles such as trees, buildings, or surrounding walls, as well as environmental conditions such as clouds, snow, and dust debris. Additionally, installation problems, such as incorrect panel orientation, can exacerbate shading issues. Finally, solar panels are prone to cracking and breakage, which can severely impact their performance. Mechanical forces, such as wind, snow, hail, and physical impacts, pose a significant risk to the integrity of the panels. Furthermore, temperature fluctuations and manufacturing defects can cause stress, leading to structural damage. These cracks and breaks make the panels vulnerable to external factors such as moisture, dust, rain, snow, and ice, thus exacerbating their deterioration.

[0019] S12. Meteorological Data Collection: Weather data, air pollution, and external environmental factors, such as the presence of shadows cast by buildings or trees on solar panels, can have a combined impact. Combined with meteorological data from the NASAPOWER project database (Global Energy Resources Forecasting), which provides high-resolution spatiotemporal climate data to assess renewable energy potential and design sustainable energy systems, this project includes a meteorological database called the "POWER Data Access Viewer," allowing users to access and download weather data from around the world.

[0020] Dust buildup on photovoltaic (PV) panels can indeed negatively impact their performance and energy production. This is because dust reduces the amount of solar energy reaching the PV cells. The model will be able to predict the condition of PV panels by combining meteorological data with thermal and RGB images of each panel. The advantage of this combination is the ability to collect the largest possible amount of digital data, which is either open-source or easily accessible. This data includes temperature, relative humidity, wind speed, wind direction, precipitation, cleanliness, and more.

[0021] The data was generated by Xlsat, using meteorological data as input. This data was cleaned, and then the correlations between the different parameters were calculated. Values ​​on the diagonal represent the correlation between each variable and itself (always equal to 1), while values ​​outside the diagonal represent the correlation between two different variables. To calculate the regression matrix for one output variable and multiple input variables, multiple linear regression was used. The steps included preparing the data into the input variable matrix by adding a bias column, and then calculating the regression coefficients as follows: Where X is the input variable matrix with an additional bias column, and y is the output vector.

[0022] For example, the correlation between temperature and relative humidity is -0.9122, indicating a moderate negative correlation between the two variables. A strong correlation (0.912) can be observed between temperature and relative humidity, indicating a significant effect of temperature on relative humidity. Other interactions between variables cannot be captured by simple linear correlations, which is why it is necessary to use nonlinear models, such as the recently developed multiple linear regression (MLR) model with a correlation coefficient of 0.85. This phenomenon can be detected using normal RGB imaging or thermal imaging, which will improve the accuracy of dust detection and related phenomena such as dirt or snail tracks. For the "cleanliness" parameter, it represents the reciprocal of the dirt ratio (SRi), a parameter used to assess the degree of dirt or grime buildup on solar panels of photovoltaic devices. The dirt ratio is determined by the ratio of the maximum current of the dirty column to the maximum current of the clean column.

[0023] In the formula SRi,PV The pollution ratio of solar panels; Imax_s This represents the daily average of the maximum current in the contaminated string; Imax_c This represents the daily average value of the maximum current in a clean string.

[0024] Pollution rate (SRi, PV) is usually expressed as a percentage. A higher pollution rate indicates greater dirt or grime on the solar panels, leading to reduced power generation from the photovoltaic (PV) system. Therefore, maintaining proper cleanliness of solar panels to maximize power generation and ensure efficient operation of the PV system is crucial, primarily because of the strong correlation between pollution rate and meteorological data.

[0025] S2. Perform data processing, specifically including the following steps: S21. Eliminate any unwanted areas in the image and select the focal point as the center of the thermal image. This process involves analyzing image features to identify defects and their surroundings, applying image processing algorithms, and continuously filling the affected areas with visually realistic data.

[0026] First, defect detection techniques are used to locate and identify affected areas in the image. This can be achieved through color analysis, image segmentation, or CNN-based methods, providing accurate information about the location and size of defects. Once a defect is detected, the surrounding environment is analyzed to understand patterns and structures within the image. A mask is applied to the image, using pixel-by-pixel operations to hide or remove unwanted objects. If both the image and mask are grayscale, each pixel in the image is simply multiplied by the corresponding pixel in the mask. If the mask pixel is white, the pixel value in the image is preserved. If the mask pixel is black, the pixel value in the image is replaced with a selected value (e.g., a pixel value of 0 for a grayscale image). If both the image and mask are color (e.g., using an RGB model), operations need to be applied separately to each color channel (red, green, blue). Using the same method as for grayscale images, each color channel of the image is multiplied by the corresponding pixel of the mask. The result will be a new image where the unwanted objects specified by the mask are either removed or hidden.

