Photovoltaic array cleaning strategy recommendation system and method based on feature analysis

Abstract

This application discloses a photovoltaic array cleaning strategy recommendation system and method based on feature analysis. It constructs a registered point cloud by acquiring 3D images of the photovoltaic array and registering them with a digital twin model of the cleaning system. Utilizing the difference in normal height between the 3D point cloud and the digital twin model, it first geometrically segments thick contaminants with physical thickness from thin shadows or flat areas. For the segmented flat areas, it further compares the color difference between the on-site texture and the standard texture of the digital twin to accurately quantify the distribution of micron-level thin dust layers, compensating for the insensitivity of simple geometric measurements to minute dust accumulations. Finally, it fuses the volumetric features of thick contaminants with the texture index of thin contaminants using multimodal hybrid features, and estimates power loss through a pre-trained model. This achieves accurate decoupling and quantitative evaluation of different types of contaminants in complex environments, and generates a cleaning strategy that considers both economic benefits and practicality.

Description

Technical Field

[0001] This application relates to the fields of photovoltaic power generation and operation and maintenance technology, and more specifically, to a photovoltaic array cleaning strategy recommendation system and method based on feature analysis. Background Technology

[0002] As a core component of the renewable energy sector, photovoltaic (PV) power generation's operational efficiency directly impacts the economic benefits of a power plant. However, the accumulation of contaminants on the surface of PV modules reduces light transmittance, leading to a decrease in power generation efficiency. To maintain efficient power plant operation, maintenance personnel need to develop cleaning plans based on accurate contamination status assessments to strike a balance between cleaning costs and power generation gains.

[0003] Currently, pollution monitoring technologies for photovoltaic arrays mainly include manual inspection, reference component comparison, and UAV-based two-dimensional image analysis. Among these, two-dimensional image-based visual analysis technology has attracted widespread attention due to its non-contact and wide coverage characteristics. However, in practical applications, it still faces significant challenges such as high environmental sensitivity and insufficient quantification accuracy. The pixel information of a two-dimensional image is the result of the coupling of multiple physical factors, such as pollution occlusion, lighting conditions, and shooting angle. Conventional algorithms struggle to effectively decouple these factors, often failing to accurately distinguish between shadows cast by clouds or surrounding objects and actual pollution. For example, shadows appear as dark areas in a two-dimensional image, easily misjudged as mud or bird droppings, while glare from strong light reflections can mask actual dust accumulation. This pathological inversion problem based on planar vision leads to frequent false alarms or missed alarms. Furthermore, relying solely on single two-dimensional visual features makes it difficult to establish an accurate mapping relationship from images to electrical power loss. While introducing three-dimensional geometric detection can identify thick dirt such as bird droppings or mud clumps with physical volume through height differences, thus eliminating interference from thin shadows, it is difficult to effectively segment and quantify micron-sized uniform thin layers of dust based solely on geometric shape because their thickness variation is less than the measurement accuracy of conventional depth sensors. Existing technologies cannot simultaneously achieve accurate differentiation and quantitative assessment of thick dirt and thin layers of dust within a unified framework, and therefore cannot provide economical cleaning strategy recommendations based on accurate power loss prediction.

[0004] Therefore, an optimized photovoltaic array cleaning strategy is desired. Summary of the Invention

[0005] To address the aforementioned technical problems, this application is proposed. Embodiments of this application provide a photovoltaic array cleaning strategy recommendation system and method based on feature analysis.

[0006] According to one aspect of this application, a photovoltaic array cleaning strategy recommendation method based on feature analysis is provided, comprising: Obtain the left and right images of the photovoltaic array object; Three-dimensional point cloud reconstruction and digital twin registration are performed on the left and right images of the photovoltaic array object to obtain the registered field point cloud; Based on the height difference threshold, the registered field point cloud is segmented into thick contamination based on 3D geometric difference to obtain thick contamination mask and flat area mask. Based on flat region masking, thin-layer contamination quantization based on 2D texture contrast is performed on the registered field point cloud to obtain thin-layer contamination feature map. A hybrid feature vector is obtained by fusing the features of the thick dirt mask and the thin dirt feature map. Power loss estimation and strategy generation are performed based on hybrid feature vectors to obtain the percentage of power loss due to contamination and recommendations for cleaning actions.

[0007] According to another aspect of this application, a photovoltaic array cleaning strategy recommendation system based on feature analysis is provided, comprising: The image acquisition module is used to acquire the left image and the right image of the photovoltaic array object; The 3D point cloud reconstruction and digital twin registration module is used to perform 3D point cloud reconstruction and digital twin registration on the left and right images of the photovoltaic array object to obtain the registered field point cloud. The heavy contamination segmentation module is used to perform heavy contamination segmentation on the registered field point cloud based on 3D geometric difference, based on the height difference threshold, to obtain heavy contamination mask and flat area mask. The thin-layer contamination quantization module is used to perform thin-layer contamination quantization on the registered field point cloud based on 2D texture contrast, based on flat region mask, to obtain thin-layer contamination feature map. The hybrid feature fusion module is used to perform hybrid feature fusion on the thick dirt mask and the thin dirt feature map to obtain a hybrid feature vector; The power loss estimation and strategy generation module is used to estimate power loss and generate strategies based on hybrid feature vectors to obtain the percentage of power loss due to contamination and recommendations for cleaning actions.

[0008] Compared with existing technologies, to address the technical problems of large environmental interference and low quantification accuracy in photovoltaic array contamination monitoring, this invention utilizes binocular stereo vision and digital twin technology. First, it reconstructs the three-dimensional point cloud of the photovoltaic array and registers it with a digital twin model of an ideal clean state. By calculating the normal distance from the on-site point cloud to the twin model, it uses physical height differences to geometrically segment thick contamination with volume from thin shadows or clean areas, thereby directly eliminating shadow interference and identifying thick contamination. Subsequently, for geometrically flat areas, it achieves precise quantification of thin-layer contamination by comparing the differences between the on-site texture and the reference texture of the twin model in the perceptual color space. Finally, by fusing the volumetric characteristics of thick contamination and the color index of thin-layer contamination, a pre-trained regression model is used to accurately estimate power loss, and combined with economic parameters, an optimal cleaning strategy is dynamically generated, achieving accurate decision-making against interference in complex environments. Attached Figure Description

[0009] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0010] Figure 1 This is a flowchart of a photovoltaic array cleaning strategy recommendation method based on feature analysis according to an embodiment of this application; Figure 2 This is a data flow diagram illustrating the photovoltaic array cleaning strategy recommendation method based on feature analysis according to an embodiment of this application; Figure 3 This is a flowchart illustrating the process of reconstructing and registering three-dimensional point clouds and digital twins of the left and right images of a photovoltaic array object to obtain a registered field point cloud, according to the photovoltaic array cleaning strategy recommendation method based on feature analysis in this application embodiment. Figure 4 This is a flowchart illustrating the process of performing heavy contamination segmentation on a registered field point cloud based on 3D geometric difference to obtain a heavy contamination mask and a flat region mask, according to the photovoltaic array cleaning strategy recommendation method based on feature analysis in the embodiments of this application. Figure 5 This is a flowchart illustrating the process of using a flat region mask to perform thin-layer contamination quantization on a registered field point cloud based on 2D texture comparison to obtain a thin-layer contamination feature map, according to a photovoltaic array cleaning strategy recommendation method based on feature analysis in an embodiment of this application. Figure 6This is a flowchart illustrating the generation of a thin-layer contamination feature map based on a reference flat area texture map and an on-site flat area texture map, according to a photovoltaic array cleaning strategy recommendation method based on feature analysis according to embodiments of this application. Figure 7 This is a block diagram of a photovoltaic array cleaning strategy recommendation system based on feature analysis according to an embodiment of this application. Detailed Implementation

[0011] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0012] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" are not specifically singular and may include plural forms. Generally speaking, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0013] While this application makes various references to certain modules of the systems according to embodiments of this application, any number of different modules can be used and run on user terminals and / or servers. The modules described are merely illustrative, and different aspects of the systems and methods may use different modules.

