A photovoltaic power station light resource estimation method and device, computer equipment and medium

Abstract

This invention provides a method, apparatus, computer equipment, and medium for estimating solar resources in a photovoltaic power station, belonging to the field of photovoltaic power generation. The method includes: acquiring aerial photographs of a photovoltaic power station in a target area with complex terrain; reconstructing a 3D point cloud based on a structure-reconstruction-motion algorithm; performing 3D terrain reconstruction and 3D photovoltaic module reconstruction based on the reconstructed 3D point cloud; constructing a 3D scene of the photovoltaic power station and setting the spectral attributes of scene elements; dynamically adjusting the solar zenith angle and azimuth angle parameters to simulate the transmission process of downward solar radiation in the 3D scene, and outputting the absorptive photosynthetically active radiation value at the photovoltaic module level; and determining the impact of terrain shading on the light-receiving efficiency of the photovoltaic modules based on multiple instantaneous simulations. This effectively quantifies the impact of terrain shading on the light-receiving efficiency of photovoltaic modules, thereby significantly improving the estimation accuracy of solar irradiance and power generation for mountainous photovoltaic power stations.

Description

Technical Field

[0001] This invention belongs to the field of photovoltaic power generation, and specifically relates to a method, device, computer equipment and medium for estimating the light resources of a photovoltaic power station. Background Technology

[0002] Energy and electricity are the foundation of economic development. With rapid economic and social development, the demand for electricity across society has continued to grow beyond expectations, leading to a surge in fossil fuel consumption. Traditional non-renewable energy sources are gradually becoming depleted, and a series of environmental problems, such as the "greenhouse effect," are becoming increasingly prominent. Finding new clean and renewable alternative energy sources is imperative. Solar energy, as one of the most common renewable energy sources, has received widespread attention due to its abundant resources, environmental friendliness, and renewability. Photovoltaic power generation, as a major method of utilizing solar energy, has experienced explosive growth in recent years.

[0003] Photovoltaic power plants mainly consist of photovoltaic arrays, converters, inverters, and other components. They convert solar energy into electrical energy through the photovoltaic effect and then feed it into the power grid. For a photovoltaic power generation system, the prediction of its power generation is crucial for the dispatching of the large power grid. Accurate estimation of solar irradiance, and based on this, prediction of photovoltaic power generation, provides an important basis for the grid-connected operation of photovoltaic power plants and the optimization of grid dispatching. It also provides a reference for timely detection of equipment anomalies and troubleshooting, which is of great significance for improving the stability of power system operation. With the increasing contradiction between the scarcity of land resources and society's demand for clean energy, plains with high-quality development conditions are becoming increasingly scarce, and complex mountainous terrain has gradually become the main choice for photovoltaic power plant development. Mountain photovoltaic power plants are built in complex terrain conditions such as mountains and hills, characterized by uneven surfaces, varying orientations, and relatively dispersed usable areas, resulting in complex downward solar radiation patterns.

[0004] Some experts and scholars have conducted in-depth research on the prediction of solar radiation in complex terrain areas, but most of these studies are based on large-scale digital elevation models (DEMs). Mountain photovoltaic power stations are located in small-scale areas with complex micro-geographical environments and uneven spatial distribution of light resources, making it difficult to accurately estimate the solar irradiance received and the power generation. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides a method, apparatus, computer equipment, and medium for estimating solar resources in photovoltaic power plants.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A method for estimating solar resources in a photovoltaic power plant, the method comprising: Aerial images of photovoltaic power plants in the target area are acquired, and 3D point clouds are reconstructed based on the structure-recovery-motion algorithm. Based on the reconstructed 3D point cloud, a cloth simulation filtering algorithm is used to separate ground points from non-ground points. A triangular mesh structure terrain model is generated based on the ground point cloud. The photovoltaic module point cloud is segmented from the non-ground point cloud. A spatial clustering algorithm is used to process the photovoltaic module point cloud to obtain an aggregated point cloud. The aggregated point cloud is then fitted into a simplified photovoltaic panel model based on principal component analysis. The triangular mesh terrain model and simplified photovoltaic panel model are imported into the three-dimensional radiation transfer model LESS to construct a three-dimensional scene of the photovoltaic power station. The spectral attributes of the scene elements are set, the solar zenith angle and azimuth angle parameters are adjusted, and the effective radiation value of each photovoltaic module is output. Based on the effective radiation value, the daily variation curve of radiation for each photovoltaic module is plotted. The total daily radiation received by each photovoltaic module is calculated by integration, thereby obtaining the light resources of the photovoltaic power station in the target area.

