A Simulation and Early Warning Method for the Generation of Hot Spots on Photovoltaic Backsheets

Abstract

This invention relates to the field of photovoltaic technology and provides a generative adversarial simulation and early warning method for hot spots on photovoltaic backsheets. The method includes the following steps: acquiring infrared images of photovoltaic modules and environmental electrical parameters; processing the infrared images and environmental electrical parameters to obtain thermal infrared images, environmental electrical feature vectors, and hot spot images; constructing a conditional generative adversarial network (GAN) based on the thermal infrared images and environmental electrical feature vectors to generate simulated hot spot images; constructing and training a hot spot early warning classification neural network model, with classification results including no hot spots, minor hot spots, and severe hot spots; compressing and deploying the early warning classification neural network model, and performing remote dynamic optimization through a dynamic feedback learning mechanism. This invention, by introducing a generative adversarial network, generates diverse hot spot images and combines them with conditional information such as electrical parameters, enabling the generation of a large number of simulated images and improving the training effect of the classification network.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic technology, specifically to a method for simulation and early warning of hot spot generation on photovoltaic backsheets. Background Technology

[0002] With the rapid development of the photovoltaic industry, improving the power generation efficiency of photovoltaic systems and optimizing the operating status of photovoltaic modules have become key technical issues. However, in actual operation, the power generation capacity of photovoltaic modules is affected by many external factors, especially hot spot problems, which have a serious impact on the overall performance and long-term stability of photovoltaic systems. Currently, hot spot detection and photovoltaic backsheet fault diagnosis mainly rely on traditional image analysis methods or rule-based machine learning models. Although these methods have certain application effects in some static environments, their performance and robustness are poor in dynamic environments, especially under complex light changes and module temperature fluctuations. In addition, existing methods often fail to accurately capture and predict the formation and development trend of hot spots when data is scarce and the samples are imbalanced, thus affecting the accuracy and reliability of early warning systems for hot spots. The existing technologies mainly have the following shortcomings:

[0003] 1. Insufficient diverse hotspot image data: In existing technologies, hotspot image data samples are scarce, especially since the cost of acquiring hotspot images in complex environments is high, resulting in insufficient samples when training deep learning models.

[0004] 2. Problem of insufficient accuracy in hot spot recognition and classification: In existing technologies, hot spot recognition usually relies on simple feature extraction or classic machine learning algorithms. These methods are difficult to handle complex environmental factors and the variable morphology of hot spots, resulting in poor recognition performance.

[0005] 3. Insufficient early warning and dynamic monitoring of hot spots: In existing technologies, hot spot detection often relies on image processing methods or static models for post-analysis, lacking the ability to provide real-time early warning and dynamic monitoring of hot spots.

[0006] Therefore, there is a need to provide a simulation and early warning method for the generation of hot spots on photovoltaic backsheets, which aims to solve the above problems. Summary of the Invention

[0007] To address the shortcomings of existing technologies, the present invention aims to provide a method for countermeasure simulation and early warning of hot spot generation on photovoltaic backsheets, thereby solving the problems existing in the background technology.

[0008] This invention is implemented as follows: a method for countermeasure simulation and early warning of hot spots generated on photovoltaic backsheets, the method comprising the following steps:

[0009] Infrared images of photovoltaic modules and environmental electrical parameters are collected, and the infrared images of photovoltaic modules and environmental electrical parameters are processed to obtain thermal infrared images, environmental electrical feature vectors, and hot spot images;

[0010] A conditional generative adversarial network is constructed based on thermal infrared images and environmental electrical feature vectors to generate diverse simulated hot spot images.

[0011] A neural network model for early warning classification of hot spots was built and trained based on MobileNetV3-Small and SE module. The classification results include no hot spots, slight hot spots and severe hot spots.

[0012] The early warning classification neural network model is compressed and deployed, and then dynamically optimized remotely through a dynamic feedback learning mechanism.

[0013] As a further aspect of the present invention, the environmental electrical parameters include ambient temperature, irradiance, current, and voltage.

[0014] As a further aspect of the present invention: the step of processing the infrared image of the photovoltaic module and the environmental electrical parameters specifically includes:

[0015] The resolution of the infrared images of photovoltaic modules is standardized, and the images are subjected to grayscale normalization.