[0014] Flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, various steps can be processed in reverse order or simultaneously as needed. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0015] Current photovoltaic array visual inspection technologies are severely affected by environmental factors such as lighting, shadows, and shooting angles, making it difficult to accurately distinguish between dirt with physical volume and thin shadows. Furthermore, existing methods struggle to accurately quantify both thick dirt and thin dust layers within a unified framework, resulting in insufficient accuracy in the evaluation model from visual features to power loss. Therefore, this application proposes a photovoltaic array cleaning strategy recommendation method based on feature analysis. By integrating three-dimensional geometric reconstruction from stereo vision with two-dimensional texture analysis using digital twin technology, it achieves accurate and interference-resistant assessment and decision-making for dirt of different forms. Specifically, the scheme first reconstructs 3D point clouds from the collected 3D images of the photovoltaic array and registers them with a high-precision digital twin model of an ideal clean state. Then, using the difference in normal distance from the point cloud to the model surface, the on-site point cloud is segmented into heavily soiled areas and flat areas based on a physical height threshold. This allows for the direct removal of shadow interference and the identification of significant volumetric dirt such as bird droppings and mud clumps using 3D geometric information. For geometrically determined flat areas, the texture is further compared pixel-by-pixel with the reference clean texture of the digital twin model to quantify the distribution of micron-level thin dust layers. Finally, the extracted volumetric features of the heavily soiled areas are fused with the texture features of the thin soiled areas and input into a pre-trained regression model to accurately estimate power loss. The optimal cleaning recommendations are then dynamically generated based on economic parameters.

[0016] Figure 1 This is a flowchart of a photovoltaic array cleaning strategy recommendation method based on feature analysis according to an embodiment of this application. Figure 2 This is a schematic diagram of the data flow for a photovoltaic array cleaning strategy recommendation method based on feature analysis according to an embodiment of this application. Figure 1 and Figure 2 As shown, the photovoltaic array cleaning strategy recommendation method based on feature analysis according to an embodiment of this application includes the following steps: S100, acquiring a left image and a right image of the photovoltaic array object; S200, performing 3D point cloud reconstruction and digital twin registration on the left and right images of the photovoltaic array object to obtain a registered field point cloud; S300, performing heavy contamination segmentation based on 3D geometric difference on the registered field point cloud based on a height difference threshold to obtain a heavy contamination mask and a flat area mask; S400, performing thin-layer contamination quantization based on 2D texture contrast on the registered field point cloud based on the flat area mask to obtain a thin-layer contamination feature map; S500, performing hybrid feature fusion on the heavy contamination mask and the thin-layer contamination feature map to obtain a hybrid feature vector; S600, performing power loss estimation and strategy generation based on the hybrid feature vector to obtain the contamination power loss percentage and cleaning action recommendations.

[0017] Specifically, in step S100, the left and right images of the photovoltaic array object are acquired. It is understood that a single two-dimensional image faces the challenge of poorly distinguishing between thick, physically voluminous dirt and thin shadows or flat, color-difference regions in photovoltaic array contamination detection. Simply relying on pixel grayscale information cannot effectively decouple the complex superposition effects of illumination, angle, and material reflection properties. Therefore, in the technical solution of this application, the left and right images of the photovoltaic array object are acquired to construct a stereoscopic vision data foundation containing parallax information, providing necessary raw observation data for subsequent reconstruction of the scene's three-dimensional geometry using triangulation principles. This allows for the introduction of depth dimension information from the source, overcoming the physical limitations of two-dimensional vision in terms of environmental interference resistance, thereby ensuring reliable spatial data support for subsequent geometric segmentation and quantitative assessment of contamination types.

[0018] More specifically, in a particular example of this application, this step is implemented using a binocular stereo vision acquisition device deployed on a mobile platform such as a drone or ground inspection robot. This device comprises two precisely calibrated camera units with a fixed baseline distance and parallel optical axes. During photovoltaic array inspection, the mobile platform carries the acquisition device along a preset path across the photovoltaic array area. When reaching a designated detection point or triggered according to a preset acquisition frequency, the synchronization controller inside the acquisition device simultaneously sends exposure commands to both the left and right camera units, ensuring that both camera units open their shutters at the same time to capture optical images of the photovoltaic module surface, thereby eliminating motion blur or matching errors caused by shooting time differences. The visual information acquired by the left camera unit is converted from analog to digital and processed into a left image of the photovoltaic array object, and the visual information acquired by the right camera unit is similarly used to generate a right image of the photovoltaic array object. These two images together constitute a set of synchronized stereo image pairs containing information from different perspectives of the same scene, and are transmitted or stored in real time as input data for subsequent 3D reconstruction and feature analysis.

[0019] Specifically, in step S200, three-dimensional point cloud reconstruction and digital twin registration are performed on the left and right images of the photovoltaic array object to obtain a registered on-site point cloud. It is understandable that although the original stereo images contain parallax information, they are still in two-dimensional pixel space, lacking an intuitive three-dimensional geometric description. Furthermore, the camera pose during on-site acquisition is random, making it impossible to directly compare the acquired data point-to-point with a standard cleanliness benchmark, and making it difficult to physically distinguish between light and shadow changes and physical accumulation. Therefore, in the technical solution of this application, three-dimensional point cloud reconstruction and digital twin registration are further performed on the left and right images of the photovoltaic array object to obtain a registered on-site point cloud. This transforms the planar parallax data into a three-dimensional spatial point set with physical scale, and through rigid transformation, precisely aligns this on-site point set to the unified coordinate system of the ideal clean panel digital twin model. In this way, a geometric correspondence can be established between the field measurement data and the standard reference model, laying a precise spatial registration foundation for subsequent precise removal of shadow interference through three-dimensional geometric difference, identification of thick dirt, and quantification of thin dust based on texture contrast.

[0020] Figure 3 This is a flowchart illustrating the process of reconstructing and registering 3D point clouds and digital twins of the left and right images of a photovoltaic array object to obtain a registered on-site point cloud, according to the feature analysis-based photovoltaic array cleaning strategy recommendation method of this application. Figure 3 As shown, step S200 includes: S210, performing semi-global matching on the left image and the right image of the photovoltaic array object to obtain a disparity map; S220, performing triangulation-based 3D point cloud reconstruction on the disparity map to obtain the original 3D point cloud; S230, performing ICP (Iterative ClosestPoint)-based registration of the original 3D point cloud with the digital twin to obtain the registered on-site point cloud.

[0021] In step S210, a semi-global matching is performed on the left and right images of the photovoltaic array object to obtain a disparity map. It is understood that simply obtaining two-dimensional images from the left and right perspectives is insufficient to directly deduce the scene's depth information. It is necessary to establish pixel-level correspondences between the left and right images to resolve the positional deviation of the same physical point under different perspectives, i.e., disparity. Furthermore, the surface of photovoltaic modules often has reflective or weakly textured areas, making traditional local matching algorithms prone to mismatches. Therefore, in the technical solution of this application, a semi-global matching is further performed on the left and right images of the photovoltaic array object to obtain a disparity map. This utilizes the semi-global matching algorithm to introduce multi-directional path aggregation constraints based on pixel-level matching cost calculations, achieving near-global energy minimization optimization while ensuring computational efficiency. This effectively overcomes matching ambiguities caused by uneven illumination or monotonous texture on the photovoltaic module surface, generating dense, accurate, and clearly edge-defined disparity data, providing highly reliable intermediate layer data support for the subsequent conversion of two-dimensional images into three-dimensional spatial geometric structures.

[0022] More specifically, in a specific example of this application, this step first receives a left and a right image after epipolar correction processing, ensuring that the corresponding points to be matched are located on the same horizontal scan line, thereby simplifying the two-dimensional search to a one-dimensional search. Next, for each pixel in the left image, the matching cost is calculated within the corresponding epipolar range in the right image. This process uses a Census transform or cross-correlation algorithm to construct a three-dimensional cost space to quantify the similarity between pixels in the left and right images and enhance robustness to illumination changes. Subsequently, the core cost aggregation operation is performed. This operation performs dynamic planning of a one-dimensional path along multiple directions (e.g., 8 or 16 directions) of the image plane, accumulating the matching cost on each path, and smoothing the disparity field by applying a penalty term to disparity abrupt changes in adjacent pixels, thereby suppressing noise while preserving component edge details. After completing multi-directional cost aggregation, a winner-takes-all strategy is adopted, selecting the disparity value corresponding to the minimum aggregated cost for each pixel as the initial disparity for that point. Preferably, the smoothing term for path aggregation uses a P1 / P2 penalty (P1... P2), and after left-right consistency checks, guided filtering / bilateral filtering is used for hole filling and sub-pixel parabolic fitting refinement. Finally, disparity refinement processing is performed on the initial disparity map, which includes left-right consistency checks to remove invalid disparities in occluded areas and sub-pixel interpolation to improve quantization accuracy, ultimately outputting a dense disparity map that fully reflects the depth hierarchy of the photovoltaic array surface.