[0007] Optionally, after reconstructing the 3D point cloud based on the structure-of-motion (SOG) algorithm, the reconstructed 3D point cloud is further preprocessed, including: Denoising of the reconstructed 3D point cloud is performed based on a pre-selected filtering algorithm; The denoised point cloud is manually cropped to remove residual noise points close to the terrain surface; The spatial thinning method is adopted to reduce the point cloud density while preserving the topographic and feature characteristics by setting a minimum point spacing.

[0008] Optionally, the step of using a cloth simulation filtering algorithm to separate ground points from non-ground points includes: Invert the point cloud data and initialize a cloth mesh above the point cloud; The process of cloth protons falling under gravity and displacement caused by the internal forces of neighboring protons was simulated iteratively. In each iteration, the current position of the cloth proton is compared with the elevation value of the inverted point cloud, and when the proton touches the point cloud, it is set to an immovable state. After the iteration terminates, the distance between each point cloud data and the final cloth model is calculated, and the point cloud is classified into ground points and non-ground points based on a preset distance threshold.

[0009] Optionally, a triangular mesh terrain model based on ground point clouds is generated using the Delaunay triangulation algorithm, including: The 3D ground point cloud is projected onto the best plane fitted by the least squares method, and the Delaunay triangulation is completed in the 2D domain. Based on the spatial adjacency relationships of the original 3D points, the 3D triangular mesh topology is reconstructed to generate a triangular mesh terrain model.

[0010] Optionally, the step of using a spatial clustering algorithm to process the photovoltaic module point cloud to obtain an aggregated point cloud includes: The DBSCAN algorithm is used to perform cluster analysis on the point cloud of a single photovoltaic module, identify and remove discrete noise points that are attached to the terrain, and retain the cluster with the most points as the aggregated point cloud of the photovoltaic module.

[0011] Optionally, fitting the aggregated point cloud into a simplified photovoltaic panel model based on principal component analysis includes: Principal component analysis was performed on the aggregated point cloud of photovoltaic modules to determine the directions of its three principal components; The first two principal components are selected to form the plane of the photovoltaic panel, while the third principal component, which has the smallest change in the thickness direction, is ignored. The boundary of the point cloud in the principal component coordinate system is calculated, a rectangular plane ignoring thickness is generated, and then the plane is transformed back to the original coordinate system through inverse transformation, thereby completing the fitting of the simplified photovoltaic panel model.

[0012] Optionally, after obtaining the daily radiation variation curve for each photovoltaic module, the method further includes: By comparing the daily variation curves of radiation of photovoltaic modules at different locations, the differences in the time of radiation peak occurrence, peak size and effective light reception duration were determined, and the correlation was made with the terrain undulations and relative positions between modules in the three-dimensional scene, thereby determining the impact of terrain shading on the light reception efficiency of photovoltaic modules.

[0013] A photovoltaic power plant solar resource estimation device, the device comprising: The acquisition module is used to acquire aerial images of photovoltaic power plants in the target area and reconstruct 3D point clouds based on the structure-reconstruction-motion algorithm. The reconstruction module is used to separate ground points from non-ground points based on the reconstructed 3D point cloud using a cloth simulation filtering algorithm, and generate a triangular mesh structure terrain model based on the ground point cloud; it segments the photovoltaic module point cloud from the non-ground point cloud, processes the photovoltaic module point cloud using a spatial clustering algorithm to obtain an aggregated point cloud, and fits the aggregated point cloud into a simplified photovoltaic panel model based on principal component analysis. The simulation module is used to import the triangular mesh structure terrain model and the simplified photovoltaic panel model into the three-dimensional radiation transfer model LESS to construct a three-dimensional scene of the photovoltaic power station, set the spectral attributes of scene elements, adjust the solar zenith angle and azimuth angle parameters, and output the effective radiation value of a single photovoltaic module. The analysis module is used to plot the daily variation curve of radiation of a single photovoltaic module based on the effective radiation value, calculate the total daily received radiation by integration, and thus obtain the light resources of the photovoltaic power station in the target area.