[0016] A pre-trained detection network based on the YOLOv5 architecture automatically detects hot spot regions in images, generates corresponding binary mask images, and obtains hot spot images.

[0017] As a further aspect of the present invention: the structure of the adversarial network includes a generator G and a discriminator D. The input is a thermal infrared image and an environmental electrical feature vector, and the output is a simulated hot spot image. The discriminator is used to determine the authenticity of the image.

[0018] As a further aspect of the present invention: the input to the generator G is an infrared image I. input ∈R 256×256×1 and electrical parameter vector C = [T e ,I r ,I m V m ]∈R 4 T e I r I m V m Representing ambient temperature, irradiance, component current, and voltage, respectively, vector C is mapped to an 8×8×8 tensor through a fully connected layer and injected into the bottleneck layer in the middle of the generator as semantic control information for the image generation process.

[0019] As a further aspect of the present invention: the generator adopts a UNet structure, including a symmetrical encoder and decoder; the discriminator adopts a PatchGAN structure, used to determine whether a given image is a real hotspot image, and the input of the discriminator is the original infrared image I. input The discriminator network, composed of five convolutional layers followed by batch normalization and LeakyReLU activation functions, outputs a 30×30 two-dimensional feature map, where each unit represents the probability that a local region in the image is real. During training, a multi-objective joint loss function L is introduced. total =λ1L GAN +λ2L L1 +λ3L mask The generator is supervised, where L GAN This represents traditional adversarial loss, designed to improve image realism. L1 Representing the global pixel difference between the generated image and the real hotspot image helps preserve the basic image structure; L mask The perception loss for the hot spot region is calculated within the mask region to improve the clarity and semantic consistency of local hot spot generation. λ1, λ2, and λ3 are loss weight coefficients.

[0020] As a further aspect of the present invention: the early warning classification neural network model includes an input layer, a backbone network, and an output layer. The input is a thermal infrared image of size 256×256×1. The backbone network is MobileNetV3-Small, which includes depthwise separable convolutions and SE modules.

[0021] As a further aspect of this invention: when training the hotspot early warning classification neural network model, a weighted focus loss is used to solve the class imbalance problem. The focus loss function is defined as: L focal =-α t (1-p t ) γ log(p t ), where p t α represents the model's predicted probability for class t. t The class weights, set to [0.3, 0.4, 0.3], are used to balance the influence of different classes, corresponding to no hotspots, slight hotspots, and severe hotspots, respectively. γ = 2 is a modulation factor to enhance the learning of difficult samples. During model training, the AdamW optimizer is used, with an initial learning rate set to 1 × 10⁻⁶. -3 The learning rate was dynamically adjusted using a cosine annealing learning rate scheduling strategy; the batch size was set to 32 during training; the number of training rounds was set to 80, and an early stopping strategy was adopted to prevent overfitting; the training data included real infrared images collected and simulated hot spot images generated.

[0022] As a further aspect of the present invention: the step of compressing and deploying the early warning classification neural network model specifically includes:

[0023] Channel pruning technique was used to compress the number of parameters of the early warning classification neural network model by 30% to reduce the amount of inference computation. Based on the compression, the TensorRT framework was used to convert the model into FP16 precision format.

[0024] The optimized model is deployed to the NVIDIA Jetson Xavier NX edge computing module, where infrared images are acquired by drones or automated inspection robots, and hot spot recognition models are executed.

[0025] Compared with the prior art, the beneficial effects of the present invention are:

[0026] This invention addresses the problems of data scarcity and imbalanced samples: By introducing a generative adversarial network (GAN), it generates diverse hotspot images and combines them with electrical parameters and other conditional information, enriching the training data. Especially when hotspot image samples are scarce, it can generate a large number of simulated images, compensating for insufficient data and improving the training effect of the classification network. This invention enhances adaptability to complex environments: By introducing diverse samples during model training, it strengthens the model's ability to adapt to environmental changes. Using a deep learning network increases the focus on hotspot regions, thereby achieving more accurate detection in dynamic environments and solving the problem of poor environmental adaptability in existing technologies. This invention achieves real-time hotspot detection and early warning. By combining infrared images and electrical feature data, it can predict the formation trend of hotspots in advance, providing early warnings and thus improving the safety and stability of photovoltaic systems. Attached Figure Description

[0027] Figure 1 This is a flowchart of a simulation and early warning method for combating the generation of hot spots on photovoltaic backsheets.