[0023] In step S220, the disparity map is reconstructed into a 3D point cloud based on triangulation to obtain the original 3D point cloud. It is understood that since the disparity map only represents the pixel-level relative offset between the left and right views, it lacks absolute spatial scale information that can directly measure the physical thickness of dirt on the photovoltaic module surface, and cannot be directly used to distinguish between bird droppings accumulations with millimeter-level height features and shadow areas with only optical differences. Therefore, in the technical solution of this application, the disparity map is further reconstructed into a 3D point cloud based on triangulation to obtain the original 3D point cloud. This is then combined with camera intrinsic parameters and binocular baseline parameters to map the disparity data on the 2D image plane into 3D spatial coordinates in the camera coordinate system. This endows the visual data with true physical measurement attributes, generating a spatial point set containing accurate depth information, providing a quantitative basis for subsequent dirt segmentation based on geometric height differences.

[0024] More specifically, in a concrete example of this application, this step first retrieves the calibration parameters of the binocular camera, including the camera's focal length, principal point coordinates, and baseline distance between the optical centers of the left and right cameras. Then, it iterates through each valid pixel in the disparity map, performing spatial coordinate calculation based on the principle of triangulation. For any pixel, the depth value in the camera coordinate system is calculated by dividing the product of its disparity value and the baseline distance by the focal length; this depth value is inversely proportional to the disparity. After determining the depth value, the spatial coordinates of the pixel in the horizontal and vertical directions are derived using the inverse transformation formula of perspective projection, combining the pixel's horizontal and vertical coordinates on the image plane with the camera's principal point coordinates. During this process, the RGB color information of the corresponding pixel position in the left or right image is simultaneously indexed and mapped onto the calculated three-dimensional spatial point. Finally, all the calculated three-dimensional spatial points with color attributes are aggregated to form an original three-dimensional point cloud data structure that truly reflects the geometric topography of the photovoltaic array surface.

[0025] In step S230, the original 3D point cloud is registered with the digital twin based on ICP to obtain a registered on-site point cloud. It is understood that since the original 3D point cloud data obtained through visual reconstruction is defined based on the camera's local coordinate system, while the digital twin model used as the cleaning reference is in a fixed global model coordinate system, there are positional and orientation deviations between the two. Direct geometric comparison cannot accurately reflect the true physical changes on the photovoltaic module surface. Therefore, in the technical solution of this application, the original 3D point cloud is further registered with the digital twin based on ICP to obtain a registered on-site point cloud. This allows for the calculation and application of the optimal rigid transformation matrix, accurately mapping and aligning the on-site collected point cloud data to the unified coordinate space of the digital twin model. This eliminates spatial misalignment caused by differences in shooting angle and position, ensuring that subsequent geometric difference calculations only reflect physical height abrupt changes caused by dirt accumulation, rather than systematic errors caused by coordinate system misalignment.

[0026] More specifically, in a concrete example of this application, this step employs an iterative nearest neighbor algorithm to find the optimal matching relationship between the source point cloud and the target model. First, a spatial index structure such as a K-Dimensional Tree (K-D Tree) is constructed for the point cloud data of the digital twin model to accelerate the search efficiency of nearest neighbors. Then, an iterative optimization process is initiated. In each iteration, every data point in the original 3D point cloud is traversed, and the point with the closest Euclidean distance in the digital twin model is searched as the corresponding point, thus establishing a set of point pairs. Based on this set of point pairs, the rotation matrix and translation vector that minimize the mean square error between point pairs are solved using singular value decomposition or least squares. Next, the spatial coordinates of the original 3D point cloud are updated using the calculated transformation parameters, and new corresponding point pairs and error values ​​are recalculated. This process of finding corresponding points, solving transformation parameters, and updating coordinates is repeated until the error change between two adjacent iterations is less than a preset convergence threshold or the maximum number of iterations is reached. Preferably, point-to-plane ICP is employed, and an M-estimation loss such as Huber or Cauchy is introduced to suppress the influence of outliers caused by reflection. Furthermore, coarse registration initialization using FPFH (Fast Point Feature Histograms) / SIFT (Scale Invariant Feature Transform) features is performed before ICP to improve convergence stability. The final output is a registered field point cloud that spatially coincides precisely with the digital twin model.

[0027] Specifically, in step S300, based on a height difference threshold, the registered field point cloud is segmented into thick dirt based on 3D geometric difference to obtain a thick dirt mask and a flat area mask. It is understandable that, since visual color or grayscale information alone is insufficient to effectively distinguish between physically sized, accumulated dirt (such as bird droppings or mud) and thin-film interference that only causes optical brightness attenuation (such as cloud shadows or tree occlusion), the two are easily confused in two-dimensional images, leading to misjudgments in cleaning decisions. Therefore, in the technical solution of this application, the registered field point cloud is further segmented into thick dirt based on 3D geometric difference using a height difference threshold to obtain a thick dirt mask and a flat area mask. This utilizes the normal height deviation between the measured field data and the surface of the ideal digital twin model as a physical discrimination criterion, identifying protruding areas exceeding the physical tolerance range as thick dirt. In this way, interference from two-dimensional visual noise such as shadows can be directly eliminated from the geometric and physical level, accurately locking heavily soiled areas with physical thickness, and defining clear spatial boundaries for subsequent textural quantification of fine dust in the remaining flat areas, thus achieving precise layered control of soil morphology.

[0028] Figure 4 This is a flowchart illustrating the process of segmenting a registered field point cloud based on 3D geometric difference to obtain a heavy contamination mask and a flat region mask, using a height difference threshold-based method for recommending photovoltaic array cleaning strategies based on feature analysis, according to embodiments of this application. Figure 4 As shown, step S300 includes: S310, calculating the point-to-model normal distance of the registered field point cloud based on the digital twin model of the cleaning panel to obtain a normal distance array; S320, performing point cloud segmentation based on the height threshold on the registered field point cloud based on the normal distance array and the height difference threshold to obtain a heavy dirt mask and a flat area mask.

[0029] In step S310, based on the digital twin model of the cleaning panel, the point-to-model normal distance of the registered field point cloud is calculated to obtain a normal distance array. It is understood that although the registered field point cloud is spatially aligned with the digital twin model, the specific physical deviation of each independent data point in the point cloud relative to the ideal clean surface has not been quantified, and the thickness of the accumulated dirt cannot be directly represented by the three-dimensional coordinates alone. Therefore, in the technical solution of this application, the point-to-model normal distance of the registered field point cloud is further calculated based on the digital twin model of the cleaning panel to obtain a normal distance array. This accurately measures the vertical height difference of each sampling point relative to the ideal reference surface, converting the three-dimensional spatial position difference into a thickness metric with clear physical meaning. This provides a unique and accurate geometric criterion for distinguishing between flat areas adhering to the panel surface and thick dirt protruding from the panel surface.

[0030] More specifically, in a concrete example of this application, this step first preprocesses the digital twin model of the cleaning panel, constructs a KD-tree spatial index structure for the reference point cloud in the model, and pre-calculates and stores the unit normal vectors of each reference position on the model surface. During calculation, each data point in the registered field point cloud is traversed sequentially, and the reference point with the closest Euclidean distance to the field data point is retrieved from the reference point cloud of the digital twin model using the KD-tree index. Subsequently, a displacement vector is constructed from the reference point to the field data point, and this displacement vector is multiplied by the unit normal vector at the reference point to calculate the vertical projection distance of the field data point relative to the ideal cleaning surface. This calculation process is performed on all points in the field point cloud one by one, and finally, all the calculated distance values ​​are integrated into a normal distance array according to the index order of the points, thus completing the transformation from spatial coordinates to physical thickness features.

