A deep learning-based sunken area photovoltaic panel blocking identification method

Abstract

The application provides a kind of based on deep learning's sunken area photovoltaic panel obstruction identification method, it is related to photovoltaic power generation and operation and maintenance technical field.A kind of based on deep learning's sunken area photovoltaic panel obstruction identification system, comprising: image acquisition module, for collecting the visible light image and infrared thermal image of sunken area photovoltaic panel;Data preprocessing module is connected with image acquisition module, for the registration, denoising and illumination correction of the visible light image and infrared thermal image, obtain preprocessed image pair, wherein illumination correction adopts adaptive gamma transformation, by image local brightness determines.It is through double wave band image preprocessing, multi-modal feature fusion, curvature deformation detection and multi-task identification, cooperate operation and maintenance decision model, realize sunken area photovoltaic panel multi-class obstruction accurate identification and closed loop operation and maintenance, effectively solve the problem of low precision, single mode, collapse difficult to detect and response lag of prior art.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power generation and operation and maintenance technology, specifically a method for identifying photovoltaic panel shading in subsidence areas based on deep learning. Background Technology

[0002] Photovoltaic power stations in coal mining subsidence areas are photovoltaic power generation facilities built on water or land areas formed by coal mining subsidence. The unique environmental factors such as continuous terrain deformation, coal dust dispersion, and vegetation regeneration cause frequent shading of the photovoltaic panels due to vegetation cover, dust accumulation, and deformation and collapse of the panels caused by uneven ground subsidence. These shading types are diverse, vary in scale, and are randomly distributed, significantly reducing the effective light-receiving area of ​​the photovoltaic modules, causing mismatch losses and hot spot effects, and seriously affecting the power generation efficiency and operational safety of the power station. Therefore, achieving accurate and real-time identification of various types of shading on photovoltaic panels in subsidence areas has become a key requirement for improving operation and maintenance efficiency and ensuring power generation revenue.

[0003] Existing occlusion recognition methods have the following drawbacks: Firstly, traditional visible light image analysis methods rely on artificially designed shallow features such as color and texture. Under complex lighting and shadow interference in the sinkhole area, it is difficult to distinguish between dust accumulation and normal surface. Furthermore, the ability to distinguish between vegetation and background is weak, resulting in low accuracy and poor robustness in occlusion detection.

[0004] Secondly, most existing technologies only utilize visible light single-mode information and fail to integrate infrared thermal images to obtain abnormal temperature differences caused by shading. They cannot effectively identify the local temperature rise characteristics of photovoltaic panels caused by dust accumulation and shading, thus limiting the ability to accurately distinguish the type of shading.

[0005] Third, traditional methods lack the ability to detect the structural deformation of photovoltaic panels due to the unique topographic collapse and shading characteristics of subsidence areas. They cannot capture the shading areas caused by collapse deformation by analyzing the curvature changes of the panel edges, making such shading essentially unrecognizable.

[0006] Fourth, while existing deep learning detection solutions can locate occlusion positions, they typically only output a single classification label or a rough bounding box. They cannot simultaneously achieve multi-classification of occlusion types and pixel-level segmentation of occlusion areas. Furthermore, the recognition results lack quantitative indicators that can be directly used for operation and maintenance decisions, making it difficult to automatically generate cleaning priorities and inverter power control commands, resulting in a lag between recognition and execution. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a deep learning-based method for identifying photovoltaic panel shading in subsidence areas. This method solves the problems of low accuracy, poor robustness, single modality, inability to identify collapsed shading, lack of pixel-level segmentation and operation and maintenance quantification support, and slow response in existing methods for identifying photovoltaic panel shading in subsidence areas.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a deep learning-based photovoltaic panel shading identification system for subsidence areas, comprising: The image acquisition module is used to acquire visible light and infrared thermal images of photovoltaic panels in the subsidence area; The data preprocessing module, connected to the image acquisition module, is used to register, denoise, and correct the illumination of the visible light image and the infrared thermal image to obtain preprocessed image pairs. The illumination correction employs adaptive gamma transform. , Determined by the local brightness of the image; A photovoltaic panel region extraction module, connected to a data preprocessing module, is used to segment out the photovoltaic panel region mask from a preprocessed image pair. The multimodal occlusion feature extraction module is connected to the photovoltaic panel region extraction module. It has a built-in dual-stream convolutional network to extract texture features and thermal difference features from the visible light and infrared image branches, respectively, and calculates deformation features based on the edge curvature of the photovoltaic panel region, and fuses them to form multi-dimensional occlusion features.