[0028] Figure 2 This is an architecture diagram of a simulation and early warning method for the generation of hot spots on photovoltaic backsheets. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0030] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0031] like Figure 1and Figure 2 As shown, this embodiment of the invention provides a method for countermeasure simulation and early warning of hot spots generated on photovoltaic backsheets. The method includes the following steps:

[0032] S100 acquires infrared images of photovoltaic modules and environmental electrical parameters, processes the infrared images of photovoltaic modules and environmental electrical parameters to obtain thermal infrared images, environmental electrical feature vectors and hot spot images;

[0033] S200 constructs a conditional generative adversarial network based on thermal infrared images and environmental electrical feature vectors to generate diverse simulated hot spot images;

[0034] S300, based on MobileNetV3-Small and SE module, builds and trains a neural network model for early warning classification of hot spots. The classification results include no hot spots, slight hot spots and severe hot spots.

[0035] S400 compresses and deploys early warning classification neural network models, and performs remote dynamic optimization through a dynamic feedback learning mechanism.

[0036] In this embodiment of the invention, infrared images of photovoltaic modules and environmental electrical parameters are first acquired and processed. The purpose is to construct a high-quality training dataset for subsequent hot spot image generation and recognition models. Specifically, representative thermal infrared images of the modules need to be acquired in actual photovoltaic scenarios, and environmental and electrical parameters around the modules, including ambient temperature, irradiance, current, and voltage, need to be recorded simultaneously. The resolution of the acquired thermal infrared images needs to be standardized and set to 256×256 pixels. During image preprocessing, the images are first subjected to grayscale normalization using the following standardization formula: Where I represents the original pixel value, μ represents the overall grayscale mean of the image, and σ represents the standard deviation. This processing method helps improve the numerical stability during neural network training.

[0037] To enhance the model's robustness to image distortion and angular changes, data augmentation operations are introduced during preprocessing, including image rotation, horizontal / vertical flipping, and brightness perturbation. Furthermore, a pre-trained detection network based on the YOLOv5 architecture is used to automatically detect hotspot regions in the image, generating corresponding binary masks for subsequent local region supervision of the adversarial generative model. Step S100 ultimately outputs a structured data triple: (thermal infrared image, electrical parameter vector, hotspot mask (hotspot image)), providing standard input for subsequent model training.

[0038] In this embodiment of the invention, a Conditional Generative Adversarial Network (cGAN) is then constructed to generate simulated hot spot images. Since real hot spot images suffer from problems such as scarcity, uneven distribution, and scene limitations in practical applications, they cannot directly meet the sample requirements for deep learning training. Therefore, this step uses a cGAN to generate high-quality, diverse simulated hot spot images, assisting the classification model in learning hot spot features and improving recognition robustness. This cGAN mainly consists of two sub-modules: a generator network (G) and a discriminator network (D). The input is a normal infrared image and environmental electrical feature vectors, and the output is a simulated hot spot image. The discriminator is used to determine whether the image is real or fake.

[0039] Specifically, the input to generator G is an infrared image I. input ∈R 256×256×1 and electrical parameter vector C = [T e ,I r ,I m V m ]∈R 4 The components represent ambient temperature, irradiance, component current, and voltage, respectively. This conditional vector C is mapped to an 8×8×8 tensor through a fully connected layer and injected into the bottleneck layer in the middle of the generator as semantic control information for the image generation process.

[0040] The generator adopts a UNet architecture, consisting of a symmetrical encoder and decoder. The encoder comprises four convolutional layers, each containing a convolution operation (stride of 2), batch normalization (BatchNorm), and LeakyReLU activation function, extracting 64, 128, 256, and 512 channel features respectively. Channel concatenation is performed on the electrical feature tensor mapped to the deepest layer of the encoder and the fully connected layer, forming a fused vector. The decoder performs upsampling through deconvolution operations, recovering the image spatial structure layer by layer, with corresponding output channels of 256, 128, and 64. Finally, a Sigmoid or Tanh activation function compresses the image to the pixel range [-1, 1]. During decoding, the corresponding encoded features are fused using skip connections to preserve low-level texture information.