[0027] S22. Resizing the image: There are several reasons to resize the image before training: First, popular models are already pre-trained on 224x224 pixels. Resizing the image to the same size is necessary for optimal use of these models. Secondly, resizing reduces image complexity. High-resolution images can be complex and require significant computational resources during model training, especially high-resolution RGB images. Thirdly, resizing helps normalize the size of the input data. By using the "Resize" function as a data preprocessing step, the image size can be adjusted to suit model requirements or other considerations. Besides 224x224 pixels, other sizes can be used depending on specific requirements.

[0028] S23. Normalize and preprocess meteorological data. Preprocessing meteorological data is a crucial step to ensure the accuracy and reliability of subsequent analysis or modeling results. In this embodiment, the preprocessing of CSV format meteorological data includes two main steps: normalization and missing value handling. These steps are essential for preparing meteorological data for tasks such as weather forecasting or climate trend analysis.

[0029] S231. Data Normalization: The purpose is to bring all variables to the same scale, preventing some variables from dominating others in magnitude. It ensures that all variables contribute to the analysis in a balanced way. Normalization includes scaling values ​​between 0 and 1. This step standardizes the data, facilitating comparisons between different variables.

[0030] S232. Handling Missing Values: Missing values ​​in meteorological data can pose a challenge during analysis. They may introduce bias or errors into the results. Several methods exist for handling missing values, such as deleting rows or columns with missing values, using techniques like mean, median, or regression to handle missing values, or utilizing more advanced methods based on ML models. In this embodiment, missing values ​​are handled by deleting affected rows or columns. However, ensuring the overall integrity of the dataset is crucial in this step.

[0031] S3. Describe the model, such as... Figure 1 As shown, the proposed model has a specific objective: to detect and classify heat leakage from solar panels using thermal and RGB images, while simultaneously making anomaly predictions based on weather data. It comprises four parts: The first part involves loading a pre-trained model specifically designed for classifying thermal images. To accomplish this task, the powerful ResNet architecture, which has proven effective in image classification, is utilized. The second part deals with loading a pre-trained model designed for classifying RGB images. It uses the AlexNet architecture, renowned for its superior performance in image classification tasks. The third part focuses on using weather data for anomaly prediction. In this case, the program reads relevant data from CSV or Excel files. Preprocessing steps are applied to the weather data, and predictions are made using a specialized model corresponding to this type of input. Part Four combines the predictions generated by the three models mentioned above. To achieve this, predefined weights are assigned to each model. Thermal images and RGB images are preprocessed and then predicted using their respective models. Similarly, weather data is preprocessed and predicted using the corresponding models. These probabilities are then weighted according to predefined weights, and the class with the highest weighted probability determines the final prediction. The ultimate goal of this program is to integrate the predictions from different models, including thermal images, RGB images, and weather data. By utilizing multiple information sources, this program improves the accuracy of anomaly prediction and provides a comprehensive assessment of potential anomalies present in images. Specifically: The first part focuses on ResNet50 for thermal imaging: ResNet50 belongs to the ResNet family of models (residual networks) and is designed to address the performance degradation that occurs as network depth increases. One of the key features of ResNet models (including ResNet50) is the use of residual blocks. These blocks introduce skip connections to bypass certain layers of the network, allowing information to propagate more easily through the network. This helps alleviate the vanishing or exploding gradient problem during training in deep networks. The ResNet50 workflow for anomaly detection in photovoltaic system thermal images is as follows: Figure 2 As shown.

[0032] The ResNet50 model consists of 50 convolutional layers, pooling operations, and fully connected layers. It utilizes 3x3 convolutional and pooling layers with a subsampling factor of 2 to reduce the spatial dimensionality of features. The ResNet50 architecture has been pre-trained on a large image dataset called ImageNet. ResNet50 can be fine-tuned for specific tasks by adding additional classification or regression layers, depending on the problem at hand.