[0009] Preferably, the denoising process of the data preprocessing module employs non-local mean filtering for each pixel in the image. The estimated value Obtained through the following formula: ; in, , In pixels and The image patch centered on This represents the Gaussian-weighted Euclidean distance. These are the filter parameters. As a normalization coefficient, in the adaptive gamma transform, the parameter γ is dynamically adjusted according to the average brightness of the image blocks, and illumination balance is achieved by increasing the γ value in dark areas and decreasing the γ value in bright areas.

[0010] Preferably, the multimodal occlusion feature extraction module includes a texture branch, a thermal difference branch, and a deformation feature calculation unit. The deformation feature calculation unit extracts the edge contour point set based on the photovoltaic panel area mask. Calculate the discrete curvature at each point The formula for calculating curvature is: ; in , The curvature value is mapped to the image coordinate system to generate a deformation feature map, which is then concatenated with the texture feature and thermal difference feature in the channel dimension to obtain the multi-dimensional occlusion feature.

[0011] Preferably, the occlusion recognition and localization module includes a region proposal network, a classification branch, and a mask regression branch, and is jointly trained using a multi-task loss function, the loss function of which is: ; in Cross-entropy loss is used for occlusion type classification. For real labels, To predict probabilities, , is the Dice loss of the binary mask for the occluded region. For real mask pixels, To predict mask pixels, To balance the weights.

[0012] Preferably, it also includes an occlusion warning and operation and maintenance decision module, which is connected to the occlusion identification and positioning module. It is used to receive the occlusion type and the binary mask of the occlusion area, calculate the occlusion area ratio, and generate cleaning priority instructions and inverter power control instructions based on a preset loss model.

[0013] A deep learning-based method for identifying photovoltaic panel shading in subsidence areas includes the following steps: Step S1: Obtain visible light and infrared thermal images of the photovoltaic panels in the subsidence area; Step S2: Register, denoise, and perform adaptive gamma correction on the visible light image and the infrared thermal image to obtain a preprocessed image pair, wherein the adaptive gamma correction parameters are dynamically determined by the local brightness mean of the image. Step S3: Extract the photovoltaic panel region mask from the preprocessed image pairs using a semantic segmentation network; Step S4: Input the preprocessed image and the photovoltaic panel area mask into the dual-flow depth network to extract texture features and thermal difference features respectively. At the same time, calculate the deformation features based on the edge curvature extracted from the photovoltaic panel area mask, and fuse the texture features, thermal difference features and deformation features to generate multi-dimensional occlusion features. Step S5: Classify and locate occlusions based on multi-dimensional occlusion features, and output the occlusion type and the corresponding binary mask of the occlusion area. The occlusion types include vegetation occlusion, dust accumulation occlusion and terrain collapse occlusion.

[0014] Preferably, the adaptive gamma correction in step S2 adopts a transform-type method. ,in The parameter is a constant. The brightness of the local area of ​​the image is dynamically determined to enhance dark areas and suppress bright areas. The noise reduction process uses non-local mean filtering, and the weights are determined by Gaussian weighted similarity of image blocks.

[0015] Preferably, the deformation feature extraction process in step S4 is as follows: obtain the photovoltaic panel edge contour point sequence based on the photovoltaic panel area mask, and calculate the discrete curvature for each point on the contour. : ; in , It is a first-order difference. , The second-order difference transforms the curvature values ​​linearly to a preset grayscale range, generating a deformation feature map corresponding to the photovoltaic panel area.

[0016] Preferably, in step S5, a multi-task learning framework is used to simultaneously perform occlusion type classification and occlusion region segmentation, with a total loss function. for: ; in Cross-entropy loss for occlusion type classification and They are respectively real and predicted types. and These represent the mask pixel values ​​for the actual and predicted occluded areas, respectively. , This is the loss weighting coefficient.