[0041] The discriminator uses a PatchGAN architecture to determine whether a given image is a real hotspot image. Its input is the original infrared image I. input Compared with the hot spot image to be determined (real image I) real Or generate image I fake =G(I inputThe channel concatenation result of C). The entire discriminator network consists of five convolutional layers, each followed by batch normalization and LeakyReLU activation function, and the final output is a two-dimensional feature map of size 30×30, where each unit represents the probability that a local region of the image is real. During training, a multi-objective joint loss function is introduced to supervise the generator, which is defined as follows: L total =λ1L GAN +λ2L L1 +λ3L mask Among them, L GAN This represents traditional adversarial loss, designed to improve image realism. L1 Representing the global pixel difference between the generated image and the real hotspot image helps preserve the basic image structure; L mask For hotspot region perception loss, generation error is calculated only within the masked area to improve the sharpness and semantic consistency of local hotspot generation. The loss weight coefficients are empirically set as follows: λ1 = 0.5, λ2 = 100, λ3 = 50.

[0042] The Adam optimizer was used during training, with a learning rate set to 2×10. -4 The momentum parameters were set to β1 = 0.5 and β2 = 0.999, the batch size was 16, and the training epochs were 100. During training, the generator and discriminator were optimized alternately. After training, the quality of the generated images was evaluated using human experts and the Structural Similarity Index (SSIM). The average SSIM between the simulated images and their corresponding real images reached over 0.92, indicating that the generated hotspot images highly approximate real images in terms of structure, texture, and brightness, and possess high discriminative value. The simulated images output in step S200 will serve as important training data, equal in quantity to the real samples, and will be input into the subsequent classification model, effectively improving its generalization ability and early recognition capability for hotspot features.

[0043] In this embodiment of the invention, a hot spot early warning classification neural network model is then constructed and trained. The aim is to identify the presence of hot spots on a photovoltaic backsheet and classify their risk levels based on generated hot spot images and real images by training a deep learning classification network. The classification task outputs three categories: no hot spots, minor hot spots, and severe hot spots. Furthermore, the network structure incorporates a Squeeze-and-Excitation (SE) module to enhance the network's ability to focus on hot spot regions, thereby improving the model's classification accuracy for different types of hot spots. This classification network uses MobileNetV3-Small as the backbone network, combined with lightweight convolutional operations to reduce computational load, making it suitable for deployment on edge devices. The model structure includes the following modules: 1. Input Layer: The input is a 256×256×1 thermal infrared image. The image is standardized before being used as the network input. 2. Backbone Network (MobileNetV3-Small): Includes depthwise separable convolutions and a Squeeze-and-Excitation (SE) module. Depthwise separable convolutions reduce computational cost, maintaining good performance with low computational resource consumption. The SE module is used to enhance the learning of key regions (hotspot regions) by adaptively adjusting the weights of each channel, thereby improving the model's ability to represent hotspot features. 3. Output Layer: This layer outputs three classification results through a fully connected layer: no hotspot (0), slight hotspot (1), and severe hotspot (2). The output uses the Softmax activation function to transform the predicted probability of each category into a probability distribution of the three labels.

[0044] Step S300 employs weighted focal loss to address class imbalance, particularly when dealing with minor hotspots (classes with fewer samples). This is achieved by reducing the weight of easily classified samples and increasing the focus on difficult-to-classify samples. The focal loss function is defined as: L focal =-α t (1-p t ) γ log(p t ), where p t Let α be the predicted probability of the model for class t. t The category weights are used to balance the influence of different categories. They are set to [0.3, 0.4, 0.3], which correspond to no hotspots, slight hotspots, and severe hotspots, respectively. γ = 2 is a modulation factor to enhance the learning of difficult samples.

[0045] During model training, the AdamW optimizer was used, with an initial learning rate set to 1×10⁻⁶. -3The learning rate was dynamically adjusted using a cosine annealing scheduling strategy; the batch size was set to 32 during training; the number of training epochs was set to 80, and an early stopping strategy was adopted to prevent overfitting; the training data included real infrared images acquired in step S100 and simulated hot spot images generated in step S200. The two were mixed in a 1:1 ratio to enhance the model's adaptability to different scenarios.