[0033] This model utilizes libraries such as astensorflow and keras, popular libraries for building and training deep learning models. Next, training and testing data directories are defined. Data preprocessing is performed using the `imagedataggenera-tor` class from Keras. This class allows batch augmentation and preprocessing of images from the data directory. Images are resized, their value scales are normalized, and transformations such as rotation, scaling, and horizontal flipping are applied to increase the diversity of the training data. A pre-trained ResNet50 model is then loaded using the `ResNet50` function in Keras, and the weights are pre-trained on the ImageNet dataset. Fully connected layers are excluded, and the weights of each layer in the pre-trained model are frozen to avoid gradient backpropagation during training. The classification model is built by adding additional layers on top of the ResNet50 model. Flatten layers are used to flatten the output of the ResNet50 model, followed by two Dense (fully connected) layers with ReLU and softmax activation functions. The output layer has a softmax activation function to perform four-class classification. The model is then compiled by specifying the optimizer (Adam), loss function (sparse_categorical_crossentropy), and evaluation metric (accuracy). The model is trained using a fitting method with previously defined training and test data generators and a specified number of epochs. Finally, the trained model can be evaluated using a separate test dataset. The `predict_generator` method is used to obtain model predictions, and a confusion matrix is ​​calculated based on the predictions and the true labels. The confusion matrix can be visualized using the matplotlib library to evaluate the model's performance.

[0034] Part Two: AlexNet for RGB Images: AlexNet Model, as shown Figure 3 As shown, AlexNet is one of the most popular architectures in CNNs. It is characterized by the direct stacking of multiple convolutional layers, which enables the network to capture complex features and use them for accurate classification.

[0035] The first convolutional layer in AlexNet is an input layer that processes an input image of size 227x227x3. Here, "227x227" represents the spatial size of the image, and "3" represents the three color channels (red, green, and blue) in the RGB image. This first convolutional layer uses 96 filters of size 11x11x3. Each filter is a small weight matrix used to extract specific features from the input image. In this embodiment, the filter size is 11x11, and it is applied to all three color channels of the image. The result of this convolution produces 96 filter maps. A stride of 4 pixels means that the filter moves 4 pixels at a time during convolution. This means that the extracted features are sampled at 4-pixel intervals, reducing the spatial resolution of the filter map produced by this layer.

[0036] In AlexNet, the filters in the second, fourth, and fifth convolutional layers are connected only to the filter maps of the previous layer, which reside on the same graphics processing unit (GPU). This means that these layers effectively utilize parallelization to accelerate computation using multiple GPUs. The third convolutional layer in AlexNet is connected to all filter maps from the second convolutional layer. This means that the features extracted in this layer take into account information from all filter maps in the previous layer.

[0037] Finally, neurons in the fully connected layers of AlexNet are connected to all neurons in the previous layer. This allows for global communication between neurons, which is important for classification tasks. In summary, AlexNet is a CNN architecture that introduced innovative concepts to the field of computer vision. It uses stacked convolutional layers, filters of different sizes, and specific connections between layers to extract features and accurately classify RGB images.

[0038] The third part, a CSV dataset containing weather data, employs a decision tree algorithm, a nonparametric supervised learning algorithm for classification and regression tasks. It has a hierarchical and tree-like structure, consisting of a root node, branches, internal nodes, and leaf nodes.

[0039] The program imports necessary libraries such as pandas and matplotlib. Pyplot is used to process the data and generate visualizations. It then loads the data from an Excel file using the `pd.read_excel()` function. The data is stored in a DataFrame object named "data". Next, the program splits the data into independent variables "X" and dependent variables "y". The columns in "X" contain all the features or explanatory variables, except for the last column which contains the target label or class in "y". The program displays the first few rows of "X" and "Y" to provide an overview of the data. It also uses the `info()` method to display information about "X", showing the total number of entries, data type, and memory consumption. Furthermore, it uses the `isnull().sum()` method to check for any null values ​​in "X" and `sort_values(ascending=False)`. Then, the program performs feature scaling using "MinMaxScaler" from "sklearn.preprocessing". The "fit_transform()" method is used to scale the values ​​in each column of "X" to the range [0,1]. The scaled data is stored in a new data frame named "min_max_Scalar_df". In summary, the program loads the data, splits it into independent and dependent variables, performs feature scaling, and displays the data before and after scaling. This prepares the data for use in ML models or for further analysis.

[0040] Part Four: Weighted Voting: A weighted voting approach – a method for multi-class classification using three different models, aiming to combine the predictions of these models by assigning weights based on their performance and reliability. First, the performance of each model is evaluated on a validation set. This allows determining the weights assigned to each model. Models that achieve better results are given higher weights, reflecting their confidence and accuracy. Once weights are assigned, each model's prediction is weighed by multiplying each prediction by its corresponding weight. For example, if Model 1 has a weight of 0.4, Model 2 has a weight of 0.3, and Model 3 has a weight of 0.3, Model 1's prediction is adjusted by multiplying Model 1's prediction by 0.4, Model 2's by 0.3, and Model 3's by 0.3. By combining the weighted predictions from each model, a final prediction for each data instance is obtained. This weighted voting approach enables leveraging the individual strengths of each model while compensating for any potential weaknesses they may possess.