[0017] Preferably, after steps S1 to S5 are performed, step S6 is further included: Based on the identified shading type and the binary mask of the shading area, the proportion of the shading area to the total area of ​​the photovoltaic panel is calculated. Combined with the output characteristic curve of the photovoltaic module, the power loss is estimated. With the goal of minimizing power loss and cleaning cost, a cleaning scheduling strategy is generated, and the corresponding inverter maximum power point tracking adjustment reference signal is output.

[0018] This invention provides a deep learning-based method for identifying photovoltaic panel shading in sunken areas. It offers the following advantages: 1. This invention performs illumination correction and noise reduction preprocessing on dual-band images through adaptive gamma transform and nonlocal mean filtering, and completes the high-robust occlusion recognition effect by combining deep texture features extracted by deep networks. It effectively reduces the interference of complex lighting, terrain shadows and dust noise in subsidence areas, and solves the defects of low recognition accuracy and poor environmental adaptability of traditional shallow feature methods.

[0019] 2. This invention constructs a dual-modal input system by simultaneously acquiring visible light and infrared thermal images, and uses a dual-stream convolutional network to extract texture features and thermal difference features respectively to achieve multi-source information fusion and recognition. This fully captures the temperature distribution differences corresponding to different occlusion types and makes up for the insufficiency of single visible light modality in distinguishing the causes of occlusion.

[0020] 3. This invention extracts edge contour point sets by using a photovoltaic panel area mask and calculates discrete curvature to obtain deformation features. It also uses a multi-dimensional feature fusion framework to achieve accurate detection of deformation occlusion of terrain collapse-type structures, thus realizing effective perception of hidden panel surface deformation and filling the technical gap that traditional methods cannot identify collapse-type occlusion.

[0021] 4. This invention synchronously outputs occlusion type classification results and pixel-level region masks through a multi-task joint training network, and generates cleaning priorities and control instructions in conjunction with a preset loss model to complete the integrated closed-loop effect of identification and maintenance, providing quantitative decision-making basis and eliminating the response lag problem between identification and handling in traditional solutions. Attached Figure Description

[0022] Figure 1 This is a system framework diagram of the present invention; Figure 2 This is a flowchart of the steps of the present invention; Figure 3 This is a diagram of the multimodal feature extraction architecture of the present invention; Figure 4 This is a comparison chart of the occlusion recognition accuracy under different occlusion area ratios of the present invention. Figure 5 This is a comparison chart showing the accuracy of occlusion type classification under different light intensities according to the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Example: like Figure 1-5 As shown, this embodiment of the invention provides a deep learning-based photovoltaic panel shading recognition system for subsidence areas, comprising: The image acquisition module is used to acquire visible light and infrared thermal images of photovoltaic panels in the subsidence area; The data preprocessing module, connected to the image acquisition module, is used to register, denoise, and correct the illumination of visible light and infrared thermal images, resulting in preprocessed image pairs. Illumination correction employs adaptive gamma transform. , Determined by the local brightness of the image, the denoising process in the data preprocessing module uses non-local mean filtering to filter the pixels in the image. The estimated value Obtained through the following formula: ,in, , In pixels and The image patch centered on This represents the Gaussian-weighted Euclidean distance. These are the filter parameters. As the normalization coefficient, in the adaptive gamma transform, the parameter γ is dynamically adjusted according to the average brightness of the image blocks. Illumination balance is achieved by increasing the γ value in dark areas and decreasing the γ value in bright areas. A photovoltaic panel region extraction module, connected to a data preprocessing module, is used to segment out the photovoltaic panel region mask from a preprocessed image pair. A multimodal occlusion feature extraction module, connected to the photovoltaic panel region extraction module, incorporates a dual-stream convolutional network. It extracts texture and thermal difference features from visible light and infrared image branches respectively, and calculates deformation features based on the edge curvature of the photovoltaic panel region. These features are then fused to form a multi-dimensional occlusion feature. The multimodal occlusion feature extraction module includes texture branches, thermal difference branches, and a deformation feature calculation unit. The deformation feature calculation unit extracts the edge contour point set based on the photovoltaic panel region mask. Calculate the discrete curvature at each point The formula for calculating curvature is: ;in , The curvature value is mapped to the image coordinate system to generate a deformation feature map, which is then spliced ​​with texture features and thermal difference features in the channel dimension to obtain multi-dimensional occlusion features. The occlusion recognition and localization module, connected to the multimodal occlusion feature extraction module, is used to identify vegetation occlusion, dust accumulation occlusion, or terrain collapse occlusion types based on multi-dimensional occlusion features, and outputs a binary mask of the occluded area. The occlusion recognition and localization module includes a region proposal network, a classification branch, and a mask regression branch, and is jointly trained using a multi-task loss function. The loss function is: ;in Cross-entropy loss for occlusion type classification For real labels, To predict probabilities, The Dice loss of the binary mask for the occluded area, For real mask pixels, To predict mask pixels, To balance the weights; It also includes an obstruction warning and operation and maintenance decision module, which is connected to the obstruction identification and positioning module. It is used to receive the obstruction type and the binary mask of the obstruction area, calculate the proportion of obstruction area, and generate cleaning priority instructions and inverter power control instructions based on the preset loss model.