[0046] The classification model outputs three categories: no hotspots (0), slight hotspots (1), and severe hotspots (2). The predicted probability for each category is also output to assess the model's confidence in the current classification result. Furthermore, if the model detects hotspots, it can further integrate time-series information for early warning, judging the potential development trend of hotspots based on historical image sequences, thus providing an early warning. The trained classification network achieved a recognition accuracy of 91.7% on the test set, with a significant improvement in the recognition accuracy of slight hotspots.

[0047] In this embodiment of the invention, model compression, deployment, and remote dynamic optimization are performed at the end. To ensure the real-time performance and deployment feasibility of the model in photovoltaic scenarios, step S400 compresses and optimizes the trained classification model. First, channel pruning technology is used to compress the model parameters by 30% to reduce the amount of inference computation. Based on the compression, the TensorRT framework is used to convert the model to FP16 precision format to further improve the execution efficiency of edge devices. The optimized model is deployed to the NVIDIA Jetson Xavier NX edge computing module. This platform is equipped with drones or automated inspection robots to collect infrared images and execute the hot spot recognition model. The inference speed is controlled within 50 milliseconds to support real-time image processing requirements. The model output is structured data, including timestamp, hot spot risk level, confidence level, and hot spot location coordinates, as shown in the following example: Time: April 1, 2025, 10:23; Hot spot level: Medium risk; Confidence level: 86%; Location coordinates: [(100,140),(120,160)]. Meanwhile, to ensure the long-term stable operation of the model and its adaptability to changes in the field environment, a dynamic feedback learning mechanism was constructed. Users can manually annotate false positives or false negatives on the platform, and the system automatically collects feedback data and performs incremental fine-tuning every 7 days. Through distillation learning, the original recognition capabilities are retained while integrating knowledge from new samples, ensuring continuous optimization and iteration of the model.

[0048] This invention presents a hot spot generation and simulation method based on Generative Adversarial Networks (GANs). It can generate diverse hot spot images and incorporate electrical parameters for conditional control, thus solving the problem of insufficient training samples caused by the scarcity of hot spot data in existing technologies. By generating high-quality simulated hot spot images, this invention improves the generalization ability of the hot spot classification network, enabling the model to maintain high detection accuracy under different photovoltaic system environments.

[0049] This invention combines a hotspot classification method using a deep learning network (MobileNetV3-Small and the SE module). By introducing the Squeeze-and-Excitation (SE) module, it enhances attention to hotspot regions and improves the network's ability to identify small hotspots. Compared to traditional methods, this invention can more accurately detect hotspots and classify their severity when dealing with complex hotspot backgrounds (such as shadows, lighting changes, etc.), avoiding the limitations of traditional image processing methods in handling complex environmental changes.

[0050] This invention combines a joint training strategy of generative adversarial networks (GANs) and deep learning classification networks. By jointly training the generator and classification network, it addresses the structural differences between generated images and real hotspot images, improving the robustness and accuracy of the classification network. Compared to traditional single-model training methods, this invention enables efficient training with limited data resources, enhancing the performance of hotspot detection systems in practical applications.

[0051] The above description only details the preferred embodiments of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0052] It should be understood that although the steps in the flowcharts of the various embodiments of the present invention are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the various embodiments may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps.

[0053] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0054] Other embodiments of this disclosure will readily occur to those skilled in the art upon consideration of the disclosure in the specification and embodiments. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of this disclosure are indicated by the claims.

Claims

1. A method for countermeasure simulation and early warning of hot spot generation on photovoltaic backsheets, characterized in that, The method includes the following steps: Infrared images of photovoltaic modules and environmental electrical parameters are collected, and the infrared images of photovoltaic modules and environmental electrical parameters are processed to obtain thermal infrared images, environmental electrical feature vectors, and hot spot images; A conditional generative adversarial network (GAN) is constructed based on thermal infrared images and environmental electrical feature vectors to generate diverse simulated hot spot images. The GAN structure includes a generator and a discriminator. During training, a multi-objective joint loss function L is introduced. total =λ1L GAN +λ2L L1 +λ3L mask The generator is supervised, where L GAN This represents traditional adversarial loss, designed to improve image realism. L1 Representing the global pixel difference between the generated image and the real hotspot image helps preserve the basic image structure; L mask For the perception loss of hot spot region, the generation error is calculated in the mask region to improve the clarity and semantic consistency of local hot spot generation. λ1, λ2 and λ3 are loss weight coefficients. A neural network model for early warning classification of hot spots was constructed and trained based on MobileNetV3-Small and the SE module. The classification results include no hot spots, slight hot spots, and severe hot spots. A weighted focus loss was used to address the class imbalance problem during the training of the neural network model. The focus loss function is defined as: L focal =-α t (1-p t ) γ log(p t ), where p t α represents the model's predicted probability for class t. t The class weights, set to [0.3, 0.4, 0.3], are used to balance the influence of different classes, corresponding to no hotspots, slight hotspots, and severe hotspots, respectively; γ = 2 is a modulation factor to enhance the learning of difficult samples; The early warning classification neural network model is compressed and deployed, and then dynamically optimized remotely through a dynamic feedback learning mechanism.