[0014] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned method for estimating solar resources in a photovoltaic power plant.

[0015] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the aforementioned method for estimating solar resources in a photovoltaic power plant.

[0016] The photovoltaic power plant solar resource estimation method provided by this invention has the following beneficial effects: This invention acquires aerial photographs of the target area and reconstructs 3D point clouds based on a motion-reconstruction structure algorithm. Then, it separately reconstructs the 3D terrain and photovoltaic (PV) modules. A triangular mesh terrain model is generated using a cloth-based simulation filtering algorithm, and a simplified PV panel model is fitted using spatial clustering and principal component analysis. These two models are then imported into a 3D radiative transfer model (LESS) to construct a 3D scene of the PV power station and set spectral attributes. Solar parameters are dynamically adjusted within the model to simulate radiative transfer, outputting the absorptive photosynthetically active radiation (SEPA) value at the PV module level. Based on multiple simulations, daily variation curves of radiation received by individual PV modules are plotted, and the total daily received radiation is calculated through integration. This method can more accurately characterize the complex terrain and spatial distribution features of mountainous PV power stations, dynamically simulate the radiative transfer process under different solar angles, and effectively quantify the impact of terrain shading on the light-receiving efficiency of PV modules, thereby significantly improving the estimation accuracy of solar irradiance and power generation for mountainous PV power stations. Attached Figure Description

[0017] To more clearly illustrate the embodiments and design schemes of the present invention, the accompanying drawings required for this embodiment will be briefly described below. The drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of a photovoltaic power plant solar resource estimation technology route provided by the present invention according to an exemplary embodiment.

[0019] Figure 2 This is a flowchart illustrating a photovoltaic power plant solar resource estimation method according to an exemplary embodiment of the present invention.

[0020] Figure 3 This is a schematic diagram illustrating the principle of obtaining terrain by inverting point clouds according to an exemplary embodiment of the present invention.

[0021] Figure 4 This invention provides a triangular mesh structure terrain based on ground point cloud generation according to an exemplary embodiment.

[0022] Figure 5 This is a training diagram of a deep learning model for processing point cloud problems according to an exemplary embodiment of the present invention.

[0023] Figure 6 This is a deep learning model inference graph provided by the present invention according to an exemplary embodiment.

[0024] Figure 7 This is a comparison diagram of the original point cloud, the labeled point cloud, and the prediction result provided by the present invention according to an exemplary embodiment.

[0025] Figure 8 This invention provides a pseudo-image generated based on 0.5m and 1m voxel grids according to an exemplary embodiment.

[0026] Figure 9 This is a schematic diagram of the point cloud of all photovoltaic modules obtained by manual cropping according to an exemplary embodiment of the present invention.

[0027] Figure 10 This is a schematic diagram illustrating the principle of a DBSCAN algorithm according to an exemplary embodiment of the present invention.

[0028] Figure 11 This invention provides a three-dimensional mountain photovoltaic power station model in LESS according to an exemplary embodiment.

[0029] Figure 12 This is a schematic diagram of a hierarchical grouping simulation APAR provided by the present invention according to an exemplary embodiment.

[0030] Figure 13 This is a schematic diagram illustrating the basic principle of uplink and downlink radiation simulation in LESS according to an exemplary embodiment of the present invention.

[0031] Figure 14 This is a schematic diagram of instantaneous solar radiation simulation at noon on the summer solstice, provided by the present invention according to an exemplary embodiment.

[0032] Figure 15 This is a schematic diagram of instantaneous solar radiation simulation at noon on the winter solstice, provided by the present invention according to an exemplary embodiment.