[0025] Example 1: This embodiment is the basic deployment implementation plan of the deep learning-based photovoltaic panel shading recognition system for subsidence areas. The system's image acquisition module consists of a high-definition visible light industrial camera, an uncooled infrared thermal imager, and a matching motorized pan-tilt bracket. All equipment is fixedly installed on galvanized steel monitoring towers along the photovoltaic array in the subsidence area. One tower is set up for every two strings along the string direction, and the acquisition angle completely covers all photovoltaic panels in the corresponding area. The acquisition cycle is set to synchronously acquire once per hour. The acquired visible light images and infrared thermal images are transmitted to a dustproof and waterproof edge computing industrial control computer deployed on site via an industrial Ethernet bus.

[0026] The data preprocessing module, photovoltaic panel area extraction module, multimodal shading feature extraction module, and shading identification and positioning module are all integrated into the edge computing industrial control computer as executable program units. The shading early warning and operation and maintenance decision module is connected to the existing SCADA monitoring system of the photovoltaic power station through industrial optical fiber. When the system is running, the synchronously acquired dual-band raw images first enter the preprocessing stage to complete image quality optimization and spatial alignment, and then sequentially complete photovoltaic panel area segmentation, multi-dimensional shading feature extraction, shading type determination and area segmentation. Finally, the operation and maintenance decision module generates corresponding control commands. The entire processing link is completed locally at the edge, without relying on cloud computing power, and is suitable for the field environment with large fluctuations in network conditions in the subsidence area.

[0027] Example 2: This embodiment, based on the basic deployment architecture, further specifies the details of the data preprocessing and multimodal feature extraction stages. In the image registration stage of the data preprocessing module, the endo- and extrinsic parameters of the visible light camera and infrared thermal imager are pre-calibrated using a checkerboard calibration board. After acquisition, pixel-level spatial alignment of the dual-band images is achieved based on the calibration parameters. The noise reduction process sets up a search window and a neighborhood window of fixed size, effectively suppressing particle noise caused by sandstorms in the subsidence area. The illumination correction stage divides the image into equal-sized sub-blocks, calculates the average brightness of each sub-block, and dynamically matches corresponding correction parameters, effectively adapting to uneven illumination scenarios in the subsidence area, such as backlighting in the morning and evening, and terrain shadows.

[0028] The photovoltaic panel region extraction module employs a processing path combining edge detection and morphological closing operations to extract the complete outer contour of the photovoltaic panel from the preprocessed image and generate a region mask, filtering out irrelevant areas such as soil and weeds in the background. The multimodal occlusion feature extraction module uses a lightweight backbone structure in its dual-stream convolutional network. The texture branch extracts features such as grayscale distribution and texture roughness from the visible light image, which can distinguish between uniform dust coverage and patchy vegetation distribution. The thermal difference branch extracts features such as temperature gradient and the proportion of hot and cold areas from the infrared image, which can identify abnormal temperature rises in the panel structure caused by terrain collapse. The deformation feature calculation unit calculates the discrete curvature point by point based on the extracted photovoltaic panel edge contour point set to generate the corresponding deformation feature map. The three types of features are stitched together in the channel dimension to form a complete multi-dimensional occlusion feature, which is then sent to the subsequent recognition stage.

