Photovoltaic module temperature monitoring method and system

Abstract

The present invention provides a photovoltaic module temperature monitoring method and system, which relates to the field of temperature monitoring technology, including obtaining standardized temperature data of photovoltaic modules through infrared thermal imaging and visible light image acquisition, combined with spatial registration algorithm and edge computing. A temperature prediction model is constructed using wavelet transform and graph convolutional neural network to realize the identification of abnormal temperature areas. Hierarchical clustering and graph neural network are used for fault analysis to generate warning information including fault location, warning level, fault type and processing guidance, and send it to the operation and maintenance management platform to realize intelligent temperature monitoring and fault warning of photovoltaic modules.

Description

Technical Field

[0001] The present invention relates to temperature monitoring technology, and in particular to a photovoltaic module temperature monitoring method and system. Background Art

[0002] Single-point temperature sensors are difficult to fully reflect the temperature distribution on the component surface, and the images collected by infrared thermal imagers often lack spatial correspondence with the actual components, making it difficult to accurately locate the temperature data to a specific location. This makes it difficult to accurately identify and locate abnormal areas.

[0003] Existing methods usually only focus on current temperature data and lack in-depth analysis and utilization of historical data. This limits the system's ability to grasp temperature trends and predict future temperature conditions, making it difficult to detect potential problems in a timely manner and take preventive measures.

[0004] Therefore, there is an urgent need for an improved photovoltaic module temperature monitoring method and system to achieve high-precision temperature monitoring and fault diagnosis. Summary of the invention

[0005] The embodiments of the present invention provide a photovoltaic module temperature monitoring method and system, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention,

[0007] A photovoltaic module temperature monitoring method is provided, comprising:

[0008] Acquire the temperature distribution image of the surface of the photovoltaic module through the infrared thermal imaging array, synchronously collect the visible light image of the surface of the photovoltaic module, register and map the temperature distribution image with the visible light image based on the spatial registration algorithm to obtain a registered image, use the edge computing unit to pre-process the registered image, extract the edge contour features of the photovoltaic module, generate the standardized temperature distribution data of the photovoltaic module, and collect the environmental monitoring data of the photovoltaic field at the same time, and store the standardized temperature distribution data and the environmental monitoring data in a distributed database to form a historical data set;

[0009] Performing wavelet transform processing on the standardized temperature distribution data in the historical data set to obtain temperature change characteristic data, constructing a graph convolutional neural network model based on the temperature change characteristic data to obtain heat conduction relationship parameters between photovoltaic modules, inputting the obtained temperature change characteristic data, heat conduction relationship parameters and environmental monitoring data into a deep neural network for training to obtain a temperature prediction model, inputting the real-time collected environmental monitoring data into the temperature prediction model to generate predicted temperature distribution data;

[0010] The standardized temperature distribution data and the predicted temperature distribution data are input into the hierarchical clustering model, the temperature anomaly area is identified and the temperature gradient value of the temperature anomaly area is calculated, the spatial position information of the temperature anomaly area and the temperature gradient value are input into the propagation analysis model, the propagation analysis model calculates the abnormal diffusion range data based on the graph neural network, extracts the characteristic parameters of the temperature anomaly area and compares them with the historical fault data, outputs the fault type determination result and the fault confidence value, generates the warning information including the fault location information, warning level information, fault type information and processing guidance information according to the temperature gradient value, the abnormal diffusion range data and the fault confidence value, and sends the warning information to the operation and maintenance management platform.

[0011] In an optional embodiment,

[0012] Preprocessing the registration image using an edge computing unit to extract edge contour features of the photovoltaic module and generate standardized temperature distribution data of the photovoltaic module includes:

[0013] Calculating similarity values ​​of image blocks in the registered image, determining a search window size based on the similarity values, constructing an adaptive non-local means filtering model according to the search window size, and inputting the registered image into the adaptive non-local means filtering model for denoising to obtain a denoised image;

[0014] A calibration plate is used to obtain internal parameters and external parameters of a camera, a radial distortion correction model is constructed according to the internal parameters, a tangential distortion correction model is constructed according to the external parameters, and the denoised image is sequentially input into the radial distortion correction model and the tangential distortion correction model for correction to obtain a corrected image;

[0015] Constructing a four-layer Gaussian pyramid for the corrected image to obtain first to fourth layer multi-scale images, respectively setting a dual-threshold adaptive edge detection operator on the first to fourth layer multi-scale images to extract edge feature points, performing morphological processing on the edge feature points using an opening operation and a closing operation to obtain optimized edge feature points, and reconstructing the contour of the optimized edge feature points using an ellipse fitting algorithm and a least squares curve fitting algorithm to obtain a boundary contour line of the photovoltaic module;

[0016] Dividing the photovoltaic module area into regular grid units based on the photovoltaic module boundary contour line, wherein the size of the regular grid unit is determined according to the spatial resolution of the temperature distribution image of the photovoltaic module surface acquired in advance;

[0017] A bilinear interpolation algorithm is used to calculate the measured temperature value of each regular grid unit, and the ambient temperature value and the standard operating temperature value of the photovoltaic module are obtained. The measured temperature value is subtracted from the ambient temperature value and then combined with the standard operating temperature value to calculate a temperature normalization value;

[0018] A time series temperature database is established, and the normalized temperature value is associated with the acquisition time information and the spatial location information for storage to generate a standardized temperature data set.

[0019] In an optional embodiment,

[0020] Constructing a four-layer Gaussian pyramid for the corrected image to obtain first to fourth layer multi-scale images, setting a dual threshold adaptive edge detection operator on the first to fourth layer multi-scale images to extract edge feature points, using open and close operations to perform morphological processing on the edge feature points to obtain optimized edge feature points, and using an ellipse fitting algorithm and a least squares curve fitting algorithm to reconstruct the contour of the optimized edge feature points to obtain a photovoltaic module boundary contour line, including:

[0021] Using a five-by-five Gaussian kernel to perform Gaussian filtering on the corrected image to obtain a filtered image, performing two-by-two regional downsampling on the filtered image to obtain a first layer image, repeatedly performing the Gaussian filtering and downsampling operations to sequentially obtain a second layer image, a third layer image, and a fourth layer image, and constructing a four-layer Gaussian pyramid structure;

[0022] Calculate the horizontal gradient components and vertical gradient components of the first to fourth layers of images respectively, calculate the gradient amplitude and gradient direction according to the horizontal gradient components and the vertical gradient components, perform non-maximum suppression on the gradient amplitude to obtain an edge intensity map, calculate the mean and standard deviation of the edge intensity map, set a high threshold and a low threshold according to the mean and standard deviation, and construct a dual-threshold adaptive edge detection operator;

[0023] Applying the dual threshold adaptive edge detection operator to the first to fourth layer images respectively, extracting strong edge points that meet the high threshold, and extracting weak edge points that meet the low threshold; judging whether there is a strong edge point in the eight neighborhoods of the weak edge point, and if so, merging the weak edge point with the strong edge point to obtain an edge feature point;

[0024] Using a three-by-three rectangular structure element to perform an opening operation on the edge feature point to obtain an initial edge point; using a five-by-five diamond structure element to perform a closing operation on the initial edge point to obtain an optimized edge point, wherein the optimized edge point includes position coordinates and direction information;

[0025] Inputting the optimized edge points into an ellipse fitting model, establishing an ellipse equation based on the position coordinates, and solving the ellipse equation parameters using a least squares method to obtain an initial contour line;

[0026] A sequence of contour points is uniformly sampled on the initial contour line, a cubic spline function is constructed with the interval between two adjacent contour points, the direction information is converted into a tangent constraint condition at the contour point, and the polynomial coefficients of the cubic spline function are solved by the least squares method to obtain the boundary contour line of the photovoltaic module.

[0027] In an optional embodiment,

[0028] The obtained temperature change characteristic data, heat conduction relationship parameters and environmental monitoring data are input into a deep neural network for training to obtain a temperature prediction model, and the real-time collected environmental monitoring data is input into the temperature prediction model to generate predicted temperature distribution data, including:

[0029] The temperature change characteristic data, heat conduction relationship parameters and environmental monitoring data are spliced ​​in the channel dimension to form a fusion feature vector;

[0030] Constructing a deep neural network model, the deep neural network model includes a feature extraction module, a time series modeling module and a temperature distribution reconstruction module, wherein the feature extraction module includes a plurality of residual blocks, each residual block includes a convolution layer and a batch normalization layer; the time series modeling module includes a forward long short-term memory network unit and a reverse long short-term memory network unit; the temperature distribution reconstruction module includes a multi-layer transposed convolution layer;

[0031] Input the fused feature vector into a feature extraction module, process it through a convolution layer and a batch normalization layer in a residual block to obtain a feature sequence, and perform a residual connection between the feature sequence and the fused feature vector to obtain a residual feature sequence;

[0032] Input the residual feature sequence into the time series modeling module, process it through the forward long short-term memory network unit and the reverse long short-term memory network unit to obtain forward features and reverse features respectively, and splice the forward features and the reverse features in the channel dimension to form a time series feature sequence;

[0033] Rearranging the time series feature sequence into a feature map, and upsampling the feature map through a transposed convolution layer in the temperature distribution reconstruction module to obtain a predicted temperature distribution;

[0034] Constructing a loss function, wherein the loss function includes a mean square error term between the predicted temperature distribution and the actual temperature distribution, and a spatial smoothness constraint term of the predicted temperature distribution, and using an adaptive optimization algorithm to train the deep neural network model to obtain a temperature prediction model;

[0035] The real-time collected environmental monitoring data is spliced ​​with the temperature change characteristic data and heat conduction relationship parameters of the previous moment to form a real-time fused feature vector, and the real-time fused feature vector is input into the temperature prediction model. The real-time residual feature sequence is obtained through the feature extraction module in turn, the real-time time series feature sequence is obtained through the time series modeling module, and the predicted temperature distribution data is generated through the temperature distribution reconstruction module.