[0031] In step S320, based on the normal distance array and the height difference threshold, the registered field point cloud is segmented using a height threshold to obtain a heavy dirt mask and a flat area mask. It is understood that while the normal distance array quantifies the physical height deviation of each point relative to the ideal surface, it does not yet form a qualitative judgment of the dirt type. Furthermore, a clear physical boundary must be used to distinguish between severely obstructing deposits and surface attachments or shadows that only affect optical properties; shadows do not geometrically possess thickness characteristics. Therefore, in the technical solution of this application, the registered field point cloud is further segmented using a height threshold based on the normal distance array and the height difference threshold to obtain a heavy dirt mask and a flat area mask. This introduces a physical thickness determination criterion, discretizing continuous distance values ​​into mutually exclusive region categories. In this way, based on geometric morphological characteristics, it is possible to physically decouple bird droppings and mud clumps with solid thickness from shadows and micron-sized dust particles with no thickness, completely eliminating the interference of shadows on the determination of heavy pollution, and providing an accurate spatial index for subsequent differentiated analysis strategies for different areas.

[0032] More specifically, in a concrete example of this application, the implementation of this step first sets a height difference threshold, which is selected based on the manufacturing tolerance of the photovoltaic module and the minimum physical thickness of the accumulated contaminants that require special attention, serving as a critical value to distinguish between three-dimensional protrusions and geometric planes. Then, two empty collection containers for storing point indices are initialized, corresponding to a heavy contaminant mask and a flat area mask, respectively. Next, a linear traversal process of the normal distance array is initiated, reading the normal distance value of each sampling point one by one, and logically comparing this value with the preset height difference threshold. During the comparison, if the normal distance value of a point is found to be strictly greater than the height difference threshold, it indicates that there is a physical protrusion exceeding the plane tolerance range at the spatial location corresponding to that point; therefore, the index of that point is marked and stored in the heavy contaminant mask. Conversely, if the normal distance value of that point is less than or equal to the height difference threshold, it indicates that the point is geometrically close to the module surface and belongs to a clean surface, shaded area, or thin dust-covered area; therefore, the index of that point is assigned to the flat area mask. After all data points have been traversed and judged, these two mask sets are output, thus realizing the binarized geometric segmentation of complex scenes on the photovoltaic array surface. Preferably, three-dimensional morphological closing operations and voxel downsampling denoising are performed on the thick dirt mask, and connected component analysis is performed to eliminate false detections less than a set voxel number threshold, thereby improving the robustness of subsequent total volume estimation.

[0033] Specifically, in step S400, based on a flat region mask, the registered field point cloud is subjected to thin-layer contamination quantization based on 2D texture contrast to obtain a thin-layer contamination feature map. It is understood that because the uniform thin-layer dust at the micrometer level is far below the depth resolution limit of a stereo vision system in terms of physical thickness, it cannot be effectively identified through the aforementioned geometric height threshold segmentation. However, such contamination significantly alters the spectral reflectance characteristics and color texture of the component surface, and relying solely on a single field image makes it difficult to quantify its specific attenuation relative to an ideal clean state in the absence of a reference. Therefore, in the technical solution of this application, the registered field point cloud is further subjected to thin-layer contamination quantization based on 2D texture contrast using a flat region mask to obtain a thin-layer contamination feature map. This focuses on geometrically determined flat areas, utilizing the supplementary role of optical texture information to physical morphology information, and analyzing the distribution density of thin-layer contamination by calculating the difference in texture between the field visual data and the digital twin reference. This effectively compensates for the blind spots of three-dimensional geometric analysis in the detection of tiny particles, enabling highly sensitive and accurate quantification of thin-layer ash accumulation without significant volume characteristics, thereby constructing a comprehensive pollution assessment system covering both heavy accumulation and thin-layer adhesion.

[0034] Figure 5This is a flowchart illustrating the process of using a flat region mask to perform thin-layer contamination quantization on a registered field point cloud based on 2D texture contrast to obtain a thin-layer contamination feature map, according to a photovoltaic array cleaning strategy recommendation method based on feature analysis in an embodiment of this application. Figure 5 As shown, step S400 includes: S410, extracting a flat area texture map from the registered field point cloud based on a flat area mask; S420, extracting a reference flat area texture map from the digital twin model of the cleaning panel based on a flat area mask; S430, generating a thin-layer dirt feature map based on the reference flat area texture map and the field flat area texture map.

[0035] In step S410, a texture map of the flat area is extracted from the registered point cloud based on a flat area mask. It is understood that while the registered point cloud contains color information, it is essentially a discrete set of three-dimensional spatial data. The analysis of fine features for thin-layer contamination relies on a continuous and regularly arranged two-dimensional image pixel array, and color abrupt changes in thick contamination areas can interfere with the overall evaluation of thin-layer dust accumulation in flat areas. Therefore, in this application's technical solution, a texture map of the flat area is further extracted from the registered point cloud based on a flat area mask. This allows for spatial filtering of the point cloud data using a mask, retaining only the color attributes of geometrically flat areas and mapping them back from three-dimensional space to a two-dimensional image plane. This generates a texture image containing only the surface details of the area to be detected, providing standardized input data for subsequent pixel-by-pixel precise alignment and difference calculation with the reference texture of the digital twin model.

[0036] More specifically, in a concrete example of this application, this step first initializes a blank two-dimensional image buffer based on the surface dimensions of the digital twin model and a preset texture resolution. This buffer has the same pixel dimensions and number of channels as the reference texture. Then, a traversal program is initiated for the set of indices in the flat region mask, extracting the three-dimensional spatial coordinates and RGB color components of the registered field point cloud data points corresponding to each index. For each extracted data point, according to a pre-determined projection mapping relationship, the three-dimensional coordinates of the point in a unified coordinate system are projected and transformed into pixel coordinates on the two-dimensional image plane. This projection transformation maintains the topological consistency of the spatial relative positions. Next, the RGB color value carried by the data point is assigned to the corresponding pixel position in the two-dimensional image buffer. If multiple points are projected to the same pixel, a weighted average or nearest neighbor strategy is used to determine the final pixel value. After all data points belonging to the flat region mask have been processed, the filled two-dimensional image buffer is output as the field flat region texture map, thereby achieving dimensionality reduction transformation from three-dimensional unstructured data to two-dimensional structured image data.

[0037] In step S420, a reference flat area texture map is extracted from the digital twin model of the cleaning panel based on a flat area mask. It is understood that accurate quantification of thin-layer contamination relies on the difference analysis between on-site measured data and the ideal cleaning state. Although the digital twin model stores full-size cleaning texture information, to ensure the effectiveness of the difference calculation, areas identified as heavily contaminated must be removed, and a point-to-point comparison benchmark must be established only within geometrically flat areas. Therefore, in the technical solution of this application, a reference flat area texture map is further extracted from the digital twin model of the cleaning panel based on a flat area mask. This utilizes the spatial gating properties of the mask to accurately filter and reconstruct an ideal texture image that is strictly consistent with the spatial distribution of the on-site flat area from the full-domain texture data of the digital twin model. This ensures accurate spatial alignment between the reference texture and the on-site texture in the pixel coordinate system, eliminates interference from non-interesting regions, and provides a standard reference for subsequent calculations of optical attenuation caused purely by thin-layer contamination.

[0038] More specifically, in a concrete example of this application, this step first loads the high dynamic range texture map and texture coordinate mapping table associated with the digital twin model of the cleaning panel. This mapping table defines the geometric correspondence between the three-dimensional surface of the model and the two-dimensional texture map. Then, a parsing process for the flat region mask is initiated, traversing each spatial point index marked as a flat region in the mask set. For each index point, based on its three-dimensional geometric position in the digital twin model, the texture coordinate mapping table is queried to obtain its UV coordinates (two-dimensional texture mapping coordinates) on the source texture map, and the corresponding standard RGB color value is sampled accordingly. Next, the projection transformation rules, identical to those used in the on-site flat region texture map generation process, are applied to fill the corresponding pixel positions in the newly created target image buffer with the sampled standard color values. After completing the texture sampling and reprojection operations for all mask points, the reference flat region texture map is output, thereby restoring the visual representation of the test area under ideal cleaning conditions on the same two-dimensional image plane.