[0029] Example 3: This embodiment focuses on the engineering implementation of shading identification, localization, and operation and maintenance decision-making. The shading identification and localization module uses an integrated detection and segmentation network, which is pre-trained offline using a multi-scenario shading sample dataset collected in the subsidence area. The training samples cover three typical operating conditions under different seasons and lighting conditions: vegetation growth shading, surface dust shading, and terrain collapse shading. In the real-time inference stage, a region proposal network generates candidate shading region boxes, which are then output as shading type labels via a classification branch and pixel-level shading binary masks via a mask regression branch. The shading warning and operation and maintenance decision-making module has a built-in power loss mapping table corresponding to the three types of shading. It estimates the power generation loss of the current string by combining the calculated shading area ratio, and then generates cleanup priority instructions according to the loss level from high to low. The instructions can be directly pushed to the mobile terminal of the operation and maintenance personnel, and the specific string location and type of the corresponding shading are marked simultaneously. At the same time, the module can generate inverter power adjustment instructions, which are sent to the inverter of the corresponding photovoltaic string through the SCADA system to adjust the output power to avoid the hot spot effect and component damage risk caused by shading. For obstructions identified as terrain collapse, the module can also directly access the power plant's safety alarm, prompting maintenance personnel to prioritize conducting safety checks on the panel structure.

[0030] Example 4: This embodiment further supplements the system's adaptable structure and functions to address the special working conditions of continuous terrain subsidence and uneven surface undulation in coal mining subsidence areas. The image acquisition module can also be mounted on a track-mounted inspection robot. The robot travels along a dedicated rust-proof inspection track laid between the rows of photovoltaic arrays. It can automatically adjust the acquisition height and gimbal angle according to the terrain undulations of the subsidence area, achieving comprehensive image acquisition of the entire array of photovoltaic panels without blind spots, and avoiding the perspective shift and acquisition blind spots caused by terrain subsidence of fixed towers.

[0031] The data preprocessing module incorporates a terrain correction unit, which can access real-time data from existing surface subsidence monitoring systems in the subsidence area. It corrects elevation deviations in infrared image temperature measurements, eliminating temperature measurement errors caused by slab tilt and height variations. The deformation characteristic calculation unit can also access a historical curvature database. By comparing multi-period edge curvature change data, it determines the development rate of slab deformation and subsidence trends, providing early warnings of terrain collapse risks before deformation becomes visible to the naked eye. The shading warning and operation and maintenance decision-making module can connect to the subsidence area ecological restoration and management platform, synchronizing vegetation shading data to the restoration system. This, combined with vegetation pruning and ecological maintenance operations, enables the coordinated advancement of photovoltaic operation and maintenance and subsidence area ecological restoration.