[0036] In an optional embodiment,

[0037] The standardized temperature distribution data and the predicted temperature distribution data are input into the hierarchical clustering model, the temperature anomaly area is identified and the temperature gradient value of the temperature anomaly area is calculated, and the spatial position information of the temperature anomaly area and the temperature gradient value are input into the propagation analysis model. The propagation analysis model calculates the abnormal diffusion range data based on the graph neural network, including:

[0038] Constructing a hierarchical clustering model based on the minimum variance criterion, calculating the distance matrix between the standardized temperature distribution data and the predicted temperature distribution data, inputting the distance matrix into the hierarchical clustering model to construct a hierarchical clustering tree, determining the optimal number of clusters based on the ratio of the sum of intra-class distances to the sum of inter-class distances, and clustering the standardized temperature distribution data to obtain temperature anomaly areas;

[0039] The temperature change rates of the temperature abnormal area in the horizontal direction and the vertical direction are calculated by using a differential calculation method, the temperature change rates in the horizontal direction and the vertical direction are synthesized to obtain a temperature gradient vector, and the temperature gradient characteristics are calculated according to the temperature gradient vector, wherein the temperature gradient characteristics include a temperature gradient amplitude and a temperature gradient direction;

[0040] A graph structure is constructed based on the spatial distribution of the temperature anomaly region, the center position of the temperature anomaly region is set as a graph node, a connection relationship is established between corresponding nodes according to the boundary distances of adjacent temperature anomaly regions to form a graph edge, and the center position coordinates, region area, region perimeter, and region maximum inscribed circle radius of the temperature anomaly region are combined with the temperature gradient feature to form a node feature;

[0041] Construct a graph attention neural network, calculate the attention coefficient between nodes through the attention calculation unit, perform weighted combination of adjacent node features based on the attention coefficient, and obtain updated node representation through nonlinear transformation and residual connection;

[0042] The node representation output by the graph attention neural network is combined with the temperature gradient feature of the temperature anomaly area, and input into a pre-trained propagation analysis model for feature mapping to obtain the diffusion range data of the temperature anomaly area.

[0043] In an optional embodiment,

[0044] The node representation output by the graph attention neural network is combined with the temperature gradient feature of the temperature anomaly area, and input into a pre-trained propagation analysis model for feature mapping, so as to obtain the diffusion range data of the temperature anomaly area, including:

[0045] The node representation vector output by the graph attention neural network is nonlinearly fused with the temperature gradient feature of the temperature anomaly area to obtain a fused feature vector.

[0046] Constructing a propagation analysis model with direction perception capability, the propagation analysis model comprising multiple direction perception layers, each direction perception layer performing a tensor multiplication operation on the fused feature vector and the node unit direction vector, and obtaining a direction perception feature through a nonlinear activation function;

[0047] An adaptive weight calculation module is constructed based on the local temperature gradient, wherein the adaptive weight calculation module comprises multiple fully connected layers and a normalization layer, and the local temperature gradient is input into the adaptive weight calculation module to obtain an adaptive weight coefficient matrix;

[0048] The adaptive weight coefficient matrix is ​​weightedly combined with the direction perception feature, and the original features are fused through a residual connection structure to obtain a diffusion prediction feature of the temperature anomaly area;

[0049] The regional overlap loss is obtained by calculating the intersection-and-union ratio of the predicted diffusion area and the real area, and a multi-scale loss function is constructed by combining the boundary prediction loss of the predicted boundary and the real boundary and the gradient continuity loss of the predicted temperature field gradient and the actual temperature field gradient. The propagation analysis model is trained by the multi-scale loss function, and a random dropout mechanism is introduced in the hidden layer of the propagation analysis model to obtain a trained propagation analysis model;

[0050] The diffusion prediction features are input into the trained propagation analysis model to obtain diffusion range data and diffusion direction data that characterize the diffusion trend of the temperature anomaly area. The diffusion range data includes the spatial boundary coordinates of the abnormal area, and the diffusion direction data includes the main direction and secondary direction of heat propagation.

[0051] In an optional embodiment,

[0052] Extracting characteristic parameters of the temperature anomaly area and comparing them with historical fault data, outputting fault type determination results and fault confidence values, generating early warning information including fault location information, early warning level information, fault type information, and processing guidance information according to temperature gradient values, abnormal diffusion range data, and fault confidence values, and sending the early warning information to the operation and maintenance management platform includes:

[0053] Extract characteristic parameters of the temperature anomaly region, wherein the characteristic parameters include morphological characteristics, temperature distribution characteristics and dynamic evolution characteristics, wherein the morphological characteristics include the area ratio, perimeter ratio, roundness coefficient and irregularity of the abnormal region, wherein the temperature distribution characteristics include the highest temperature, average temperature, temperature standard deviation and temperature distribution skewness in the region, wherein the dynamic evolution characteristics include the temperature rise rate, abnormal diffusion rate and diffusion direction stability, and the morphological characteristics, temperature distribution characteristics and dynamic evolution characteristics are combined to construct a multidimensional feature vector;

[0054] Constructing a deep metric learning model based on a twin neural network, wherein the deep metric learning model includes multiple feature subspaces, each feature subspace corresponds to a different type of feature parameter, and a weight coefficient of each feature subspace is calculated through an attention mechanism;

[0055] Input the multidimensional feature vector into the deep metric learning model, calculate the similarity between the multidimensional feature vector and the standard feature vector in the historical fault database, weight the influence of the time decay factor on the historical data, and obtain the fault type determination result and the fault confidence value;

[0056] Inputting the temperature gradient value, the abnormal diffusion range data and the fault confidence value into a fuzzy inference system, calculating the warning level according to a preset fuzzy rule, and retrieving corresponding processing guidance information from a preset knowledge base based on the warning level;

[0057] The fault location information, the warning level, the fault type determination result and the processing guidance information are hierarchically organized to generate warning information, the fault location information includes the string number and physical coordinates, the processing guidance information includes step-by-step operation suggestions, and the warning information is sent to the operation and maintenance management platform through an encrypted channel. In the second aspect of the embodiment of the present invention,

[0058] A photovoltaic module temperature monitoring system is provided, comprising:

[0059] The first unit is used to obtain a temperature distribution image of the surface of a photovoltaic module through an infrared thermal imaging array, synchronously collect a visible light image of the surface of the photovoltaic module, register and map the temperature distribution image with the visible light image based on a spatial registration algorithm to obtain a registered image, pre-process the registered image using an edge computing unit, extract edge contour features of the photovoltaic module, generate standardized temperature distribution data of the photovoltaic module, and simultaneously collect environmental monitoring data of the photovoltaic field area, and store the standardized temperature distribution data and the environmental monitoring data in a distributed database to form a historical data set;

[0060] The second unit is used to perform wavelet transform processing on the standardized temperature distribution data in the historical data set to obtain temperature change characteristic data, build a graph convolutional neural network model based on the temperature change characteristic data to obtain heat conduction relationship parameters between photovoltaic modules, input the obtained temperature change characteristic data, heat conduction relationship parameters and environmental monitoring data into a deep neural network for training to obtain a temperature prediction model, input the real-time collected environmental monitoring data into the temperature prediction model, and generate predicted temperature distribution data;

[0061] The third unit is used to input the standardized temperature distribution data and the predicted temperature distribution data into the hierarchical clustering model, identify the temperature anomaly area and calculate the temperature gradient value of the temperature anomaly area, input the spatial position information of the temperature anomaly area and the temperature gradient value into the propagation analysis model, the propagation analysis model calculates the abnormal diffusion range data based on the graph neural network, extracts the characteristic parameters of the temperature anomaly area and compares them with the historical fault data, outputs the fault type determination result and the fault confidence value, generates early warning information including fault location information, early warning level information, fault type information, and processing guidance information according to the temperature gradient value, the abnormal diffusion range data, and the fault confidence value, and sends the early warning information to the operation and maintenance management platform.

[0062] According to a third aspect of the embodiments of the present invention,

[0063] An electronic device is provided, comprising:

[0064] processor;

[0065] a memory for storing processor-executable instructions;

[0066] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0067] A fourth aspect of the embodiments of the present invention is:

[0068] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0069] In this embodiment, by registering and mapping the infrared thermal imaging array and the visible light image, combined with the preprocessing of the edge computing unit, the edge contour features of the photovoltaic module can be accurately extracted to generate standardized temperature data. This method significantly improves the accuracy of temperature monitoring, and at the same time uses a distributed database to store historical data, thereby improving the efficiency of data processing and analysis. The temperature distribution data is processed by wavelet transform, and a temperature prediction model that can consider the heat conduction relationship between photovoltaic modules is constructed by combining graph convolutional neural network and deep neural network. This method can more accurately predict the temperature distribution of photovoltaic modules and provide a reliable basis for timely detection of potential problems. The temperature abnormality area is identified by a hierarchical clustering model, and the abnormal diffusion range is analyzed by using a graph neural network. Combined with historical fault data, this system can intelligently determine the fault type and give a confidence level. The warning information finally generated contains comprehensive fault information and processing guidance, which greatly improves the operation and maintenance efficiency and reliability of photovoltaic modules. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 Schematic diagram of the process of the photovoltaic module temperature monitoring method according to an embodiment of the present invention;

[0071] Figure 2 Schematic diagram of the structure of a photovoltaic module temperature monitoring system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0072] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0073] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0074] Figure 1 FIG. 1 is a flow chart of a photovoltaic module temperature monitoring method according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0075] S101. Acquire the temperature distribution image of the surface of the photovoltaic module through the infrared thermal imaging array, and synchronously collect the visible light image of the surface of the photovoltaic module. Based on the spatial registration algorithm, the temperature distribution image and the visible light image are registered and mapped to obtain a registered image. The registered image is preprocessed by the edge computing unit to extract the edge contour features of the photovoltaic module and generate the standardized temperature distribution data of the photovoltaic module. At the same time, the environmental monitoring data of the photovoltaic field is collected, and the standardized temperature distribution data and the environmental monitoring data are stored in a distributed database to form a historical data set.

[0076] First, the temperature distribution image of the photovoltaic module surface is obtained through the infrared thermal imaging array. The infrared thermal imaging array can use an uncooled infrared detector with a resolution of 640×480, an operating wavelength of 8-14μm, and a temperature resolution of 0.05°C. The visible light image of the photovoltaic module surface is collected simultaneously, and the visible light camera uses a CMOS sensor with a resolution of 1920×1080.

[0077] Then, based on the spatial registration algorithm, the temperature distribution image and the visible light image are registered and mapped to obtain a registered image. Specifically, a registration method based on feature points can be used. First, SIFT feature points are extracted from the two images, and then the RANSAC algorithm is used to match the feature points. Finally, the affine transformation matrix is ​​calculated to achieve image registration.