[0039] In step S430, a thin-layer contamination feature map is generated based on the reference flat area texture map and the on-site flat area texture map. It is understood that although micron-sized thin dust layers cannot trigger three-dimensional threshold segmentation in terms of geometric height, their coverage on the photovoltaic module surface causes light absorption and scattering effects, resulting in a significant deviation of the on-site texture from the ideal cleanliness state in terms of color depth and contrast. This deviation is positively correlated with the contamination density. Therefore, in the technical solution of this application, a thin-layer contamination feature map is further generated based on the reference flat area texture map and the on-site flat area texture map to establish a pixel-by-pixel comparative analysis mechanism, converting the optical differences between the on-site image and the digital twin benchmark into a quantified contamination severity value. This outputs a two-dimensional feature map that intuitively reflects the distribution of thin-layer dust accumulation at various points within the flat area, providing high-precision, fine-grained data support for subsequent assessment of transmittance loss.

[0040] Figure 6 This is a flowchart illustrating the process of using a flat region mask to perform thin-layer contamination quantization on a registered field point cloud based on 2D texture contrast to obtain a thin-layer contamination feature map, according to a photovoltaic array cleaning strategy recommendation method based on feature analysis in an embodiment of this application. Figure 6 As shown, step S430 includes: S431, performing pixel-by-pixel color difference calculation on the reference flat area texture map and the on-site flat area texture map to obtain a color difference matrix; S432, generating a thin-layer dirt feature map based on the color difference matrix and the flat area mask.

[0041] In step S431, pixel-by-pixel color difference calculations are performed on the reference flat area texture map and the on-site flat area texture map to obtain a color difference matrix. It is understood that thin layers of dust adhering to the surface of photovoltaic modules cause non-uniform changes in their spectral reflectance. These subtle changes in optical properties are difficult to capture through macroscopic geometric measurements, and the degree of dirt accumulation in different areas often varies. Visual observation or simple global threshold determination cannot accurately quantify the transmittance attenuation of each tiny area. Therefore, in the technical solution of this application, pixel-by-pixel color difference calculations are further performed on the reference flat area texture map and the on-site flat area texture map to obtain a color difference matrix. This transforms the visually intuitive differences into mathematical quantitative deviations, establishing a numerical comparison relationship between the measured color value and the ideal clean color value for each pixel in the image. This generates a high-resolution difference data matrix that accurately maps the severity of dirt at any location within the flat area, providing a fine-grained data foundation for subsequently constructing feature maps reflecting the details of dirt distribution and accurately estimating power loss caused by spectral absorption.

[0042] More specifically, in a specific example of this application, pixel-by-pixel color difference calculation is performed on the reference flat area texture map and the on-site flat area texture map to obtain a color difference matrix, including: performing color space conversion on the reference flat area texture map and the on-site flat area texture map to obtain a CIELAB color space reference flat area texture map and a CIELAB color space on-site flat area texture map; calculating the CIEDE2000 color difference between the CIELAB color space reference flat area texture map and the CIELAB color space on-site flat area texture map to obtain a color difference matrix.

[0043] Accordingly, the reference flat area texture map and the on-site flat area texture map are color space converted to obtain the CIELAB color space reference flat area texture map and the CIELAB color space on-site flat area texture map. It is understandable that, since the original texture image typically uses the device-dependent, non-uniform RGB color space, its geometric distance cannot linearly reflect the differences in human eye's color perception, and there is a high correlation between the three RGB channels, making it easy for changes in light intensity to be mixed into color information determination, making it difficult to accurately quantify the minute changes in spectral reflectance caused by a thin layer of dust. Therefore, in the technical solution of this application, the reference flat area texture map and the on-site flat area texture map are further color space converted to obtain the CIELAB color space reference flat area texture map and the CIELAB color space on-site flat area texture map, thereby converting the image data to the device-independent and perceptually more uniform CIELAB color space. This effectively decouples brightness and chromaticity information, utilizing... Components characterize brightness. and The component characterizes chromaticity, thus providing an accurate measurement benchmark that conforms to visual psychophysics for subsequent calculations of minute color differences based on the characteristics of human eye perception.

[0044] More specifically, in a concrete example of this application, this step first iterates through each valid pixel in the reference flat area texture map and the field flat area texture map, reading its normalized R, G, and B component values. Next, an inverse Gamma correction operation is performed on the R, G, and B components, converting them from the non-linear sRGB space to the linear RGB space, eliminating the influence of the non-linear photoelectric conversion characteristics of the display device. Subsequently, using a standard linear transformation matrix, the linear RGB values ​​are mapped to the CIE XYZ tristimulus value space, which serves as a bridge connecting the physical spectrum and color perception. Based on this, the standard illuminant D65 is selected as the reference white point, and according to the definition formula of the CIELAB color model, the XYZ tristimulus values ​​are converted to... , , Coordinates. Specifically, brightness is calculated by transforming the normalized Y value using a non-linear function. The red-green axis chromaticity components are calculated using the difference between the nonlinear transformations of X and Y. The yellow-blue axis chromaticity components are calculated using the difference between the nonlinear transformations of Y and Z. Finally, the calculated... , , The values ​​are stored in the corresponding output image channels to generate CIELAB color space reference flat area texture map and CIELAB color space field flat area texture map, which can accurately reflect the color perception characteristics of the photovoltaic module surface.

[0045] Accordingly, the CIEDE2000 color difference between the CIELAB color space reference flat area texture map and the CIELAB color space field flat area texture map is calculated to obtain a color difference matrix. It is understandable that although the CIELAB color space is perceptually relatively uniform, its simple Euclidean distance metric still exhibits perceptual non-uniformity when dealing with the near-neutral color variations caused by the dark blue background and dust commonly found in photovoltaic modules. This fails to accurately reflect the subtle spectral attenuation caused by micron-level thin-layer contamination, resulting in insufficient quantification sensitivity for early, slight dust accumulation. Therefore, in the technical solution of this application, pixel-by-pixel color difference calculations are further performed on the reference flat area texture map and the field flat area texture map to obtain a color difference matrix. This allows the introduction of the CIEDE2000 color difference formula, which includes brightness, chroma, hue weighting functions, and rotation terms, to perform refined, high-order correction calculations on the differences between the two images in the perceptual space. This eliminates the nonlinear error of human visual characteristics in different color regions, and transforms the influence of thin-layer contamination on the optical properties of the component surface into a precise numerical index that is highly positively correlated with the degree of physical shading, thereby significantly improving the resolution accuracy of low-density contamination distribution.

[0046] More specifically, in a concrete example of this application, this step first establishes a floating-point matrix container with the same size as the flat area texture map to store the calculation results. Then, a double loop traversal of the image pixel array is initiated. For each pair of corresponding pixel positions, its lightness, red-green axis chroma, and yellow-blue axis chroma components in the CIELAB space are obtained. For the extracted component values, the average chroma is first calculated, and a non-linear adjustment factor is generated accordingly. The red-green axis chroma is compressed and corrected to improve the calculation performance in the neutral gray area, thereby deriving the corrected chroma value and hue angle. Based on this, the differences between two pixels in lightness, corrected chroma, and corrected hue are calculated, and the arithmetic mean of each component is calculated to determine its specific position in the color space. Next, based on these position parameters, lightness weight, chroma weight, and hue weight are calculated using specific weighting functions. These weighting functions are used to adjust the differences in visual sensitivity under different brightness and saturation levels. In particular, for the blue area unique to photovoltaic panels, a rotation term coefficient is calculated to correct the interaction between hue and chroma. Finally, substitute all the above differences, weighting factors, and parameter factors into the CIEDE2000 color difference formula to perform weighted sum of squares and square root operations to obtain the final color difference value of the pixel and fill it into the matrix container. This process continues until all pixels have been processed, and a complete color difference matrix is ​​output.

[0047] In step S432, a thin-layer contamination feature map is generated based on the color difference matrix and the flat region mask. It is understood that while the color difference matrix quantifies the degree of optical deviation per pixel, the calculation range it covers may include non-photovoltaic module areas or areas already identified as heavily contaminated. Directly using the original full-image difference data for subsequent evaluation would lead to repeated inclusion of color abrupt changes in heavily contaminated areas in the thin-layer contamination feature map, causing index coupling and decreased accuracy. Therefore, in the technical solution of this application, a thin-layer contamination feature map is further generated based on the color difference matrix and the flat region mask. This utilizes the spatial gating effect of the flat region mask to strictly limit the effective data in the color difference matrix to geometrically flat areas and filter out noise interference from all non-flat areas. This allows the construction of a dedicated feature map that only reflects the distribution of micron-level thin-layer dust, ensuring that the thin-layer contamination features and the heavily contamination features maintain strict orthogonality and independence in spatial and physical properties during the subsequent feature fusion process.