[0032] A deep learning-based method for identifying photovoltaic panel shading in subsidence areas includes the following steps: Step S1: Obtain visible light and infrared thermal images of the photovoltaic panels in the subsidence area; Step S2: Register, denoise, and perform adaptive gamma correction on the visible light image and the infrared thermal image to obtain a preprocessed image pair. The adaptive gamma correction parameters are dynamically determined by the local brightness mean of the image, and the adaptive gamma correction adopts a transform-based method. ,in The parameter is a constant. The image is dynamically determined by the local brightness mean, which enhances dark areas and suppresses bright areas. The noise reduction process uses non-local mean filtering, and the weights are determined by Gaussian weighted similarity of image blocks. Step S3: Extract the photovoltaic panel region mask from the preprocessed image pairs using a semantic segmentation network; Step S4: Input the preprocessed image and the photovoltaic panel area mask into a dual-flow depth network to extract texture features and thermal difference features respectively. Simultaneously, calculate deformation features based on the edge curvature extracted from the photovoltaic panel area mask. Fuse texture features, thermal difference features, and deformation features to generate multi-dimensional occlusion features. The deformation feature extraction process is as follows: Obtain the photovoltaic panel edge contour point sequence based on the photovoltaic panel area mask, and calculate the discrete curvature for each point on the contour. : ;in , It is a first-order difference. , The second-order difference method linearly transforms the curvature values ​​to a preset grayscale range, generating a deformation feature map corresponding to the photovoltaic panel area. Step S5: Based on multi-dimensional occlusion features, perform occlusion classification and localization, outputting the occlusion type and the corresponding binary mask of the occlusion region. Occlusion types include vegetation occlusion, dust accumulation occlusion, and terrain collapse occlusion. A multi-task learning framework is used to simultaneously perform occlusion type classification and occlusion region segmentation. The total loss function is... for: ;in Cross-entropy loss for occlusion type classification and They are respectively real and predicted types. and These represent the mask pixel values ​​for the actual and predicted occluded areas, respectively. , This is the loss weighting coefficient.

[0033] Step S6: Calculate the proportion of the shading area to the total area of ​​the photovoltaic panel based on the identified shading type and the binary mask of the shading area. Estimate the power loss by combining the output characteristic curve of the photovoltaic module. Generate a cleaning and scheduling strategy with the goal of minimizing power loss and cleaning cost, and output the corresponding inverter maximum power point tracking adjustment reference signal.

[0034] Example 5: This embodiment is a complete implementation plan for the identification method. Before implementation, a checkerboard calibration board is used to complete the joint calibration of the internal and external parameters of the visible light acquisition device and the infrared acquisition device to establish the pixel-level mapping relationship of the dual-band images. Simultaneously, dual-band image samples of photovoltaic power stations in subsidence areas under different seasons and weather conditions are collected, and the photovoltaic panel area and various shading areas are labeled. A localized training dataset is constructed and the offline training and optimization of the model are completed. The trained model is deployed on the on-site edge computing device.

[0035] During routine operation, visible light and infrared thermal images of the target photovoltaic string are acquired synchronously according to a preset acquisition cycle. The image preprocessing, photovoltaic panel area segmentation, multi-dimensional shading feature extraction, shading classification and area segmentation are completed sequentially according to the established process. Finally, the shading type, area mask and power loss estimation results are output, and corresponding operation and maintenance scheduling and inverter control signals are generated. The entire identification process is completed in closed loop at the field edge, without relying on cloud transmission. It is suitable for the field environment where network coverage is unstable in coal mining subsidence areas. All identification and control records are automatically stored in the local storage unit for subsequent review and model optimization.

[0036] This embodiment further refines the specific implementation of the image preprocessing and feature extraction stages. The image registration stage, based on pre-calibrated mapping parameters, uniformly aligns the infrared thermal image to the coordinate system of the visible light image, ensuring a one-to-one correspondence between pixels in the dual-band images.

[0037] In the denoising process, to address the image particle noise easily caused by sandstorms in the subsidence area, appropriate search window and neighborhood window sizes are set. Noise suppression is achieved through image block similarity weighting while preserving the edge details of the panel. In the illumination correction process, the image is divided into non-overlapping sub-blocks of fixed size. After calculating the local brightness mean of each sub-block, correction parameters are dynamically matched to achieve a balanced effect of brightening dark areas and suppressing bright areas, weakening the illumination interference caused by terrain shadows and backlighting in the morning and evening. In the feature extraction process, the edge contour of the photovoltaic panel area mask is first smoothly sampled, then the discrete curvature is calculated point by point to generate a deformation feature map. Finally, the texture features, thermal difference features, and deformation feature map are stitched together in the channel dimension to form a unified multi-dimensional occlusion feature, which is then sent to the subsequent recognition process.