[0078] Next, the edge computing unit is used to preprocess the registered image. In the photovoltaic temperature monitoring scenario, the edge computing unit will be deployed directly near the photovoltaic module to process the image data obtained from the infrared thermal imaging device and the visible light camera. For example, by performing image registration, filtering and edge detection by the edge computing unit, the temperature abnormality area and the edge features of the photovoltaic module can be quickly identified, and standardized temperature data can be generated. This real-time computing capability reduces the delay of data transmission and can issue early warning signals in time when abnormal situations occur. The image is denoised by Gaussian filtering, and then the edge detection operator is used for edge detection to extract the edge contour features of the photovoltaic module. Based on the extracted edge features, the position of the photovoltaic module in the image is located, and the temperature data is standardized to the range of 0-1 to generate standardized temperature data.

[0079] At the same time, environmental monitoring data of the photovoltaic field is collected, including ambient temperature, humidity, wind speed, light intensity, etc. The standardized temperature distribution data and environmental monitoring data are stored in a distributed database (such as HBase) to form a historical data set.

[0080] In an optional embodiment,

[0081] Preprocessing the registration image using an edge computing unit to extract edge contour features of the photovoltaic module and generate standardized temperature distribution data of the photovoltaic module includes:

[0082] Calculating similarity values ​​of image blocks in the registered image, determining a search window size based on the similarity values, constructing an adaptive non-local means filtering model according to the search window size, and inputting the registered image into the adaptive non-local means filtering model for denoising to obtain a denoised image;

[0083] A calibration plate is used to obtain internal parameters and external parameters of a camera, a radial distortion correction model is constructed according to the internal parameters, a tangential distortion correction model is constructed according to the external parameters, and the denoised image is sequentially input into the radial distortion correction model and the tangential distortion correction model for correction to obtain a corrected image;

[0084] Constructing a four-layer Gaussian pyramid for the corrected image to obtain first to fourth layer multi-scale images, respectively setting a dual-threshold adaptive edge detection operator on the first to fourth layer multi-scale images to extract edge feature points, performing morphological processing on the edge feature points using an opening operation and a closing operation to obtain optimized edge feature points, and reconstructing the contour of the optimized edge feature points using an ellipse fitting algorithm and a least squares curve fitting algorithm to obtain a boundary contour line of the photovoltaic module;

[0085] Dividing the photovoltaic module area into regular grid units based on the photovoltaic module boundary contour line, wherein the size of the regular grid unit is determined according to the spatial resolution of the temperature distribution image of the photovoltaic module surface acquired in advance;

[0086] A bilinear interpolation algorithm is used to calculate the measured temperature value of each regular grid unit, and the ambient temperature value and the standard operating temperature value of the photovoltaic module are obtained. The measured temperature value is subtracted from the ambient temperature value and then combined with the standard operating temperature value to calculate a temperature normalization value;

[0087] A time series temperature database is established, and the normalized temperature value is associated with the acquisition time information and the spatial location information for storage to generate a standardized temperature data set.

[0088] Exemplarily, first, calculate the similarity value of the image blocks in the registered image. Specifically, the structural similarity index (SSIM) can be used to measure the similarity between image blocks. The registered image is divided into multiple image blocks of equal size, such as image blocks of 8x8 pixels. For each image block, calculate its SSIM value with the surrounding image blocks. The closer the SSIM value is to 1, the more similar the two image blocks are.

[0089] Then, the search window size is determined based on the calculated similarity value. When the similarity value is high, it means that the local area of ​​the image is highly similar, and a smaller search window can be used; otherwise, a larger search window is used. For example, when the average SSIM value is greater than 0.9, the search window size is set to 7x7; when the average SSIM value is between 0.7-0.9, the search window size is set to 11x11; when the average SSIM value is less than 0.7, the search window size is set to 15x15.

[0090] Next, an adaptive non-local mean filtering model is constructed based on the determined search window size. This model is based on the non-local mean filtering algorithm, but the search window size is adaptively adjusted according to the local features of the image. For each pixel point, similar pixel blocks are found in its search window, similarity weights are calculated, and then these weights are used to perform weighted averaging of the pixel values ​​in the window to obtain the filtered pixel value.

[0091] The registered image is input into the constructed adaptive non-local mean filtering model for denoising to obtain a denoised image. This step can effectively remove the noise in the image while retaining the edge and texture information of the image.

[0092] Next, use the calibration board to obtain the internal and external parameters of the camera. The internal parameters include focal length, principal point coordinates, pixel size, etc.; the external parameters include the position and attitude information of the camera. You can use a checkerboard calibration board, take multiple calibration board images at different positions, and then use the calibration method to calculate the camera parameters.

[0093] A radial distortion correction model is constructed based on the internal parameters obtained. Radial distortion is mainly caused by the imperfection of the lens, which causes the image to produce barrel or pincushion-shaped deformation. The radial distortion correction model usually uses a polynomial function to describe the degree of distortion and corrects the pixel coordinates according to the degree of distortion.

[0094] At the same time, a tangential distortion correction model is constructed based on external parameters. Tangential distortion is mainly caused by the non-parallelism between the lens and the image plane, which causes the image to tilt and deform. The tangential distortion correction model usually uses a bilinear function to describe the degree of distortion and corrects the pixel coordinates according to the degree of distortion.

[0095] The denoised image is input into the radial distortion correction model and the tangential distortion correction model for correction to obtain a corrected image. This step can eliminate the geometric deformation introduced by the camera lens in the image and make the image closer to the real scene.

[0096] A four-layer Gaussian pyramid is constructed for the corrected image to obtain multi-scale images from the first to the fourth layers. The Gaussian pyramid is obtained by downsampling and Gaussian filtering layer by layer, and the resolution of each layer of the image is 1 / 4 of the previous layer. This allows image features to be analyzed at different scales, which is conducive to extracting objects of different sizes.

[0097] A dual-threshold adaptive edge detection operator is set on the first to fourth layers of multi-scale images to extract edge feature points. The dual-threshold adaptive edge detection operator can adaptively adjust the threshold according to the local features of the image to improve the accuracy of edge detection. Specifically, the Canny operator can be used, in which the high and low thresholds are dynamically determined according to the image gradient histogram.

[0098] The extracted edge feature points are morphologically processed using opening and closing operations to obtain optimized edge feature points. The opening operation can remove small noise points, and the closing operation can fill the edge gaps. This step can make the edge smoother and more continuous.

[0099] The ellipse fitting algorithm and the least squares curve fitting algorithm are used to reconstruct the contour of the optimized edge feature points to obtain the boundary contour of the photovoltaic module. The ellipse fitting algorithm is suitable for fitting the overall contour of the photovoltaic module, and the least squares curve fitting algorithm is suitable for fitting local details.

[0100] Based on the obtained PV module boundary contour, the PV module area is divided into regular grid cells. The size of the regular grid cells is determined according to the spatial resolution of the temperature distribution image of the PV module surface acquired in advance. For example, if the spatial resolution of the temperature distribution image is 0.1m / pixel, the PV module area can be divided into 0.1m x 0.1m grid cells.

[0101] The measured temperature value of each regular grid cell is calculated using the bilinear interpolation algorithm. The bilinear interpolation algorithm takes into account the temperature values ​​of the four vertices of the grid cell and can obtain a smoother temperature distribution.

[0102] Get the ambient temperature and standard operating temperature of the PV module. The ambient temperature can be measured by a temperature sensor installed near the PV module. The standard operating temperature is usually 25°C.

[0103] The temperature normalization value is calculated by subtracting the ambient temperature value from the measured temperature value and combining it with the standard working temperature value. The specific calculation formula is: temperature normalization value = (measured temperature - ambient temperature) / (standard working temperature - ambient temperature). This can eliminate the impact of ambient temperature changes on the measurement results, making the measurement results at different times and in different environments comparable.

[0104] Finally, a time series temperature database is established to associate and store the normalized temperature value with the acquisition time information and spatial location information to generate a standardized temperature data set. The time series temperature database can be implemented using a relational database (such as MySQL) or a time series database (such as InfluxDB). Each record in the database contains information such as the normalized temperature value, acquisition timestamp, and grid unit coordinates.

[0105] In this embodiment, image noise and geometric deformation are effectively removed by adaptive non-local mean filtering and geometric distortion correction, thereby improving the accuracy of subsequent processing. Multiscale analysis and morphological processing further enhance the extraction effect of edge features. Temperature normalization is used to eliminate the influence of ambient temperature changes, making the temperature data collected under different conditions comparable. Regular grid division and bilinear interpolation ensure the spatial consistency of temperature data. A time series temperature database is established to associate temperature data with spatiotemporal information for storage, which is convenient for subsequent big data analysis and mining, and provides a data basis for photovoltaic module performance evaluation and fault diagnosis.

[0106] In an optional embodiment,

[0107] Constructing a four-layer Gaussian pyramid for the corrected image to obtain first to fourth layer multi-scale images, setting a dual threshold adaptive edge detection operator on the first to fourth layer multi-scale images to extract edge feature points, using open and close operations to perform morphological processing on the edge feature points to obtain optimized edge feature points, and using an ellipse fitting algorithm and a least squares curve fitting algorithm to reconstruct the contour of the optimized edge feature points to obtain a photovoltaic module boundary contour line, including:

[0108] Using a five-by-five Gaussian kernel to perform Gaussian filtering on the corrected image to obtain a filtered image, performing two-by-two regional downsampling on the filtered image to obtain a first layer image, repeatedly performing the Gaussian filtering and downsampling operations to sequentially obtain a second layer image, a third layer image, and a fourth layer image, and constructing a four-layer Gaussian pyramid structure;

[0109] Calculate the horizontal gradient components and vertical gradient components of the first to fourth layers of images respectively, calculate the gradient amplitude and gradient direction according to the horizontal gradient components and the vertical gradient components, perform non-maximum suppression on the gradient amplitude to obtain an edge intensity map, calculate the mean and standard deviation of the edge intensity map, set a high threshold and a low threshold according to the mean and standard deviation, and construct a dual-threshold adaptive edge detection operator;

[0110] Applying the dual threshold adaptive edge detection operator to the first to fourth layer images respectively, extracting strong edge points that meet the high threshold, and extracting weak edge points that meet the low threshold; judging whether there is a strong edge point in the eight neighborhoods of the weak edge point, and if so, merging the weak edge point with the strong edge point to obtain an edge feature point;

[0111] Using a three-by-three rectangular structure element to perform an opening operation on the edge feature point to obtain an initial edge point; using a five-by-five diamond structure element to perform a closing operation on the initial edge point to obtain an optimized edge point, wherein the optimized edge point includes position coordinates and direction information;

[0112] Inputting the optimized edge points into an ellipse fitting model, establishing an ellipse equation based on the position coordinates, and solving the ellipse equation parameters using a least squares method to obtain an initial contour line;

[0113] A sequence of contour points is uniformly sampled on the initial contour line, a cubic spline function is constructed with the interval between two adjacent contour points, the direction information is converted into a tangent constraint condition at the contour point, and the polynomial coefficients of the cubic spline function are solved by the least squares method to obtain the boundary contour line of the photovoltaic module.