[0041] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A photovoltaic panel detection method based on a combination of meteorological data and image processing, characterized in that, Includes the following steps: S1. Conduct data collection and analysis, use a standard camera to acquire RGB color images, use an infrared thermal imaging device to capture infrared images emitted by solar panels, and combine meteorological data with thermal images and RGB images of each solar panel to predict the condition of the photovoltaic panels. S2. Perform data processing, eliminate unwanted areas in the image, select the focal point as the center of the thermal image, adjust the image size, normalize the meteorological data, and preprocess missing values. S3. Perform model description, use thermal images and RGB images to detect and classify heat leakage of solar panels, and make anomaly predictions based on weather data. Load a pre-trained model specifically for classifying thermal images and load a pre-trained model designed for classifying RGB images. Use the model to make anomaly predictions based on weather data. Finally, combine the prediction results generated by the above three models.

2. The photovoltaic panel detection method based on the combination of meteorological data and image processing according to claim 1, characterized in that, In step S1, the solar panels are classified into five categories based on thermal images and RGB images: standard panels, dirty panels, broken panels, shaded panels, and cracked panels.

3. The photovoltaic panel detection method based on the combination of meteorological data and image processing according to claim 1, characterized in that, In step S1, the meteorological data includes temperature, relative humidity, wind speed, wind direction, precipitation, and cleanliness. The data is generated by Xlsat. Using the meteorological data as input, it is cleaned, and then the correlation between different parameters is calculated. The values ​​on the diagonal represent the correlation between each variable and itself, and are always equal to 1. The values ​​outside the diagonal represent the correlation between two different variables. Multiple linear regression is used to calculate the regression matrix of one output variable and multiple input variables. Specifically, the data is prepared into the input variable matrix by adding a bias column, and then the regression coefficients are calculated as follows: Where X is the input variable matrix with an additional bias column, and y is the output vector.

4. The photovoltaic panel detection method based on the combination of meteorological data and image processing according to claim 3, characterized in that, The cleanliness parameter represents the reciprocal of the fouling ratio, which is determined by the ratio of the maximum current of the dirty column to the maximum current of the clean column. In the formula SRi,PV The pollution ratio of solar panels; Imax_s This represents the daily average of the maximum current in the contaminated string; Imax_c This represents the daily average value of the maximum current in a clean string.

5. The photovoltaic panel detection method based on the combination of meteorological data and image processing according to claim 1, characterized in that, In step S2, a defect detection method is used to locate and identify affected areas in the image. After detecting defects, the surrounding environment is analyzed to understand the patterns and structures within the image. A mask is applied to the image, and pixel-by-pixel operations are used to hide or delete unwanted objects. If both the image and the mask are grayscale, each pixel in the image is multiplied by the corresponding pixel in the mask. If the mask pixel is white, the pixel value in the image is retained. If the mask pixel is black, the pixel value in the image is replaced with the selected value. If both the image and the mask are color, operations need to be applied to each color channel separately.

6. The photovoltaic panel detection method based on the combination of meteorological data and image processing according to claim 1, characterized in that, In step S2, data normalization scales the data to values ​​between 0 and 1, standardizing the data and facilitating comparisons between different variables. Missing values ​​are handled by deleting rows or columns with missing values ​​and inputting missing values ​​using the mean, median, or regression.

7. The photovoltaic panel detection method based on the combination of meteorological data and image processing according to claim 1, characterized in that, In step S3, the pre-trained model for classifying thermal images uses the ResNet50 model, which consists of 50 convolutional layers, pooling operations, and fully connected layers; the classification of RGB images uses the AlexNet model; and the anomaly prediction of weather data uses the decision tree model.

8. The photovoltaic panel detection method based on the combination of meteorological data and image processing according to claim 7, characterized in that, Three different models are classified into multiple classes using a weighted voting method. The predictions of the three models are combined by assigning weights based on their performance and reliability. The performance of each model is evaluated on the validation set, and the weights assigned to each model are increased. The model that achieves better results is given higher weights. After the weights are assigned, the predictions of each model are weighed by multiplying each prediction by its corresponding weight. The final prediction for each data instance is obtained by combining the weighted predictions from each model.

Images