[0048] More specifically, in a concrete example of this application, this step first initializes a single-channel grayscale image or floating-point matrix with the same resolution and size as the flat area texture map as a thin-layer dirt feature map, and sets the initial value of all pixels to zero. Then, a traversal process for the color difference matrix is ​​initiated, or parallel processing is performed using a flat area mask as an index template. For each pixel coordinate in the image array, it is checked whether it is included in the index set of the flat area mask. If the pixel coordinate belongs to a flat area, the corresponding CIEDE2000 color difference value is extracted from the color difference matrix and directly assigned to the corresponding position in the thin-layer dirt feature map; if the pixel coordinate does not belong to a flat area (i.e., belongs to a thick dirt area or background area), the value at that position in the thin-layer dirt feature map is kept zero, thereby realizing the spatial masking operation. Based on this, in order to adapt to the input data specifications of the subsequent regression model, the extracted color difference values ​​can be normalized, and the physical color difference values ​​can be linearly mapped to the preset dimensionless feature range. Finally, a digital feature map that intuitively presents the spatial distribution of the severity of dust accumulation on the surface of photovoltaic modules can be output.

[0049] Specifically, in step S500, the thick contamination mask and the thin contamination feature map are fused to obtain a hybrid feature vector. It is understood that the impact of thick contamination on photovoltaic modules is mainly manifested as current mismatch caused by physical shading, while the impact of thin contamination is manifested as global transmittance attenuation caused by spectral absorption. The physical damage mechanisms of these two are significantly different and are stored in discrete mask sets and continuous feature maps, respectively, lacking a unified mathematical representation to support comprehensive power loss assessment. Therefore, in the technical solution of this application, the thick contamination mask and the thin contamination feature map are further fused to obtain a hybrid feature vector. This is used to extract and integrate volume features representing physical accumulation, texture index features representing optical attenuation, and area features representing the overall shading range, reducing the dimensionality of multi-source heterogeneous sensing data and standardizing it into a compact numerical vector. This provides an input interface containing all physical information for subsequent nonlinear regression models, ensuring that power loss estimation can simultaneously consider local shading effects and global radiation attenuation effects, thereby achieving a comprehensive and accurate quantification of the overall health status of the photovoltaic array.

[0050] More specifically, in a specific example of this application, a hybrid feature fusion is performed on a thick filth mask and a thin filth feature map to obtain a hybrid feature vector, including: extracting the total volume of thick filth from the thick filth mask; extracting the average index of thin filth from the thin filth feature map; calculating the total filth coverage area based on the thick filth mask and the thin filth feature map; and vectorizing the total volume of thick filth, the average index of thin filth, and the total filth coverage area to obtain the hybrid feature vector.

[0051] Accordingly, the total volume of heavy contamination is extracted from the heavy contamination mask, and the average index of thin-layer contamination is extracted from the thin-layer contamination feature map. It is understandable that while the heavy contamination mask and thin-layer contamination feature map accurately identify the distribution location and morphology of contamination in space, they are essentially high-dimensional spatial distribution data and cannot be directly input as scalar parameters into a power loss model based on numerical regression. Furthermore, the impact of heavy accumulation on power generation performance mainly depends on the physical volume of light it blocks, while the impact of thin-layer dust depends on its overall spectral attenuation. Therefore, in the technical solution of this application, the total volume of heavy contamination is further extracted from the heavy contamination mask, and the average index of thin-layer contamination is extracted from the thin-layer contamination feature map. This is used to perform feature dimensionality reduction and physical quantization on multi-source heterogeneous spatial sensing data, transforming complex geometric distribution information and texture difference information into key numerical indicators that can directly characterize the physical damage state of photovoltaic modules. This provides standardized input variables with clear physical meaning and mathematical decoupling for subsequent hybrid feature fusion and power loss estimation, ensuring that the evaluation model can accurately distinguish and quantify the specific contribution of different types of pollution to power generation.

[0052] More specifically, in a concrete example of this application, for the extraction of the total volume of thick contamination, firstly, based on the sampling density and projection parameters during 3D point cloud reconstruction, the effective area represented by a single point cloud data point on the photovoltaic module surface is determined. Then, using the point index stored in the thick contamination mask as the address key, the corresponding normal distance value is retrieved from the normal distance array generated in the previous step; this value characterizes the physical thickness of the contamination at that point. Next, the normal distance of each mask point is multiplied by the effective area to obtain the micro-volume of that point, and the micro-volumes corresponding to all mask points are accumulated to calculate the total volume of thick contamination covering the module surface. Simultaneously, for the extraction of the thin-layer contamination average index, the effective pixel region defined by the flat area mask in the thin-layer contamination feature map is traversed, and the color difference value or texture attenuation value stored at each pixel location is read. Finally, the arithmetic mean of the values ​​of all effective pixels is calculated to obtain a scalar index characterizing the average coverage density and optical attenuation of the thin-layer contamination in the entire flat area, which serves as the thin-layer contamination average index. Accordingly, the total contamination coverage area is calculated based on the thick contamination mask and the thin contamination feature map. It is understood that the output power of a photovoltaic module depends not only on the thickness of the contaminant or the degree of light transmittance attenuation, but also directly on the effective light-emitting area of ​​the cells obscured by the contaminant. Furthermore, the complete shading caused by thick contamination and the partial shading caused by thin contamination together spatially constitute the total physical obstacle to the photovoltaic array's reception of radiant energy. Therefore, in the technical solution of this application, the total contamination coverage area is further calculated based on the thick contamination mask and the thin contamination feature map, thereby statistically representing the total physical surface area actually occupied by various types of contaminants on the photovoltaic array surface. This introduces a key area normalization parameter into the power loss assessment model, ensuring that the model can correctly distinguish the different impact modes of localized high-density contamination and large-area low-density contamination on power generation performance, thereby improving the accuracy of filling factor reduction and current mismatch estimation.

[0053] More specifically, in a specific example of this application, this step first determines the unit projected area of ​​a single pixel or point cloud data point on the physical plane of the photovoltaic module based on the calibration parameters of the imaging system and the spatial resolution of the 3D reconstruction. Then, for a thick contamination mask, the total number of valid point indices contained therein is counted, and this number is multiplied by the unit projected area to calculate the physical coverage area of ​​the thick contamination region. Simultaneously, for a thin contamination feature map, its pixel array is traversed, and the total number of pixels with feature values ​​greater than a preset noise threshold within the flat area mask is counted, and this number is multiplied by the unit projected area to obtain the physical coverage area of ​​the thin contamination region. Finally, the calculated physical coverage area of ​​the thick contamination is added to the physical coverage area of ​​the thin contamination to obtain the total contamination coverage area. This value represents the sum of the effective light-receiving surface area of ​​the photovoltaic array affected by contamination and serves as one of the dimensions for subsequently constructing the hybrid feature vector.

[0054] Accordingly, the total volume of heavy contamination, the average index of thin contamination, and the total contamination coverage area are vectorized to obtain a hybrid feature vector. It is understandable that, since the total volume of heavy contamination, the average index of thin contamination, and the total contamination coverage area are three independent physical quantities with different physical dimensions (such as volume units, dimensionless exponents, and area units) and numerical orders of magnitude, this data heterogeneity makes it impossible to directly input them in parallel into a unified regression model based on gradient descent or matrix operations for efficient processing. Furthermore, without unified structured encapsulation, the model struggles to capture the nonlinear coupling effects of each feature on power loss. Therefore, in the technical solution of this application, the total volume of heavy contamination, the average index of thin contamination, and the total contamination coverage area are further vectorized to obtain a hybrid feature vector, thereby eliminating the dimensional differences between different physical features and integrating discrete scalar data into an ordered mathematical object within a unified feature space. This provides a standardized tensor input interface for subsequent pre-trained power loss regression models, ensuring that the model can process information from the three dimensions of geometric stacking, texture decay, and coverage with balanced weights, thereby improving the convergence speed and inference accuracy of power loss estimation.