[0114] Exemplarily, a four-layer Gaussian pyramid is first constructed for the corrected image to obtain multi-scale images from the first layer to the fourth layer. Specifically, a 5x5 Gaussian kernel is used to perform Gaussian filtering on the corrected image to obtain a filtered image, and then the filtered image is downsampled by 2x2 regions to obtain a first layer image. The Gaussian filtering and downsampling operations are repeated to obtain the second layer, the third layer, and the fourth layer images in turn, thereby constructing a four-layer Gaussian pyramid structure.

[0115] Next, a dual-threshold adaptive edge detection operator is set on each layer of multi-scale images to extract edge feature points. In specific implementation, the horizontal gradient component and vertical gradient component of each layer of the image are first calculated respectively. Taking the first layer of image as an example, the Sobel operator can be used to calculate the gradients in the horizontal and vertical directions. Then the gradient amplitude and gradient direction are calculated based on the horizontal and vertical gradient components. The gradient amplitude is non-maximum suppressed to obtain the edge intensity map, and the mean and standard deviation of the edge intensity map are calculated. The high threshold and low threshold are set according to the mean and standard deviation. For example, the high threshold can be set to the mean plus 1.5 times the standard deviation, and the low threshold can be set to 0.4 times the high threshold. In this way, a dual-threshold adaptive edge detection operator is constructed.

[0116] The dual threshold adaptive edge detection operator is applied to each layer of the image to extract strong edge points that meet the high threshold and weak edge points that meet the low threshold. For weak edge points, it is determined whether there are strong edge points in its 8-neighborhood. If there are, the weak edge points are merged with the strong edge points to obtain edge feature points.

[0117] In order to further optimize the edge feature points, a morphological processing method is used. First, the edge feature points are opened using a 3x3 rectangular structure element to obtain the initial edge points. The opening operation can remove small protrusions and noise. Then, the initial edge points are closed using a 5x5 diamond structure element to obtain the optimized edge points. The closing operation can fill small holes and gaps. The optimized edge points contain position coordinates and direction information.

[0118] Next, the ellipse fitting algorithm is used to reconstruct the contour of the optimized edge points. The optimized edge points are input into the ellipse fitting model, and the ellipse equation is established based on the position coordinates. The least squares method is used to solve the ellipse equation parameters to obtain the initial contour line. For example, the algebraic distance minimization method can be used to solve the ellipse parameters.

[0119] Finally, the least squares curve fitting algorithm is used for fine contour reconstruction. The contour point sequence is uniformly sampled on the initial contour line, and the cubic spline function is constructed with the interval between two adjacent contour points. The direction information of the edge point is converted into the tangent constraint condition at the contour point. The polynomial coefficients of the cubic spline function are solved by the least squares method to obtain the final boundary contour line of the photovoltaic module.

[0120] In this embodiment, a multi-scale image processing strategy is adopted, which can effectively process image features of different scales and improve the accuracy and robustness of edge detection. The adaptive dual-threshold edge detection algorithm can automatically adjust the threshold according to the image features to adapt to images under different lighting conditions. Combined with multiple algorithms such as morphological processing, ellipse fitting and curve fitting, the optimization of edge features and fine contour reconstruction are achieved, and the accuracy and smoothness of boundary extraction are improved.

[0121] S102. Perform wavelet transform processing on the standardized temperature distribution data in the historical data set to obtain temperature change characteristic data, construct a graph convolutional neural network model based on the temperature change characteristic data to obtain thermal conduction relationship parameters between photovoltaic components, input the obtained temperature change characteristic data, thermal conduction relationship parameters and environmental monitoring data into a deep neural network for training to obtain a temperature prediction model, input the real-time collected environmental monitoring data into the temperature prediction model to generate predicted temperature distribution data.

[0122] The standardized temperature distribution data in the historical data set is processed by wavelet transform to obtain temperature change characteristic data. Specifically, the db4 wavelet basis function can be used to perform three-layer wavelet decomposition and extract low-frequency coefficients as temperature change characteristics. A graph convolutional neural network model is constructed based on the temperature change characteristic data to obtain the thermal conduction relationship parameters between photovoltaic modules. The graph convolutional network adopts a three-layer GraphConv structure, with a hidden layer dimension of 64 and an activation function of ReLU.

[0123] The obtained temperature change characteristic data, heat conduction relationship parameters and environmental monitoring data are input into the deep neural network for training to obtain the temperature prediction model. The deep neural network adopts a 4-layer fully connected structure, with hidden layer dimensions of 256, 128, and 64 respectively, the activation function is ReLU, the loss function is the mean square error, and the optimizer is Adam. The training data set contains 6 months of historical data, and the training set and the validation set are divided into 8:2 ratio. The environmental monitoring data collected in real time are input into the temperature prediction model to generate predicted temperature distribution data. The mean absolute error (MAE) of the prediction result is less than 1°C. The standardized temperature distribution data and the predicted temperature distribution data are input into the hierarchical clustering model to identify the temperature anomaly area. The hierarchical clustering model adopts the Ward method, sets the number of clusters to 3, and clusters the temperature data into three categories: normal, slightly abnormal, and severely abnormal.

[0124] In an optional embodiment,

[0125] The obtained temperature change characteristic data, heat conduction relationship parameters and environmental monitoring data are input into a deep neural network for training to obtain a temperature prediction model, and the real-time collected environmental monitoring data is input into the temperature prediction model to generate predicted temperature distribution data, including:

[0126] The temperature change characteristic data, heat conduction relationship parameters and environmental monitoring data are spliced ​​in the channel dimension to form a fusion feature vector;

[0127] Constructing a deep neural network model, the deep neural network model includes a feature extraction module, a time series modeling module and a temperature distribution reconstruction module, wherein the feature extraction module includes a plurality of residual blocks, each residual block includes a convolution layer and a batch normalization layer; the time series modeling module includes a forward long short-term memory network unit and a reverse long short-term memory network unit; the temperature distribution reconstruction module includes a multi-layer transposed convolution layer;

[0128] Input the fused feature vector into a feature extraction module, process it through a convolution layer and a batch normalization layer in a residual block to obtain a feature sequence, and perform a residual connection between the feature sequence and the fused feature vector to obtain a residual feature sequence;

[0129] Input the residual feature sequence into the time series modeling module, process it through the forward long short-term memory network unit and the reverse long short-term memory network unit to obtain forward features and reverse features respectively, and splice the forward features and the reverse features in the channel dimension to form a time series feature sequence;

[0130] Rearranging the time series feature sequence into a feature map, and upsampling the feature map through a transposed convolution layer in the temperature distribution reconstruction module to obtain a predicted temperature distribution;

[0131] Constructing a loss function, wherein the loss function includes a mean square error term between the predicted temperature distribution and the actual temperature distribution, and a spatial smoothness constraint term of the predicted temperature distribution, and using an adaptive optimization algorithm to train the deep neural network model to obtain a temperature prediction model;

[0132] The real-time collected environmental monitoring data is spliced ​​with the temperature change characteristic data and heat conduction relationship parameters of the previous moment to form a real-time fused feature vector, and the real-time fused feature vector is input into the temperature prediction model. The real-time residual feature sequence is obtained through the feature extraction module in turn, the real-time time series feature sequence is obtained through the time series modeling module, and the predicted temperature distribution data is generated through the temperature distribution reconstruction module.

[0133] For example, firstly, temperature change characteristic data, heat conduction relationship parameters and environmental monitoring data are obtained, and these data are spliced ​​in the channel dimension to form a fused feature vector. The fused feature vector contains the historical information of temperature change, the heat conduction characteristics between objects and the current environmental conditions, providing comprehensive input information for subsequent temperature prediction.

[0134] Next, we build a deep neural network model, which consists of a feature extraction module, a time series modeling module, and a temperature distribution reconstruction module. The feature extraction module contains multiple residual blocks, each of which consists of a convolutional layer and a batch normalization layer. The convolutional layer is used to extract spatial features, and the batch normalization layer helps to accelerate network training and improve the generalization ability of the model. The design of residual connections can alleviate the gradient vanishing problem of deep networks and is conducive to training deeper network structures.

[0135] The time series modeling module adopts a bidirectional long short-term memory network (LSTM) structure, which includes forward LSTM units and reverse LSTM units. LSTM networks are good at capturing long-term dependencies and can effectively model the change of temperature over time. The bidirectional structure can use past and future information at the same time to further improve the accuracy of time series modeling.

[0136] The temperature distribution reconstruction module consists of multiple layers of transposed convolutional layers. The transposed convolutional layers can gradually upsample the low-dimensional features to the original resolution and reconstruct the complete temperature distribution map. The multi-layer design makes the reconstruction process smoother and helps to generate more detailed temperature distribution results.

[0137] The specific processing flow is as follows: First, the fused feature vector is input into the feature extraction module. Taking 5 residual blocks as an example, each residual block contains two 3x3 convolutional layers with 64 convolution kernels. The batch normalization layer is located after each convolutional layer. The output of the residual block is elementwise added to the input to form a residual connection. After being processed by 5 residual blocks, a feature sequence is obtained. The feature sequence is residually connected with the original fused feature vector to obtain a residual feature sequence.

[0138] The residual feature sequence is then input into the time series modeling module. Both the forward LSTM unit and the reverse LSTM unit use 128 hidden units. The forward LSTM processes the sequence from left to right, and the reverse LSTM processes the sequence from right to left. The outputs of the two directions are concatenated in the channel dimension to obtain a 256-channel time series feature sequence.