[0033] Figure 16 This is a schematic diagram of instantaneous solar radiation simulation at 16:00 on the afternoon of the summer solstice, provided by the present invention according to an exemplary embodiment.

[0034] Figure 17 This is a schematic diagram of instantaneous solar radiation simulation at 8:00 AM on the winter solstice, provided according to an exemplary embodiment of the present invention.

[0035] Figure 18This is a data map showing the daily solar radiation received by photovoltaic modules at five typical locations, according to an exemplary embodiment of the present invention.

[0036] Figure 19 This is a block diagram of a photovoltaic power plant solar resource estimation device provided by the present invention according to an exemplary embodiment. Detailed Implementation

[0037] To enable those skilled in the art to better understand and implement the technical solutions of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and should not be construed as limiting the scope of protection of the present invention.

[0038] This invention innovatively proposes the use of the three-dimensional radiative transfer model (LESS) to achieve refined modeling and accurate estimation of solar downflow radiation for mountain photovoltaic power plants. The main innovation lies in the first application of three-dimensional radiative transfer theory to the prediction of solar radiation received by photovoltaic power plants. A high-precision three-dimensional scene model is established in LESS to simulate the solar radiation transfer process under complex terrain conditions.

[0039] This invention focuses on a typical mountain photovoltaic power station within a photovoltaic base for ecological restoration of a mine. This project is the first large-scale photovoltaic base in the region, making full use of existing barren mountains and wastelands, including areas of rocky desertification, heavy metal pollution, and mining subsidence, and actively implementing integrated development of "agricultural-photovoltaic complementarity" and "forest-photovoltaic complementarity," thus possessing high research value. Preliminary analysis of aerial photographs of the study area revealed that existing solar photovoltaic modules installed on the slopes exhibit shading, resulting in suboptimal light energy utilization efficiency. Furthermore, the terrain partially blocks sunlight, further limiting the amount of sunlight received by lower-lying photovoltaic modules throughout the day. Therefore, accurately estimating the solar irradiance received by a specific module within the power station, and using this information to optimize photovoltaic string arrangement and improve power generation prediction accuracy, will be of great significance for improving the utilization rate of solar resources and promoting the transformation and upgrading of the energy structure in resource-depleted areas represented by the study area.

[0040] The research team first reconstructed a 3D point cloud of the study area using aerial photographs taken by UAVs. Based on this, they extracted and reconstructed the point cloud of the terrain and photovoltaic strings using a cloth-based simulation filtering algorithm and a deep learning network. Using the reconstructed 3D scene model, they dynamically simulated the transmission process of solar radiation under complex terrain conditions at different times (considering diurnal and seasonal variations) by adjusting key illumination parameters such as the solar zenith angle and azimuth angle in LESS (Light Element System). This allowed them to analyze the spatiotemporal distribution characteristics of the solar irradiance received by each module. By summing the instantaneous radiation values ​​received at each moment, the total radiant energy received by the photovoltaic modules over a period of time (one day or one year) can be calculated.

[0041] The refined solar resource assessment method proposed in this invention has advantages in three main aspects: First, the refined 3D modeling fully considers the terrain features and component layout of actual photovoltaic power plants. Second, the 3D radiative transfer model based on the real structure can effectively simulate the transmission process of solar shortwave downflow radiation under complex terrain, achieving accurate estimation of solar radiation down to a specific photovoltaic module. Third, this method has good 3D visualization effects, providing an intuitive reference for power plant design and operation and maintenance management; the solar power generation potential of complex terrain areas is expected to be further explored, which will further promote the development of my country's photovoltaic industry and the transformation of its energy structure.

[0042] The project, in its actual implementation, mainly consists of a 3D reconstruction component and a radiation simulation component. The 3D reconstruction component can be further divided into terrain reconstruction and photovoltaic module reconstruction. A brief technical roadmap is as follows: Figure 1 As shown.

[0043] The technical solutions provided by the various embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0044] First, this invention provides a method for estimating the solar resources of a photovoltaic power plant, specifically as follows: Figure 2 As shown, it includes the following steps: S101. Acquire aerial images of photovoltaic power plants in the target area and reconstruct 3D point clouds based on the structure-reconstruction-motion algorithm.