[0055] More specifically, in a concrete example of this application, this step first establishes a one-dimensional array or tensor container whose dimension matches the number of nodes in the input layer of the pre-trained model. Then, the total volume of the thick layer of filth, the average index of the thin layer of filth, and the total filth coverage area calculated in the previous steps are retrieved. Before filling these raw values ​​into the container, to prevent gradient bias during model training or inference due to large differences in numerical magnitudes (e.g., the volume value may be much larger than the texture index), Z-score normalization or Min-Max normalization is performed on these three features using the mean and standard deviation parameters obtained beforehand from the training dataset, mapping them to the same order of magnitude (e.g., between 0 and 1 or a standard normal distribution). Next, following a fixed order determined during model training (e.g., volume for the first dimension, texture index for the second dimension, and area for the third dimension), the three normalized values ​​are sequentially filled into the corresponding positions in the array container. Finally, the filled ordered array is encapsulated into a hybrid feature vector and given a data type (such as 32-bit floating-point) that is compatible with the subsequent calculation framework. This vector is then output to the power loss estimation module as a complete feature descriptor describing the current state of contamination of the photovoltaic array.

[0056] Specifically, in step S600, power loss estimation and strategy generation are performed based on hybrid feature vectors to obtain the percentage of pollution power loss and cleaning action recommendations. It is understood that simple physical feature quantification is insufficient to intuitively reflect the actual damage to the electrical output capability of the photovoltaic array caused by pollution, and static physical threshold determination cannot resolve the complex nonlinear mapping relationship between pollution morphology and power generation efficiency. Furthermore, it is difficult to combine dynamically changing electricity market economic parameters and future weather conditions to formulate optimal operation and maintenance decisions. Therefore, in the technical solution of this application, power loss estimation and strategy generation are further performed based on hybrid feature vectors to obtain the percentage of pollution power loss and cleaning action recommendations. This utilizes a pre-trained nonlinear regression model to establish a mapping mechanism from multidimensional physical features to precise electrical loss, and comprehensively considers current economic benefits and future weather trends to dynamically evaluate the input-output ratio of cleaning actions. This enables high-precision prediction of the current power loss state of the photovoltaic array and generates intelligent cleaning strategies that conform to the principle of maximizing economic benefits throughout the entire life cycle, effectively avoiding waste of operation and maintenance costs due to blind cleaning or loss of power generation revenue due to delayed cleaning.

[0057] More specifically, in a specific example of this application, power loss estimation and policy generation based on a hybrid feature vector to obtain the percentage of pollution power loss and cleaning action recommendations include: inputting the hybrid feature vector into a pre-trained power loss regression model to obtain the percentage of pollution power loss; and generating cleaning action recommendations based on the percentage of pollution power loss, a set of economic parameters, and weather forecast data.

[0058] Accordingly, the hybrid feature vector is input into a pre-trained power loss regression model to obtain the percentage of pollution power loss. It is understandable that the local shading caused by heavy pollution and the spectral transmittance attenuation caused by thin pollution are fundamentally different in their physical mechanisms. Furthermore, these two effects act on the IV (current-voltage) characteristic curve of the photovoltaic module in a complex nonlinear coupling manner, making it difficult to derive the accurate power generation efficiency attenuation based solely on a single geometric or texture feature through simple linear superposition calculations. Therefore, in the technical solution of this application, the hybrid feature vector is further input into the pre-trained power loss regression model to obtain the percentage of pollution power loss. This utilizes the powerful nonlinear fitting capability of machine learning models to establish an end-to-end mapping relationship between multi-dimensional physical morphological features and the final electrical performance indicators. This effectively shields the interference of ambient temperature and irradiance fluctuations on real-time power measurement, directly outputting the theoretical power loss ratio caused solely by pollution accumulation, providing a highly confident quantitative basis for subsequent economic decisions.

[0059] More specifically, in a concrete example of this application, this step first loads a pre-trained and validated power loss regression model. This model is built upon a massive amount of historically accumulated sample data, which covers contamination states of varying thickness, coverage area, and texture features, along with their corresponding measured power loss labels. The model architecture employs a gradient boosting decision tree or a multilayer perceptron network to ensure regression accuracy. During inference, the standardized hybrid feature vector generated in the preceding steps is fed as an input tensor to the model's input layer. Internally, the model performs forward propagation calculations on the input features based on fixed weight parameters and activation functions, analyzing the contribution weights of thick contamination volume to current mismatch and thin contamination texture indices to the decrease in fill factor. After feature cross-interaction and nonlinear transformation in the hidden layers, the output layer calculates and generates a continuous scalar value. This value represents the predicted percentage of contamination power loss, indicating the proportion of power generation capacity lost by the photovoltaic array relative to an ideal clean state if cleaning is not performed at the current moment.

[0060] Accordingly, cleaning action recommendations are generated based on the percentage of power loss due to pollution, a set of economic parameters, and weather forecast data. It is understandable that simply calculating the physical power loss is insufficient to directly support economically viable O&M decisions, and fixed loss thresholds ignore revenue differences caused by electricity market price fluctuations and the potential zero-cost cleaning effect of upcoming natural rainfall, easily leading to ineffective or inefficient cleaning operations. Therefore, in the technical solution of this application, cleaning action recommendations are further generated based on the percentage of power loss due to pollution, a set of economic parameters, and weather forecast data to construct a multi-dimensional dynamic decision-making model that comprehensively weighs the physical performance restoration requirements with the input-output ratio of cleaning operations and future weather probabilities. This transforms the O&M strategy from passive fault response to proactive profit optimization, ensuring that each cleaning action is triggered based on a clear expectation of positive economic returns, thereby minimizing the total lifecycle O&M costs of the photovoltaic power plant.

[0061] More specifically, in a concrete example of this application, this step first retrieves a set of economic parameters from the operation and maintenance database, including the current feed-in tariff, the engineering cost of a single cleaning operation, and the rated installed capacity and average daily effective irradiance duration of the photovoltaic array. The percentage of pollution power loss output in the previous step is then converted into a specific daily expected power generation revenue loss amount. Subsequently, the daily revenue loss amount is divided by the cost of a single cleaning operation to calculate the cleaning investment payback period, i.e., the number of days required for the power generation gains recovered through cleaning to cover the cleaning costs. Simultaneously, weather forecast data for the next few days returned by the meteorological service interface is analyzed to extract the probability of rainfall and the expected rainfall amount, in order to identify the expected arrival time of the nearest effective natural rainfall event. Finally, a logical judgment procedure is executed. Only when the calculated investment payback period is less than a preset economically viable threshold and significantly less than the remaining days until the next effective natural rainfall, a cleaning action instruction is generated and output; otherwise, if the investment payback period is too long or sufficient rainfall to remove pollution is expected in the short term, a suggestion not to clean is generated, thereby avoiding waste of cleaning resources due to upcoming rainfall.

[0062] In summary, the photovoltaic array cleaning strategy recommendation method based on feature analysis according to the embodiments of this application is clarified. It constructs a registered on-site point cloud by acquiring a 3D image of the photovoltaic array and registering it with a cleaning digital twin model. Utilizing the difference in normal height between the 3D point cloud and the digital twin model, thick dirt with physical thickness is first geometrically segmented from thin shadows or flat areas, effectively eliminating the interference of illumination and shadows on dirt assessment. For the segmented flat areas, the color difference between the on-site texture and the standard texture of the digital twin is further compared to accurately quantify the distribution degree of micron-level thin dust layers, compensating for the insensitivity of simple geometric measurements to small dust accumulations. Finally, the volumetric characteristics of thick dirt and the texture index of thin dirt layers are fused using multimodal hybrid features. Power loss is estimated through a pre-trained model, thereby achieving accurate decoupling and quantitative assessment of different types of dirt in complex environments, and generating a cleaning strategy that considers both economic benefits and practicality.