[0139] Next, the time series feature sequence is rearranged into a feature map and input into the temperature distribution reconstruction module. This module contains 3 layers of transposed convolutional layers, with kernel size of 4x4 and stride of 2. The number of output channels of the three layers are 128, 64 and 1 respectively. The output of the last layer is the predicted temperature distribution map.

[0140] During the training phase, a loss function is constructed to guide model learning. The loss function consists of two parts: the mean square error term between the predicted temperature distribution and the actual temperature distribution, and the spatial smoothness constraint term of the predicted temperature distribution. The mean square error term measures the difference between the predicted value and the true value, and the spatial smoothness constraint term encourages the temperature prediction values ​​at adjacent locations to be similar, which helps to generate a smoother and more natural temperature distribution. The Adam optimizer is used to train the model, the learning rate is initially set to 0.001, and the learning rate decay strategy is used.

[0141] After training is completed, the model can be used for real-time temperature prediction. The real-time collected environmental monitoring data is spliced ​​with the temperature change feature data and heat conduction relationship parameters of the previous moment to form a real-time fusion feature vector. The real-time fusion feature vector is input into the trained model, and then passes through the feature extraction module, time series modeling module and temperature distribution reconstruction module in turn to finally generate the predicted temperature distribution data.

[0142] Taking an industrial production environment as an example, suppose you need to predict the temperature distribution of a 10mx10m area. The environmental monitoring data includes the current ambient temperature, humidity and air pressure; the temperature change characteristic data includes the temperature distribution map for each hour in the past 24 hours; the heat conduction relationship parameters include the thermal conductivity coefficients of different materials in the area. These data are integrated into a 100x100x30 fusion feature vector (spatial resolution is 100x100, number of feature channels is 30). After model processing, a 100x100x1 temperature distribution prediction map is output, indicating the temperature distribution in the next hour.

[0143] In this embodiment, by fusing multi-source data and adopting a deep learning model, the method can make full use of information such as historical temperature changes, heat conduction characteristics, and environmental conditions to achieve high-precision temperature distribution prediction. Using a residual network structure and a bidirectional LSTM network, the method can effectively extract spatial-temporal features and capture the long-term dependencies of temperature changes. By introducing spatial smoothing constraints, the temperature distribution generated by this method is smoother and more natural, avoiding unreasonable jumps in the prediction results.

[0144] S103. Input the standardized temperature distribution data and the predicted temperature distribution data into the hierarchical clustering model, identify the temperature anomaly area and calculate the temperature gradient value of the temperature anomaly area, input the spatial position information of the temperature anomaly area and the temperature gradient value into the propagation analysis model, the propagation analysis model calculates the abnormal diffusion range data based on the graph neural network, extracts the characteristic parameters of the temperature anomaly area and compares them with the historical fault data, outputs the fault type determination result and the fault confidence value, generates the warning information including the fault location information, warning level information, fault type information and processing guidance information according to the temperature gradient value, abnormal diffusion range data and fault confidence value, and sends the warning information to the operation and maintenance management platform.

[0145] The temperature gradient value of the temperature anomaly area is calculated. The temperature gradient uses the Sobel operator to calculate the gradient in the x and y directions, and then takes the gradient amplitude as the temperature gradient value.

[0146] The characteristic parameters of the temperature abnormal area include the shape characteristics (area, perimeter, roundness, etc.), temperature distribution characteristics (maximum temperature, average temperature, temperature gradient, etc.) and time series characteristics (temperature change rate, diffusion rate, etc.) of the abnormal area. These characteristic parameters are compared and matched with the standard characteristics in the historical fault database, and the similarity score is calculated to output the most likely fault type judgment result and its corresponding confidence value.

[0147] Generate early warning information based on temperature gradient value, abnormal diffusion range data, and fault confidence value. The early warning level is divided into four levels: normal, slight abnormality, moderate abnormality, and severe abnormality. Fault types include hot spots, cold solder joints, cracks, etc. Processing guidance information includes recommended inspection items, processing methods, etc.

[0148] In an optional embodiment,

[0149] The standardized temperature distribution data and the predicted temperature distribution data are input into the hierarchical clustering model, the temperature anomaly area is identified and the temperature gradient value of the temperature anomaly area is calculated, and the spatial position information of the temperature anomaly area and the temperature gradient value are input into the propagation analysis model. The propagation analysis model calculates the abnormal diffusion range data based on the graph neural network, including:

[0150] Constructing a hierarchical clustering model based on the minimum variance criterion, calculating the distance matrix between the standardized temperature distribution data and the predicted temperature distribution data, inputting the distance matrix into the hierarchical clustering model to construct a hierarchical clustering tree, determining the optimal number of clusters based on the ratio of the sum of intra-class distances to the sum of inter-class distances, and clustering the standardized temperature distribution data to obtain temperature anomaly areas;

[0151] The temperature change rates of the temperature abnormal area in the horizontal direction and the vertical direction are calculated by using a differential calculation method, the temperature change rates in the horizontal direction and the vertical direction are synthesized to obtain a temperature gradient vector, and the temperature gradient characteristics are calculated according to the temperature gradient vector, wherein the temperature gradient characteristics include a temperature gradient amplitude and a temperature gradient direction;

[0152] A graph structure is constructed based on the spatial distribution of the temperature anomaly region, the center position of the temperature anomaly region is set as a graph node, a connection relationship is established between corresponding nodes according to the boundary distances of adjacent temperature anomaly regions to form a graph edge, and the center position coordinates, region area, region perimeter, and region maximum inscribed circle radius of the temperature anomaly region are combined with the temperature gradient feature to form a node feature;

[0153] Construct a graph attention neural network, calculate the attention coefficient between nodes through the attention calculation unit, perform weighted combination of adjacent node features based on the attention coefficient, and obtain updated node representation through nonlinear transformation and residual connection;

[0154] The node representation output by the graph attention neural network is combined with the temperature gradient feature of the temperature anomaly area, and input into a pre-trained propagation analysis model for feature mapping to obtain the diffusion range data of the temperature anomaly area.

[0155] For example, the standardized temperature distribution data and the predicted temperature distribution data are first input into the hierarchical clustering model to identify the temperature anomaly area and calculate the temperature gradient value of the temperature anomaly area. Then the spatial location information of the temperature anomaly area and the temperature gradient value are input into the propagation analysis model, and the abnormal diffusion range data is calculated based on the graph neural network. The specific implementation steps are as follows:

[0156] First, a hierarchical clustering model based on the minimum variance criterion was constructed. The Euclidean distance between the standardized temperature distribution data and the predicted temperature distribution data was calculated to obtain a distance matrix. The distance matrix was input into the hierarchical clustering model, and the hierarchical clustering tree was constructed using the Ward method. The optimal number of clusters was determined by calculating the ratio of the sum of the intra-class distance to the sum of the inter-class distance. The standardized temperature distribution data was clustered with the optimal number of clusters as a parameter to obtain the temperature anomaly area.

[0157] For example, assume that the standardized temperature distribution data is a 10×10 matrix, and the predicted temperature distribution data is also a 10×10 matrix. Calculate the Euclidean distance between the corresponding elements of the two matrices to obtain a 100×100 distance matrix. Input the distance matrix into the hierarchical clustering model to construct a hierarchical clustering tree. By calculating the ratio of the sum of the intra-class distance to the sum of the inter-class distance under different cluster numbers, the optimal number of clusters is determined to be 3. Finally, cluster the standardized temperature distribution data to obtain 3 temperature anomaly areas.

[0158] Next, the temperature change rate in the horizontal and vertical directions of the temperature anomaly area is calculated using the differential calculation method. Specifically, for each point in the temperature anomaly area, the temperature difference between it and the adjacent point in the horizontal and vertical directions is calculated, and divided by the corresponding spatial distance to obtain the temperature change rate in the horizontal and vertical directions. The temperature change rates in the horizontal and vertical directions are synthesized to obtain the temperature gradient vector. The temperature gradient characteristics are calculated based on the temperature gradient vector, including the temperature gradient amplitude and the temperature gradient direction.

[0159] Taking the above 10×10 matrix as an example, assume that the first temperature anomaly area contains 16 points. The temperature change rates in the horizontal and vertical directions are calculated for these 16 points to obtain 16 temperature gradient vectors. The average amplitude and average direction of these vectors are calculated as the temperature gradient characteristics of the temperature anomaly area.

[0160] Then, a graph structure is constructed based on the spatial distribution of the temperature anomaly area. The center position of the temperature anomaly area is set as a graph node. A connection relationship is established between corresponding nodes according to the boundary distance of adjacent temperature anomaly areas to form a graph edge. The center position coordinates, area, perimeter, and maximum inscribed circle radius of the temperature anomaly area are combined with the temperature gradient features to form a node feature.

[0161] For example, for the three temperature anomaly areas mentioned above, their center position coordinates, area, perimeter, and maximum inscribed circle radius are calculated respectively. These features are combined with the previously calculated temperature gradient features to obtain the feature vectors of the three nodes. Based on the boundary distances between the three areas, the connection relationship between the nodes is established.

[0162] Next, a graph attention neural network is constructed. The attention coefficients between nodes are calculated by the attention calculation unit. Specifically, for each pair of connected nodes, the attention coefficients are calculated based on their feature vectors. The features of adjacent nodes are weighted combined based on the attention coefficients, and the updated node representation is obtained through nonlinear transformation and residual connection.

[0163] In this example, the attention coefficients between the three nodes and their adjacent nodes are calculated respectively. Then, based on these coefficients, the node features are weighted combined and nonlinearly transformed to obtain the updated node representation.

[0164] Finally, the node representation output by the graph attention neural network is combined with the temperature gradient feature of the temperature anomaly area and input into the pre-trained propagation analysis model for feature mapping to obtain the diffusion range data of the temperature anomaly area. The propagation analysis model can adopt a multi-layer perceptron structure to map the input features to the diffusion range data through multiple nonlinear transformations.

[0165] For each temperature anomaly area, its updated node representation is spliced ​​with the temperature gradient feature and input into the propagation analysis model. The model outputs the predicted value of the diffusion range of the area at different time steps, such as the diffusion area after 1 hour, 3 hours and 6 hours.