[0038] This embodiment provides supplementary explanations regarding model training and online inference in the occlusion recognition stage. During model training, data augmentation operations such as rotation, scaling, and brightness perturbation are introduced to expand the sample size based on the localized dataset. A multi-task joint learning framework is used to simultaneously optimize the classification and segmentation branches. By adjusting the loss weight coefficients, the training progress of the two tasks is balanced, enabling the model to possess both accurate type determination and fine-grained region segmentation capabilities. In the online inference stage, morphological opening operations are performed on the initial output binary mask of the occluded region to remove isolated noise pixels, making the mask boundaries more closely match the actual occlusion contours. Simultaneously, corresponding feature verification thresholds are set for the three types of occlusion. For example, when determining terrain collapse occlusion, the amplitude threshold of deformation features is simultaneously matched; when determining dust accumulation occlusion, the distribution uniformity of thermal difference features is referenced. Multi-feature cross-validation reduces the probability of misjudgment, ensuring the reliability of recognition in complex subsidence area scenarios.

[0039] This embodiment focuses on the specific implementation logic of operation and maintenance scheduling and closed-loop optimization. After obtaining the shading type and area mask, the proportion of pixels in the shading area to the effective area pixels of the photovoltaic panel is first calculated. Combined with the component output characteristic curve of the corresponding photovoltaic string and historical power generation data, the instantaneous power loss caused by the current shading is estimated. Then, with the goal of reducing power generation loss and controlling operation and maintenance costs, a priority-based cleaning scheduling plan is generated based on the geographical location, operation and maintenance path distance, and operation difficulty of each shading point. High-loss and easy-to-operate points are prioritized for cleaning work. Simultaneously, the maximum power point tracking adjustment reference signal of the corresponding string is generated and sent to the string inverter to avoid the risk of hot spots caused by shading. During operation, the method can periodically retrieve historical identification results and compare them with the actual power generation data after cleaning, and iteratively correct the relevant parameters of the power estimation model and the loss weight of the identification model, so that the whole method can continuously adapt to the seasonal vegetation changes, dust accumulation patterns, and topographic subsidence process in the subsidence area.

[0040] Table 1:

[0041] Table 2:

[0042] Table 3:

[0043] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A deep learning-based photovoltaic panel shading recognition system for subsidence areas, characterized in that, include: The image acquisition module is used to acquire visible light and infrared thermal images of photovoltaic panels in the subsidence area; The data preprocessing module, connected to the image acquisition module, is used to register, denoise, and correct the illumination of the visible light image and the infrared thermal image to obtain preprocessed image pairs. The illumination correction employs adaptive gamma transform. , Determined by the local brightness of the image; A photovoltaic panel region extraction module, connected to a data preprocessing module, is used to segment out the photovoltaic panel region mask from a preprocessed image pair. The multimodal occlusion feature extraction module is connected to the photovoltaic panel region extraction module. It has a built-in dual-stream convolutional network to extract texture features and thermal difference features from the visible light and infrared image branches respectively, and calculates deformation features based on the edge curvature of the photovoltaic panel region, and fuses them to form multi-dimensional occlusion features. The occlusion recognition and localization module is connected to the multimodal occlusion feature extraction module. It is used to identify the type of occlusion, such as vegetation occlusion, dust accumulation occlusion, or terrain collapse occlusion, based on multi-dimensional occlusion features, and output a binary mask of the occlusion area.

2. The photovoltaic panel shading recognition system based on deep learning in a subsidence area according to claim 1, characterized in that: The data preprocessing module employs non-local mean filtering for noise reduction of pixels in the image. The estimated value Obtained through the following formula: ; in, , In pixels and The image patch centered on This represents the Gaussian-weighted Euclidean distance. These are the filter parameters. As a normalization coefficient, in the adaptive gamma transform, the parameter γ is dynamically adjusted according to the average brightness of the image blocks, and illumination balance is achieved by increasing the γ value in dark areas and decreasing the γ value in bright areas.

3. The photovoltaic panel shading recognition system for subsidence areas based on deep learning according to claim 1, characterized in that: The multimodal occlusion feature extraction module includes a texture branch, a thermal difference branch, and a deformation feature calculation unit. The deformation feature calculation unit extracts the edge contour point set based on the photovoltaic panel area mask. Calculate the discrete curvature at each point The formula for calculating curvature is: ; in , The curvature value is mapped to the image coordinate system to generate a deformation feature map, which is then concatenated with the texture feature and thermal difference feature in the channel dimension to obtain the multi-dimensional occlusion feature.