[0045] In this step, aerial photographs of the study area are first acquired using UAV photogrammetry technology. Then, scene point clouds are reconstructed based on SfM (Structure from Motion, which can be implemented by professional mapping software such as Pix4Dmapper and is simple to operate; it will not be discussed in detail in this invention). The resulting point cloud of the study area can reach the order of millions.

[0046] In one embodiment, after reconstructing the 3D point cloud based on the structure-of-motion (SOG) algorithm, the reconstructed 3D point cloud is further preprocessed. In this embodiment, the point cloud denoising is performed by comparing the denoising effects of radius filtering, statistical filtering, and Gaussian filtering, and selecting the filtering algorithm that best preserves terrain features and effectively removes large-scale noise. The filtered point cloud is then manually cropped to remove residual noise close to the terrain surface. Spatial thinning is used to reduce the point cloud density while preserving terrain and feature characteristics by setting a minimum point spacing.

[0047] For example, before formally reconstructing the 3D of a photovoltaic power station, the point cloud data of the study area needs to be preprocessed. The main steps include point cloud denoising and point cloud thinning. Point cloud denoising adopts a combination of radius filtering algorithm and manual denoising, which effectively reduces the interference of noise and outliers on the terrain modeling effect. Point cloud thinning reduces the point cloud density from 40 points / square meter to about 5 points / square meter, which greatly reduces the difficulty of subsequent calculations.

[0048] 1. Point cloud filtering and noise reduction.

[0049] Due to measurement errors, environmental interference, or limitations of the sensors themselves, point cloud data often contains noise and outliers, affecting the accuracy of subsequent processing and analysis. For terrain modeling, noise and outliers can lead to unsatisfactory results (specifically, uneven terrain areas) or even incorrect terrain modeling. Therefore, point cloud data preprocessing is a crucial step in ensuring the quality of terrain modeling.

[0050] (1) Statistical filtering.

[0051] Statistical filtering is a method for identifying and removing outliers (noise) based on the distance distribution between points in a point cloud and other points. Its basic idea is that most points in a point cloud should be close to each other, while points far from other points are likely noise or outliers. The statistical filter performs spatial feature statistical analysis on the neighborhood of each point, calculates the average distance from that point to all points in the neighborhood, and then compares it to the overall average distance. If the average distance of a point is significantly greater than the overall average distance, that point is likely noise and can be removed.

[0052] (2) Radius filtering.

[0053] Radius filtering is a simple filtering method that identifies outliers based on the number of neighboring points within a point cloud. If a point has fewer than a certain threshold of neighboring points within a specified search radius, it is considered an outlier and can be removed. This method is very effective for removing isolated points or small noise clusters. Radius filtering requires setting a radius threshold *r* and a threshold *n* for the number of neighboring points within that radius. If the number of neighboring points within a radius of *r* centered on the point is less than or equal to *n*, the point is considered noise and removed; otherwise, it is retained.

[0054] (3) Gaussian filtering.

[0055] Gaussian filtering is a technique for smoothing data using a Gaussian function. A Gaussian filter is a symmetric, centered kernel function whose shape is determined by a Gaussian distribution. In point cloud processing, Gaussian filtering is commonly used to reduce noise. It is achieved by replacing the position of each point with a weighted average of its neighborhood points, with the weights determined by the Gaussian kernel. The weights of the Gaussian kernel decrease exponentially with distance, thus positions closer to the center of the neighborhood receive a larger weight.

[0056] Observations revealed that the study area contained a significant amount of large-scale noise. Using MATLAB 2023a, code was written to implement the three filtering methods mentioned above, taking the original point cloud as input data, and the filtering effects were compared. Opening the filtered point cloud in CloudCompare and comparing the results showed that statistical filtering and radius filtering were more effective at removing large-scale noise than Gaussian filtering. Furthermore, compared to statistical filtering, radius filtering preserved terrain features, making it the most suitable denoising method in this study.

[0057] 2. Manual noise reduction of point clouds.