[0166] In this embodiment, the temperature anomaly area is identified by a hierarchical clustering model, and combined with the temperature gradient characteristics, the abnormal pattern of temperature distribution can be accurately captured. The graph attention neural network is used to model the relationship between the temperature anomaly areas, making full use of the spatial distribution information of the area. This modeling method based on the graph structure can effectively capture the mutual influence between abnormal areas and improve the accuracy of abnormal diffusion analysis. Combining the graph neural network with the propagation analysis model realizes end-to-end analysis from anomaly detection to diffusion prediction. This integrated method can not only identify temperature anomaly areas, but also predict their future diffusion trends, providing an important basis for taking preventive measures in a timely manner.

[0167] In an optional embodiment,

[0168] The node representation output by the graph attention neural network is combined with the temperature gradient feature of the temperature anomaly area, and input into a pre-trained propagation analysis model for feature mapping, so as to obtain the diffusion range data of the temperature anomaly area, including:

[0169] The node representation vector output by the graph attention neural network is nonlinearly fused with the temperature gradient feature of the temperature anomaly area to obtain a fused feature vector.

[0170] Constructing a propagation analysis model with direction perception capability, the propagation analysis model comprising multiple direction perception layers, each direction perception layer performing a tensor multiplication operation on the fused feature vector and the node unit direction vector, and obtaining a direction perception feature through a nonlinear activation function;

[0171] An adaptive weight calculation module is constructed based on the local temperature gradient, wherein the adaptive weight calculation module comprises multiple fully connected layers and a normalization layer, and the local temperature gradient is input into the adaptive weight calculation module to obtain an adaptive weight coefficient matrix;

[0172] The adaptive weight coefficient matrix is ​​weightedly combined with the direction perception feature, and the original features are fused through a residual connection structure to obtain a diffusion prediction feature of the temperature anomaly area;

[0173] The regional overlap loss is obtained by calculating the intersection-and-union ratio of the predicted diffusion area and the real area, and a multi-scale loss function is constructed by combining the boundary prediction loss of the predicted boundary and the real boundary and the gradient continuity loss of the predicted temperature field gradient and the actual temperature field gradient. The propagation analysis model is trained by the multi-scale loss function, and a random dropout mechanism is introduced in the hidden layer of the propagation analysis model to obtain a trained propagation analysis model;

[0174] The diffusion prediction features are input into the trained propagation analysis model to obtain diffusion range data and diffusion direction data that characterize the diffusion trend of the temperature anomaly area. The diffusion range data includes the spatial boundary coordinates of the abnormal area, and the diffusion direction data includes the main direction and secondary direction of heat propagation.

[0175] Exemplarily, first obtain the node representation output by the graph attention neural network and the temperature gradient characteristics of the temperature anomaly area. The node representation vector and the temperature gradient feature are nonlinearly fused, which can be implemented by a multi-layer perceptron. The input layer connects the two features, and several fully connected layers and ReLU activation functions are set in the middle. The last layer outputs the fused feature vector. For example, if the node representation dimension is 128 and the temperature gradient feature dimension is 64, the input layer dimension is 192, and two hidden layers can be set in the middle, with dimensions of 256 and 128 respectively, and finally a 128-dimensional fused feature vector is output.

[0176] Next, we build a propagation analysis model with direction perception. The model contains multiple direction perception layers, each of which performs tensor multiplication on the fused feature vector and the node's unit direction vector. In specific implementation, the fused feature vector can be adjusted to the shape of the batch size (i.e., the number of samples), a single dimension, and a feature dimension, and the node unit direction vector can be adjusted to the shape of the batch size, the number of directions, and the three-dimensional coordinates, and then batch matrix multiplication is performed to obtain a tensor containing the batch size, the number of directions, and the feature dimension. The direction perception feature is then obtained through a nonlinear activation function (such as the ReLU activation function). Multiple such direction perception layers can be stacked, and the output of each layer is used as the input of the next layer.

[0177] Then, an adaptive weight calculation module is constructed based on the local temperature gradient. This module contains multiple layers of fully connected layers and normalization layers, and takes the local temperature gradient as input. For example, a three-layer fully connected network can be designed, with the dimensions of the hidden layers set to 64 and 32 respectively, and the dimension of the output layer is the same as the number of directions. Each layer is followed by batch normalization and ReLU activation. The output of the last layer is normalized by Softmax to obtain the adaptive weight coefficient matrix.

[0178] The adaptive weight coefficient matrix is ​​weightedly combined with the direction-aware features. The specific operation is to adjust the weight coefficient matrix to the shape of batch size, number of directions and single dimension, multiply it element by element with the direction-aware features (batch size, number of directions and feature dimension), and then sum it in the direction number dimension to obtain the weighted features containing batch size and feature dimension. Then, through the residual connection structure, the weighted features are added to the original fusion features to generate the diffusion prediction features of the temperature anomaly area.

[0179] Next, a multi-scale loss function is constructed to train the propagation analysis model. First, the intersection over union (IoU) of the predicted diffusion area and the true area is calculated to obtain the regional overlap loss. IoU (Intersection over Union) can be used as an evaluation indicator. Secondly, the boundary prediction loss between the predicted boundary and the true boundary is calculated, and the smooth L1 loss can be used. Then the gradient continuity loss between the predicted temperature field gradient and the actual temperature field gradient is calculated, and the mean square error loss can be used. The weighted sum of these three losses is used to obtain the multi-scale loss function. During the training process, a random deactivation mechanism (Dropout) is introduced into the hidden layer of the propagation analysis model to prevent overfitting. The model parameters are trained by backpropagation and gradient descent optimization algorithms.

[0180] Finally, the diffusion prediction features are input into the trained propagation analysis model to obtain the diffusion range data and diffusion direction data that characterize the diffusion trend of the temperature anomaly area. The diffusion range data contains the spatial boundary coordinates of the abnormal area and can be expressed as a series of three-dimensional coordinate points. The diffusion direction data contains the main direction and secondary direction of heat propagation, which can be expressed by unit vectors. For example, the main direction is [0.8, 0.5, 0.3], and the secondary direction is [-0.4, 0.7, 0.6].

[0181] In this embodiment, the local and global features of the temperature anomaly area are effectively extracted through the graph attention network and nonlinear feature fusion, which enhances the model's ability to express complex heat conduction processes. The introduction of the direction perception layer and the adaptive weight mechanism enables the model to have directional heat conduction analysis capabilities and can accurately predict the diffusion trend of heat in different directions. The use of multi-scale loss functions and random inactivation mechanisms improves the generalization performance of the model, avoids overfitting problems, and makes the prediction results more stable and reliable.

[0182] In an optional embodiment,

[0183] Extracting characteristic parameters of the temperature anomaly area and comparing them with historical fault data, outputting fault type determination results and fault confidence values, generating early warning information including fault location information, early warning level information, fault type information, and processing guidance information according to temperature gradient values, abnormal diffusion range data, and fault confidence values, and sending the early warning information to the operation and maintenance management platform includes:

[0184] Extract characteristic parameters of the temperature anomaly region, wherein the characteristic parameters include morphological characteristics, temperature distribution characteristics and dynamic evolution characteristics, wherein the morphological characteristics include the area ratio, perimeter ratio, roundness coefficient and irregularity of the abnormal region, wherein the temperature distribution characteristics include the highest temperature, average temperature, temperature standard deviation and temperature distribution skewness in the region, wherein the dynamic evolution characteristics include the temperature rise rate, abnormal diffusion rate and diffusion direction stability, and the morphological characteristics, temperature distribution characteristics and dynamic evolution characteristics are combined to construct a multidimensional feature vector;

[0185] Constructing a deep metric learning model based on a twin neural network, wherein the deep metric learning model includes multiple feature subspaces, each feature subspace corresponds to a different type of feature parameter, and a weight coefficient of each feature subspace is calculated through an attention mechanism;

[0186] Input the multidimensional feature vector into the deep metric learning model, calculate the similarity between the multidimensional feature vector and the standard feature vector in the historical fault database, weight the influence of the time decay factor on the historical data, and obtain the fault type determination result and the fault confidence value;

[0187] Inputting the temperature gradient value, the abnormal diffusion range data and the fault confidence value into a fuzzy inference system, calculating the warning level according to a preset fuzzy rule, and retrieving corresponding processing guidance information from a preset knowledge base based on the warning level;

[0188] The fault location information, the warning level, the fault type determination result and the processing guidance information are hierarchically organized to generate warning information, the fault location information includes the string number and physical coordinates, the processing guidance information includes step-by-step operation suggestions, and the warning information is sent to the operation and maintenance management platform through an encrypted channel.

[0189] For example, first, a thermal imaging camera is used to collect the surface temperature image of the photovoltaic power station component. The collected temperature image is preprocessed, including noise removal, contrast enhancement and other operations to improve the image quality. Then, an image segmentation algorithm is used to extract the temperature abnormal area and separate the temperature abnormal area from the background.

[0190] Next, the characteristic parameters of the temperature anomaly area are extracted. The characteristic parameters include morphological characteristics, temperature distribution characteristics and dynamic evolution characteristics.

[0191] Morphological features include:

[0192] 1) Abnormal area ratio: Calculate the ratio of the abnormal area to the entire component area.

[0193] 2) Abnormal area perimeter ratio: Calculate the ratio of the abnormal area perimeter to the entire component perimeter.

[0194] 3) Roundness coefficient: Calculates the circularity of the abnormal area. The closer the roundness coefficient is to 1, the closer the shape is to a circle.

[0195] 4) Irregularity: describes the complexity of the shape of the abnormal area, which is obtained by calculating the ratio of the perimeter to the area of ​​the area.

[0196] Temperature distribution characteristics include:

[0197] 1) Maximum temperature in the area: the highest temperature value in the abnormal area.

[0198] 2) Average temperature: the average temperature of all pixels in the abnormal area.

[0199] 3) Temperature standard deviation: describes the degree of dispersion of temperature distribution.

[0200] 4) Temperature distribution skewness: describes the asymmetry of temperature distribution.

[0201] Dynamic evolution features include:

[0202] 1) Temperature rise rate: the increase in the average temperature of the abnormal area per unit time.

[0203] 2) Abnormal diffusion rate: the rate at which the abnormal area expands per unit time.

[0204] 3) Diffusion direction stability: describes the stability of the diffusion direction of the abnormal area.

[0205] The above feature parameters are combined to construct a multi-dimensional feature vector as the input of fault diagnosis.