4. The photovoltaic panel shading recognition system based on deep learning in a subsidence area according to claim 1, characterized in that: The occlusion recognition and localization module includes a region proposal network, a classification branch, and a mask regression branch. It is jointly trained using a multi-task loss function, which is: ; in Cross-entropy loss is used for occlusion type classification. For real labels, To predict probabilities, , is the Dice loss of the binary mask for the occluded region. For real mask pixels, To predict mask pixels, To balance the weights.

5. The photovoltaic panel shading recognition system for subsidence areas based on deep learning according to claim 1, characterized in that: It also includes an occlusion warning and operation and maintenance decision module, which is connected to the occlusion identification and positioning module. It is used to receive the occlusion type and the binary mask of the occlusion area, calculate the occlusion area ratio, and generate cleaning priority instructions and inverter power control instructions based on the preset loss model.

6. A method for identifying photovoltaic panel shading in subsidence areas based on deep learning, and a system for identifying photovoltaic panel shading in subsidence areas based on deep learning according to any one of claims 1 to 5, characterized in that, Includes the following steps: Step S1: Obtain visible light and infrared thermal images of the photovoltaic panels in the subsidence area; Step S2: Register, denoise, and perform adaptive gamma correction on the visible light image and the infrared thermal image to obtain a preprocessed image pair, wherein the adaptive gamma correction parameters are dynamically determined by the local brightness mean of the image. Step S3: Extract the photovoltaic panel region mask from the preprocessed image pairs using a semantic segmentation network; Step S4: Input the preprocessed image and the photovoltaic panel area mask into the dual-flow depth network to extract texture features and thermal difference features respectively. At the same time, calculate the deformation features based on the edge curvature extracted from the photovoltaic panel area mask, and fuse the texture features, thermal difference features and deformation features to generate multi-dimensional occlusion features. Step S5: Classify and locate occlusions based on multi-dimensional occlusion features, and output the occlusion type and the corresponding binary mask of the occlusion area. The occlusion types include vegetation occlusion, dust accumulation occlusion and terrain collapse occlusion.

7. The method for identifying photovoltaic panel shading in subsidence areas based on deep learning according to claim 6, characterized in that: The adaptive gamma correction in step S2 adopts a transform method. ,in The parameter is a constant. The brightness of the local area of ​​the image is dynamically determined to enhance dark areas and suppress bright areas. The noise reduction process uses non-local mean filtering, and the weights are determined by Gaussian weighted similarity of image blocks.

8. The method for identifying photovoltaic panel shading in subsidence areas based on deep learning according to claim 6, characterized in that: The deformation feature extraction process in step S4 is as follows: obtain the photovoltaic panel edge contour point sequence based on the photovoltaic panel area mask, and calculate the discrete curvature for each point on the contour. : ; in , It is a first-order difference. , The second-order difference transforms the curvature values ​​linearly to a preset grayscale range, generating a deformation feature map corresponding to the photovoltaic panel area.

9. A method for identifying photovoltaic panel shading in subsidence areas based on deep learning according to claim 6, characterized in that: In step S5, a multi-task learning framework is used to simultaneously perform occlusion type classification and occlusion region segmentation, with the total loss function... for: ; in Cross-entropy loss for occlusion type classification and They are respectively real and predicted types. and These represent the mask pixel values ​​for the actual and predicted occluded areas, respectively. , This is the loss weighting coefficient.

10. A method for identifying photovoltaic panel shading in subsidence areas based on deep learning according to claim 6, characterized in that: After steps S1 to S5 are performed, step S6 is also included: Based on the identified shading type and the binary mask of the shading area, the proportion of the shading area to the total area of ​​the photovoltaic panel is calculated. Combined with the output characteristic curve of the photovoltaic module, the power loss is estimated. With the goal of minimizing power loss and cleaning cost, a cleaning scheduling strategy is generated, and the corresponding inverter maximum power point tracking adjustment reference signal is output.

Images