2. The method for countermeasure simulation and early warning of hot spot generation on photovoltaic backsheets according to claim 1, characterized in that, The environmental electrical parameters include ambient temperature, irradiance, current, and voltage.

3. The method for countermeasure simulation and early warning of photovoltaic backsheet hot spot generation according to claim 1, characterized in that, The steps for processing the infrared image of the photovoltaic module and the environmental electrical parameters specifically include: The resolution of the infrared images of photovoltaic modules is standardized, and the images are subjected to grayscale normalization. A pre-trained detection network based on the YOLOv5 architecture automatically detects hot spot regions in images, generates corresponding binary mask images, and obtains hot spot images.

4. The method for countermeasure simulation and early warning of photovoltaic backsheet hot spot generation according to claim 1, characterized in that, The structure of an adversarial network includes a generator G and a discriminator D. The input is a thermal infrared image and an environmental electrical feature vector, and the output is a simulated hot spot image. The discriminator is used to determine whether the image is real or fake.

5. The method for countermeasure simulation and early warning of photovoltaic backsheet hot spot generation according to claim 4, characterized in that, The input to the generator G is an infrared image I. input ∈R 256×256×1 and electrical parameter vector C = [T e ,I r ,I m V m ]∈R 4 T e I r I m V m Representing ambient temperature, irradiance, component current, and voltage, respectively, vector C is mapped to an 8×8×8 tensor through a fully connected layer and injected into the bottleneck layer in the middle of the generator as semantic control information for the image generation process.

6. The method for countermeasure simulation and early warning of photovoltaic backsheet hot spot generation according to claim 5, characterized in that, The generator employs a UNet architecture, including a symmetrical encoder and decoder; the discriminator uses a PatchGAN architecture to determine whether a given image is a real hotspot image. The input to the discriminator is the original infrared image I. input The discriminator network consists of five convolutional layers, each followed by batch normalization and LeakyReLU activation functions. The final output is a two-dimensional feature map of size 30×30, where each unit represents the probability that a local region of the image is real.

7. The method for countermeasure simulation and early warning of photovoltaic backsheet hot spot generation according to claim 1, characterized in that, The early warning classification neural network model includes an input layer, a backbone network, and an output layer. The input is a thermal infrared image of size 256×256×1. The backbone network is MobileNetV3-Small, which includes depthwise separable convolutions and SE modules.

8. The method for countermeasure simulation and early warning of photovoltaic backsheet hot spot generation according to claim 7, characterized in that, During model training, the AdamW optimizer was used, with an initial learning rate set to 1×10⁻⁶. -3 The learning rate was dynamically adjusted using a cosine annealing learning rate scheduling strategy; the batch size was set to 32 during training; the number of training rounds was set to 80, and an early stopping strategy was adopted to prevent overfitting; the training data included real infrared images collected and simulated hot spot images generated.

9. The method for countermeasure simulation and early warning of photovoltaic backsheet hot spot generation according to claim 1, characterized in that, The steps for compressing and deploying the early warning classification neural network model specifically include: Channel pruning technique was used to compress the number of parameters of the early warning classification neural network model by 30% to reduce the amount of inference computation. Based on the compression, the TensorRT framework was used to convert the model into FP16 precision format. The optimized model is deployed to the NVIDIA Jetson Xavier NX edge computing module, where infrared images are acquired by drones or automated inspection robots, and hot spot recognition models are executed.

Images