[0206] Then, a deep metric learning model based on the twin neural network is constructed. The model contains multiple feature subspaces, each of which corresponds to a type of feature parameter. For example, the morphological feature subspace, the temperature distribution feature subspace, and the dynamic evolution feature subspace. In each subspace, a convolutional neural network is used to extract deep features. The weight coefficients of each feature subspace are calculated through the attention mechanism to achieve adaptive fusion of features.

[0207] The multidimensional feature vector is input into the deep metric learning model to calculate the similarity with the standard feature vector in the historical fault database. The similarity calculation uses cosine similarity, with a value range of 0-1. The time decay factor is introduced to give a smaller weight to the older historical data. For example, for data from 30 days ago, the weight coefficient is set to 0.8; for data from 60 days ago, the weight coefficient is set to 0.6. The final similarity score is obtained by weighted average.

[0208] The fault type determination result is determined based on the similarity score. For example, when the similarity score is greater than 0.9, it is determined to be this type of fault; when the similarity score is between 0.7-0.9, it is determined to be a suspected fault of this type. At the same time, the similarity score is used as the fault confidence value.

[0209] Next, the temperature gradient value, the abnormal diffusion range data and the fault confidence value are input into the fuzzy inference system. The fuzzy inference system calculates the warning level based on the preset fuzzy rules. The fuzzy rule example is as follows:

[0210] If the temperature gradient is large, the anomaly diffusion range is large, and the fault confidence is high, the warning level is high; if the temperature gradient is medium, the anomaly diffusion range is medium, and the fault confidence is medium, the warning level is medium; if the temperature gradient is small, the anomaly diffusion range is small, and the fault confidence is low, the warning level is low.

[0211] According to the calculated warning level, the corresponding processing guidance information is retrieved from the preset knowledge base. The processing guidance information includes step-by-step operation suggestions, such as: immediately shut down the affected string inverter, arrange professional personnel to conduct on-site inspection, prepare spare components, and replace them when necessary.

[0212] Finally, the fault location information, warning level, fault type determination results and processing guidance information are organized in layers to generate warning information. The fault location information includes the string number (such as String_01) and physical coordinates (such as X: 120m, Y: 80m). The generated warning information is sent to the operation and maintenance management platform through an SSL encrypted channel to ensure the security of data transmission.

[0213] In this embodiment, by extracting multi-dimensional feature parameters and combining them with a deep metric learning model, the characteristics of the temperature abnormality area can be fully characterized, and the accuracy of fault type determination can be improved. The warning level is automatically calculated based on the fuzzy reasoning system, and corresponding processing guidance suggestions are given to reduce the subjectivity of human judgment and improve the timeliness and effectiveness of fault handling. The modular design can easily expand new feature parameters and fault types to meet the needs of different types of photovoltaic power stations, and has good versatility and scalability.

[0214] Figure 2FIG. 1 is a schematic diagram of the structure of a photovoltaic module temperature monitoring system according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0215] The first unit is used to obtain a temperature distribution image of the surface of a photovoltaic module through an infrared thermal imaging array, synchronously collect a visible light image of the surface of the photovoltaic module, register and map the temperature distribution image with the visible light image based on a spatial registration algorithm to obtain a registered image, pre-process the registered image using an edge computing unit, extract edge contour features of the photovoltaic module, generate standardized temperature distribution data of the photovoltaic module, and simultaneously collect environmental monitoring data of the photovoltaic field area, and store the standardized temperature distribution data and the environmental monitoring data in a distributed database to form a historical data set;

[0216] The second unit is used to perform wavelet transform processing on the standardized temperature distribution data in the historical data set to obtain temperature change characteristic data, build a graph convolutional neural network model based on the temperature change characteristic data to obtain heat conduction relationship parameters between photovoltaic modules, input the obtained temperature change characteristic data, heat conduction relationship parameters and environmental monitoring data into a deep neural network for training to obtain a temperature prediction model, input the real-time collected environmental monitoring data into the temperature prediction model, and generate predicted temperature distribution data;

[0217] The third unit is used to input the standardized temperature distribution data and the predicted temperature distribution data into the hierarchical clustering model, identify the temperature anomaly area and calculate the temperature gradient value of the temperature anomaly area, input the spatial position information of the temperature anomaly area and the temperature gradient value into the propagation analysis model, the propagation analysis model calculates the abnormal diffusion range data based on the graph neural network, extracts the characteristic parameters of the temperature anomaly area and compares them with the historical fault data, outputs the fault type determination result and the fault confidence value, generates early warning information including fault location information, early warning level information, fault type information, and processing guidance information according to the temperature gradient value, the abnormal diffusion range data, and the fault confidence value, and sends the early warning information to the operation and maintenance management platform.

[0218] According to a third aspect of the embodiments of the present invention,

[0219] An electronic device is provided, comprising:

[0220] processor;

[0221] a memory for storing processor-executable instructions;

[0222] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0223] A fourth aspect of the embodiments of the present invention is:

[0224] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0225] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0226] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A photovoltaic module temperature monitoring method, characterized in that: include: Acquire the temperature distribution image of the surface of the photovoltaic module through the infrared thermal imaging array, synchronously collect the visible light image of the surface of the photovoltaic module, register and map the temperature distribution image with the visible light image based on the spatial registration algorithm to obtain a registered image, use the edge computing unit to pre-process the registered image, extract the edge contour features of the photovoltaic module, generate the standardized temperature distribution data of the photovoltaic module, and collect the environmental monitoring data of the photovoltaic field at the same time, and store the standardized temperature distribution data and the environmental monitoring data in a distributed database to form a historical data set; Performing wavelet transform processing on the standardized temperature distribution data in the historical data set to obtain temperature change characteristic data, constructing a graph convolutional neural network model based on the temperature change characteristic data to obtain heat conduction relationship parameters between photovoltaic modules, inputting the obtained temperature change characteristic data, heat conduction relationship parameters and environmental monitoring data into a deep neural network for training to obtain a temperature prediction model, inputting the real-time collected environmental monitoring data into the temperature prediction model to generate predicted temperature distribution data; The standardized temperature distribution data and the predicted temperature distribution data are input into the hierarchical clustering model, the temperature anomaly area is identified and the temperature gradient value of the temperature anomaly area is calculated, the spatial position information of the temperature anomaly area and the temperature gradient value are input into the propagation analysis model, the propagation analysis model calculates the abnormal diffusion range data based on the graph neural network, extracts the characteristic parameters of the temperature anomaly area and compares them with the historical fault data, outputs the fault type determination result and the fault confidence value, generates the warning information including the fault location information, warning level information, fault type information and processing guidance information according to the temperature gradient value, the abnormal diffusion range data and the fault confidence value, and sends the warning information to the operation and maintenance management platform.

2. The method according to claim 1, characterized in that Preprocessing the registration image using an edge computing unit to extract edge contour features of the photovoltaic module and generate standardized temperature distribution data of the photovoltaic module includes: Calculating similarity values ​​of image blocks in the registered image, determining a search window size based on the similarity values, constructing an adaptive non-local means filtering model according to the search window size, and inputting the registered image into the adaptive non-local means filtering model for denoising to obtain a denoised image; A calibration plate is used to obtain internal parameters and external parameters of a camera, a radial distortion correction model is constructed according to the internal parameters, a tangential distortion correction model is constructed according to the external parameters, and the denoised image is sequentially input into the radial distortion correction model and the tangential distortion correction model for correction to obtain a corrected image; Constructing a four-layer Gaussian pyramid for the corrected image to obtain first to fourth layer multi-scale images, respectively setting a dual-threshold adaptive edge detection operator on the first to fourth layer multi-scale images to extract edge feature points, performing morphological processing on the edge feature points using an opening operation and a closing operation to obtain optimized edge feature points, and reconstructing the contour of the optimized edge feature points using an ellipse fitting algorithm and a least squares curve fitting algorithm to obtain a boundary contour line of the photovoltaic module; Dividing the photovoltaic module area into regular grid units based on the photovoltaic module boundary contour line, wherein the size of the regular grid unit is determined according to the spatial resolution of the temperature distribution image of the photovoltaic module surface acquired in advance; A bilinear interpolation algorithm is used to calculate the measured temperature value of each regular grid unit, and the ambient temperature value and the standard operating temperature value of the photovoltaic module are obtained. The measured temperature value is subtracted from the ambient temperature value and then combined with the standard operating temperature value to calculate a temperature normalization value; A time series temperature database is established, and the normalized temperature value is associated with the acquisition time information and the spatial location information for storage to generate a standardized temperature data set.

3. The method according to claim 2, characterized in that Constructing a four-layer Gaussian pyramid for the corrected image to obtain first to fourth layer multi-scale images, setting a dual threshold adaptive edge detection operator on the first to fourth layer multi-scale images to extract edge feature points, using open and close operations to perform morphological processing on the edge feature points to obtain optimized edge feature points, and using an ellipse fitting algorithm and a least squares curve fitting algorithm to reconstruct the contour of the optimized edge feature points to obtain a photovoltaic module boundary contour line, including: Using a five-by-five Gaussian kernel to perform Gaussian filtering on the corrected image to obtain a filtered image, performing two-by-two regional downsampling on the filtered image to obtain a first layer image, repeatedly performing the Gaussian filtering and downsampling operations to sequentially obtain a second layer image, a third layer image, and a fourth layer image, and constructing a four-layer Gaussian pyramid structure; Calculate the horizontal gradient components and vertical gradient components of the first to fourth layers of images respectively, calculate the gradient amplitude and gradient direction according to the horizontal gradient components and the vertical gradient components, perform non-maximum suppression on the gradient amplitude to obtain an edge intensity map, calculate the mean and standard deviation of the edge intensity map, set a high threshold and a low threshold according to the mean and standard deviation, and construct a dual-threshold adaptive edge detection operator; Applying the dual threshold adaptive edge detection operator to the first to fourth layer images respectively, extracting strong edge points that meet the high threshold, and extracting weak edge points that meet the low threshold; judging whether there is a strong edge point in the eight neighborhoods of the weak edge point, and if so, merging the weak edge point with the strong edge point to obtain an edge feature point; Using a three-by-three rectangular structure element to perform an opening operation on the edge feature point to obtain an initial edge point; using a five-by-five diamond structure element to perform a closing operation on the initial edge point to obtain an optimized edge point, wherein the optimized edge point includes position coordinates and direction information; Inputting the optimized edge points into an ellipse fitting model, establishing an ellipse equation based on the position coordinates, and solving the ellipse equation parameters using a least squares method to obtain an initial contour line; A sequence of contour points is uniformly sampled on the initial contour line, a cubic spline function is constructed with the interval between two adjacent contour points, the direction information is converted into a tangent constraint condition at the contour point, and the polynomial coefficients of the cubic spline function are solved by the least squares method to obtain the boundary contour line of the photovoltaic module.