[0063] Figure 7 This is a block diagram of a photovoltaic array cleaning strategy recommendation system based on feature analysis according to an embodiment of this application. Figure 7As shown, the photovoltaic array cleaning strategy recommendation system 100 based on feature analysis according to an embodiment of this application includes: an image acquisition module 110, used to acquire a left image and a right image of a photovoltaic array object; a 3D point cloud reconstruction and digital twin registration module 120, used to perform 3D point cloud reconstruction and digital twin registration on the left image and the right image of the photovoltaic array object to obtain a registered on-site point cloud; and a heavy dirt segmentation module 130, used to perform heavy dirt segmentation based on 3D geometric difference on the registered on-site point cloud based on a height difference threshold. The system is segmented to obtain a thick contamination mask and a flat area mask; a thin-layer contamination quantization module 140 is used to perform thin-layer contamination quantization based on 2D texture contrast on the registered field point cloud based on the flat area mask to obtain a thin-layer contamination feature map; a hybrid feature fusion module 150 is used to perform hybrid feature fusion on the thick contamination mask and the thin-layer contamination feature map to obtain a hybrid feature vector; and a power loss estimation and strategy generation module 160 is used to perform power loss estimation and strategy generation based on the hybrid feature vector to obtain the contamination power loss percentage and cleaning action suggestions.

[0064] As described above, the photovoltaic array cleaning strategy recommendation system 100 based on feature analysis according to the embodiments of this application can be implemented in various types of computing devices or control units. For example, it can be deployed in the edge computing gateway of a photovoltaic power plant, the onboard computer of an inspection drone, the central server of a photovoltaic operation and maintenance management platform, or in a SCADA (Supervisory Control and Data Acquisition) workstation in a centralized control center. In one possible implementation, the photovoltaic array cleaning strategy recommendation system 100 based on feature analysis according to the embodiments of this application can be integrated into the computing device as a software module and / or hardware module. For example, the photovoltaic array cleaning strategy recommendation system 100 based on feature analysis can be an intelligent analysis component in the operating environment of the computing device or operation and maintenance system. This software module is configured to perform synchronous acquisition of stereo image data, accurate registration of 3D point cloud reconstruction and digital twin models, quantification of contamination layering based on geometric and texture differences, and power loss regression and economic cleaning decision generation based on hybrid features. Alternatively, it can be a dedicated photovoltaic intelligent inspection algorithm library developed for the computing device. Of course, the photovoltaic array cleaning strategy recommendation system 100 based on feature analysis can also be one of the many hardware modules of the computing device or control unit, or it can be embedded in a field-programmable gate array to accelerate the stereo matching and large-scale point cloud processing process in parallel, or it can be a photovoltaic array health diagnosis on-chip system for a specific application.

[0065] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for recommending photovoltaic array cleaning strategies based on feature analysis, characterized in that, include: Obtain the left and right images of the photovoltaic array object; Three-dimensional point cloud reconstruction and digital twin registration are performed on the left and right images of the photovoltaic array object to obtain the registered field point cloud; Based on the height difference threshold, the registered field point cloud is segmented into thick contamination based on 3D geometric difference to obtain thick contamination mask and flat area mask. Based on flat region masking, thin-layer contamination quantization based on 2D texture contrast is performed on the registered field point cloud to obtain thin-layer contamination feature map. A hybrid feature vector is obtained by fusing the features of the thick dirt mask and the thin dirt feature map. Power loss estimation and strategy generation are performed based on hybrid feature vectors to obtain the percentage of power loss due to contamination and recommendations for cleaning actions.

2. The photovoltaic array cleaning strategy recommendation method based on feature analysis according to claim 1, characterized in that, 3D point cloud reconstruction and digital twin registration are performed on the left and right images of the photovoltaic array object to obtain the registered on-site point cloud, including: Semi-global matching is performed on the left and right images of the photovoltaic array object to obtain a disparity map; The disparity map is reconstructed into a 3D point cloud based on triangulation to obtain the original 3D point cloud; The original 3D point cloud is registered with the field point cloud and digital twin based on ICP to obtain the registered field point cloud.

3. The photovoltaic array cleaning strategy recommendation method based on feature analysis according to claim 1, characterized in that, Based on the height difference threshold, the registered field point cloud is segmented into heavy contamination based on 3D geometric difference to obtain heavy contamination masks and flat region masks, including: Based on the digital twin model of the cleaning panel, the point-to-model normal distance is calculated from the registered field point cloud to obtain the normal distance array; Based on the normal distance array and the height difference threshold, the registered field point cloud is segmented based on the height threshold to obtain a heavy dirt mask and a flat area mask.

4. The photovoltaic array cleaning strategy recommendation method based on feature analysis according to claim 1, characterized in that, Based on flat region masks, thin-layer contamination quantization based on 2D texture contrast is performed on the registered field point cloud to obtain a thin-layer contamination feature map, including: Based on the flat region mask, extract the texture map of the flat region of the site from the registered site point cloud; Based on the flat area mask, extract the reference flat area texture map from the digital twin model of the clean panel; A thin-layer fouling feature map is generated based on the reference flat area texture map and the on-site flat area texture map.

5. The photovoltaic array cleaning strategy recommendation method based on feature analysis according to claim 4, characterized in that, Based on the reference flat area texture map and the on-site flat area texture map, a thin-layer fouling feature map is generated, including: A color difference matrix is ​​obtained by calculating the pixel-by-pixel color difference between the reference flat area texture map and the on-site flat area texture map. A thin-layer contamination feature map is generated based on the color difference matrix and the flat region mask.

6. The photovoltaic array cleaning strategy recommendation method based on feature analysis according to claim 5, characterized in that, A pixel-by-pixel color difference calculation is performed on the reference flat area texture map and the on-site flat area texture map to obtain a color difference matrix, including: The reference flat area texture map and the on-site flat area texture map are converted to color space to obtain the CIELAB color space reference flat area texture map and the CIELAB color space on-site flat area texture map. The CIEDE2000 color difference between the CIELAB color space reference flat area texture map and the CIELAB color space field flat area texture map is calculated to obtain the color difference matrix.

7. The photovoltaic array cleaning strategy recommendation method based on feature analysis according to claim 1, characterized in that, A hybrid feature vector is obtained by fusing the thick sludge mask and the thin sludge feature map, including: Extracting the total volume of the thick filth from the thick filth mask; Extract the average index of thin-layer fouling from the thin-layer fouling feature map; The total contamination area is calculated based on the thick contamination mask and the thin contamination feature map. The total volume of heavy filth, the average index of thin filth, and the total filth coverage area are vectorized to obtain a mixed feature vector.

8. The photovoltaic array cleaning strategy recommendation method based on feature analysis according to claim 1, characterized in that, Power loss estimation and strategy generation are performed based on hybrid feature vectors to obtain the percentage of contamination power loss and cleaning action recommendations, including: The mixed feature vectors are input into a pre-trained power loss regression model to obtain the percentage of pollution power loss. Based on the percentage of pollution power loss, economic parameter set, and weather forecast data, cleaning action recommendations are generated.

9. A photovoltaic array cleaning strategy recommendation system based on feature analysis, characterized in that, include: The image acquisition module is used to acquire the left image and the right image of the photovoltaic array object; The 3D point cloud reconstruction and digital twin registration module is used to perform 3D point cloud reconstruction and digital twin registration on the left and right images of the photovoltaic array object to obtain the registered field point cloud. The heavy contamination segmentation module is used to perform heavy contamination segmentation on the registered field point cloud based on 3D geometric difference, based on the height difference threshold, to obtain heavy contamination mask and flat area mask. The thin-layer contamination quantization module is used to perform thin-layer contamination quantization on the registered field point cloud based on 2D texture contrast, based on flat region mask, to obtain thin-layer contamination feature map. The hybrid feature fusion module is used to perform hybrid feature fusion on the thick dirt mask and the thin dirt feature map to obtain a hybrid feature vector; The power loss estimation and strategy generation module is used to estimate power loss and generate strategies based on hybrid feature vectors to obtain the percentage of power loss due to contamination and recommendations for cleaning actions.

10. The photovoltaic array cleaning strategy recommendation system based on feature analysis according to claim 9, characterized in that, The 3D point cloud reconstruction and digital twin registration module includes: A semi-global matching unit is used to perform semi-global matching on the left and right images of the photovoltaic array object to obtain a disparity map. The 3D point cloud reconstruction unit is used to perform 3D point cloud reconstruction based on triangulation on the disparity map to obtain the original 3D point cloud. The on-site point cloud and digital twin registration unit is used to perform ICP-based on-site point cloud and digital twin registration on the original 3D point cloud to obtain the registered on-site point cloud.

Images