4. The method according to claim 1, characterized in that: The obtained temperature change characteristic data, heat conduction relationship parameters and environmental monitoring data are input into a deep neural network for training to obtain a temperature prediction model, and the real-time collected environmental monitoring data is input into the temperature prediction model to generate predicted temperature distribution data, including: The temperature change characteristic data, heat conduction relationship parameters and environmental monitoring data are spliced ​​in the channel dimension to form a fusion feature vector; Constructing a deep neural network model, the deep neural network model includes a feature extraction module, a time series modeling module and a temperature distribution reconstruction module, wherein the feature extraction module includes a plurality of residual blocks, each residual block includes a convolution layer and a batch normalization layer; the time series modeling module includes a forward long short-term memory network unit and a reverse long short-term memory network unit; the temperature distribution reconstruction module includes a multi-layer transposed convolution layer; Input the fused feature vector into a feature extraction module, process it through a convolution layer and a batch normalization layer in a residual block to obtain a feature sequence, and perform a residual connection between the feature sequence and the fused feature vector to obtain a residual feature sequence; Input the residual feature sequence into the time series modeling module, process it through the forward long short-term memory network unit and the reverse long short-term memory network unit to obtain forward features and reverse features respectively, and splice the forward features and the reverse features in the channel dimension to form a time series feature sequence; Rearranging the time series feature sequence into a feature map, and upsampling the feature map through a transposed convolution layer in the temperature distribution reconstruction module to obtain a predicted temperature distribution; Constructing a loss function, wherein the loss function includes a mean square error term between the predicted temperature distribution and the actual temperature distribution, and a spatial smoothness constraint term of the predicted temperature distribution, and using an adaptive optimization algorithm to train the deep neural network model to obtain a temperature prediction model; The real-time collected environmental monitoring data is spliced ​​with the temperature change characteristic data and heat conduction relationship parameters of the previous moment to form a real-time fused feature vector, and the real-time fused feature vector is input into the temperature prediction model. The real-time residual feature sequence is obtained through the feature extraction module in turn, the real-time time series feature sequence is obtained through the time series modeling module, and the predicted temperature distribution data is generated through the temperature distribution reconstruction module.

5. The method according to claim 1, characterized in that The standardized temperature distribution data and the predicted temperature distribution data are input into the hierarchical clustering model, the temperature anomaly area is identified and the temperature gradient value of the temperature anomaly area is calculated, and the spatial position information of the temperature anomaly area and the temperature gradient value are input into the propagation analysis model. The propagation analysis model calculates the abnormal diffusion range data based on the graph neural network, including: Constructing a hierarchical clustering model based on the minimum variance criterion, calculating the distance matrix between the standardized temperature distribution data and the predicted temperature distribution data, inputting the distance matrix into the hierarchical clustering model to construct a hierarchical clustering tree, determining the optimal number of clusters based on the ratio of the sum of intra-class distances to the sum of inter-class distances, and clustering the standardized temperature distribution data to obtain temperature anomaly areas; The temperature change rates of the temperature abnormal area in the horizontal direction and the vertical direction are calculated by using a differential calculation method, the temperature change rates in the horizontal direction and the vertical direction are synthesized to obtain a temperature gradient vector, and the temperature gradient characteristics are calculated according to the temperature gradient vector, wherein the temperature gradient characteristics include a temperature gradient amplitude and a temperature gradient direction; A graph structure is constructed based on the spatial distribution of the temperature anomaly region, the center position of the temperature anomaly region is set as a graph node, a connection relationship is established between corresponding nodes according to the boundary distances of adjacent temperature anomaly regions to form a graph edge, and the center position coordinates, region area, region perimeter, and region maximum inscribed circle radius of the temperature anomaly region are combined with the temperature gradient feature to form a node feature; Construct a graph attention neural network, calculate the attention coefficient between nodes through the attention calculation unit, perform weighted combination of adjacent node features based on the attention coefficient, and obtain updated node representation through nonlinear transformation and residual connection; The node representation output by the graph attention neural network is combined with the temperature gradient feature of the temperature anomaly area, and input into a pre-trained propagation analysis model for feature mapping to obtain the diffusion range data of the temperature anomaly area.

6. The method according to claim 5, characterized in that The node representation output by the graph attention neural network is combined with the temperature gradient feature of the temperature anomaly area, and input into a pre-trained propagation analysis model for feature mapping, so as to obtain the diffusion range data of the temperature anomaly area, including: The node representation vector output by the graph attention neural network is nonlinearly fused with the temperature gradient feature of the temperature anomaly area to obtain a fused feature vector. Constructing a propagation analysis model with direction perception capability, the propagation analysis model comprising multiple direction perception layers, each direction perception layer performing a tensor multiplication operation on the fused feature vector and the node unit direction vector, and obtaining a direction perception feature through a nonlinear activation function; An adaptive weight calculation module is constructed based on the local temperature gradient, wherein the adaptive weight calculation module includes multiple fully connected layers and normalization layers, and the local temperature gradient is input into the adaptive weight calculation module to obtain an adaptive weight coefficient matrix; The adaptive weight coefficient matrix is ​​weightedly combined with the direction perception feature, and the original features are fused through a residual connection structure to obtain a diffusion prediction feature of the temperature anomaly area; The regional overlap loss is obtained by calculating the intersection-and-union ratio of the predicted diffusion area and the real area, and a multi-scale loss function is constructed by combining the boundary prediction loss of the predicted boundary and the real boundary and the gradient continuity loss of the predicted temperature field gradient and the actual temperature field gradient. The propagation analysis model is trained by the multi-scale loss function, and a random dropout mechanism is introduced in the hidden layer of the propagation analysis model to obtain a trained propagation analysis model; The diffusion prediction features are input into the trained propagation analysis model to obtain diffusion range data and diffusion direction data that characterize the diffusion trend of the temperature anomaly area. The diffusion range data includes the spatial boundary coordinates of the abnormal area, and the diffusion direction data includes the main direction and secondary direction of heat propagation.

7. The method according to claim 1, characterized in that Extracting characteristic parameters of the temperature anomaly area and comparing them with historical fault data, outputting fault type determination results and fault confidence values, generating early warning information including fault location information, early warning level information, fault type information, and processing guidance information according to temperature gradient values, abnormal diffusion range data, and fault confidence values, and sending the early warning information to the operation and maintenance management platform includes: Extract characteristic parameters of the temperature anomaly region, wherein the characteristic parameters include morphological characteristics, temperature distribution characteristics and dynamic evolution characteristics, wherein the morphological characteristics include the area ratio, perimeter ratio, roundness coefficient and irregularity of the abnormal region, wherein the temperature distribution characteristics include the highest temperature, average temperature, temperature standard deviation and temperature distribution skewness in the region, wherein the dynamic evolution characteristics include the temperature rise rate, abnormal diffusion rate and diffusion direction stability, and the morphological characteristics, temperature distribution characteristics and dynamic evolution characteristics are combined to construct a multidimensional feature vector; Constructing a deep metric learning model based on a twin neural network, wherein the deep metric learning model includes multiple feature subspaces, each feature subspace corresponds to a different type of feature parameter, and a weight coefficient of each feature subspace is calculated through an attention mechanism; Input the multidimensional feature vector into the deep metric learning model, calculate the similarity between the multidimensional feature vector and the standard feature vector in the historical fault database, weight the influence of the time decay factor on the historical data, and obtain the fault type determination result and the fault confidence value; Inputting the temperature gradient value, the abnormal diffusion range data and the fault confidence value into a fuzzy inference system, calculating the warning level according to a preset fuzzy rule, and retrieving corresponding processing guidance information from a preset knowledge base based on the warning level; The fault location information, the warning level, the fault type determination result and the processing guidance information are hierarchically organized to generate warning information, the fault location information includes the string number and physical coordinates, the processing guidance information includes step-by-step operation suggestions, and the warning information is sent to the operation and maintenance management platform through an encrypted channel.

8. A photovoltaic module temperature monitoring system, used to implement the method according to any one of claims 1 to 7, characterized in that: include: The first unit is used to obtain a temperature distribution image of the surface of a photovoltaic module through an infrared thermal imaging array, synchronously collect a visible light image of the surface of the photovoltaic module, register and map the temperature distribution image with the visible light image based on a spatial registration algorithm to obtain a registered image, pre-process the registered image using an edge computing unit, extract edge contour features of the photovoltaic module, generate standardized temperature distribution data of the photovoltaic module, and simultaneously collect environmental monitoring data of the photovoltaic field area, and store the standardized temperature distribution data and the environmental monitoring data in a distributed database to form a historical data set; The second unit is used to perform wavelet transform processing on the standardized temperature distribution data in the historical data set to obtain temperature change characteristic data, build a graph convolutional neural network model based on the temperature change characteristic data to obtain heat conduction relationship parameters between photovoltaic modules, input the obtained temperature change characteristic data, heat conduction relationship parameters and environmental monitoring data into a deep neural network for training to obtain a temperature prediction model, input the real-time collected environmental monitoring data into the temperature prediction model, and generate predicted temperature distribution data; The third unit is used to input the standardized temperature distribution data and the predicted temperature distribution data into the hierarchical clustering model, identify the temperature anomaly area and calculate the temperature gradient value of the temperature anomaly area, input the spatial position information of the temperature anomaly area and the temperature gradient value into the propagation analysis model, the propagation analysis model calculates the abnormal diffusion range data based on the graph neural network, extracts the characteristic parameters of the temperature anomaly area and compares them with the historical fault data, outputs the fault type determination result and the fault confidence value, generates early warning information including fault location information, early warning level information, fault type information, and processing guidance information according to the temperature gradient value, the abnormal diffusion range data, and the fault confidence value, and sends the early warning information to the operation and maintenance management platform.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images