[0058] After radius filtering, large-scale noise was largely removed, but some noise close to the terrain surface remained. The triangular mesh-style terrain surface generated from this insufficiently denoised point cloud data would have many bumps, affecting the 3D terrain modeling effect. Therefore, manual denoising is recommended to obtain a smoother point cloud surface. Manual denoising was achieved using the "Segment" tool in CloudCompare, which involves using a clipping box to define the noisy point cloud areas to be clipped and selecting to retain the area outside the clipping box. After seven rounds of manual denoising, a relatively smooth point cloud was obtained.

[0059] 3. Thinning of point clouds.

[0060] Due to limitations in research equipment, raw point cloud computing is large in volume and time-consuming. Therefore, it is necessary to thin the point cloud before terrain modeling. The main purpose of point cloud data thinning is to accurately represent the features of the ground and land features with fewer points, achieving a balance between point cloud density and data accuracy; at the same time, point cloud thinning can also play a certain role in noise reduction.

[0061] Point cloud thinning was performed using the "Subsampleapointcloud" tool in CloudCompare software, offering three thinning methods: "Spatial" thins the point cloud spatially, allowing you to set the minimum point spacing after thinning (e.g., setting it to 0.3 means maintaining a point spacing of 0.3 meters); "Random" thins randomly, allowing you to set the number of points to retain; and "Octree" thins proportionally. A rough estimate suggests the point cloud density in the current study area is approximately 40 points / square unit. This study used the "Spatial" tool to reduce the point cloud density to a reasonable approximately 5 points / square unit.

[0062] S102. Based on the reconstructed 3D point cloud, perform 3D terrain reconstruction and 3D photovoltaic module reconstruction respectively to generate a triangular mesh structure terrain model and a simplified photovoltaic panel model.

[0063] In this step, the terrain 3D reconstruction uses a cloth simulation filtering algorithm to separate ground points from non-ground points, and generates a triangular mesh terrain model based on the ground point cloud; the photovoltaic module 3D reconstruction includes segmenting the photovoltaic module point cloud from the non-ground point cloud, using a spatial clustering algorithm to process the photovoltaic module point cloud to obtain an aggregated point cloud, and fitting the aggregated point cloud into a simplified photovoltaic panel model based on principal component analysis.

[0064] In one embodiment, to generate a triangular mesh terrain model, the point cloud data is first inverted, and a cloth mesh is initialized above the point cloud. The falling process of cloth protons under gravity and their displacement due to the internal forces of neighboring protons are iteratively simulated. In each iteration, the current position of the cloth protons is compared with the elevation value of the inverted point cloud; when a proton touches the point cloud, it is set to an immovable state. After the iteration terminates, the distance between each point cloud data point and the final cloth model is calculated, and the point cloud is classified into ground points and non-ground points based on a preset distance threshold. Then, the 3D ground point cloud is projected onto the best plane fitted by the least squares method, and Delaunay triangulation is completed in the 2D domain. Based on the spatial adjacency relationships of the original 3D points, an accurate 3D triangular mesh topology is reconstructed, generating a triangular mesh terrain model.

[0065] For example, the terrain reconstruction section uses the CSF point cloud filtering algorithm, setting the terrain detail size to 1, the maximum number of iterations to 1000, and the classification threshold to 0.1. This extensive iteration count and stringent constraints ensure the extraction of a relatively complete ground point cloud. Finally, a .obj format triangular mesh terrain file is generated and saved in CloudCompare based on the ground point cloud. This file can be imported into LESS as terrain settings or scene elements.

[0066] Because CloudCompare and LESS use different coordinate systems, coordinate transformation and normalization are necessary before 3D reconstruction of point clouds. This step must be performed after using the CSF filtering algorithm; otherwise, filtering will fail.

[0067] 1. Separate ground points and non-ground points based on the cloth simulation filtering algorithm (CSF).

[0068] To obtain high-precision 3D terrain data for the study area and to better segment photovoltaic strings, it is necessary to first separate ground points from non-ground points. Given the complex terrain of the study area, after comparing the advantages and disadvantages of various filtering algorithms, the Cloth Simulation Point Cloud Filtering (CSF) algorithm was chosen to extract ground points. The Cloth Simulation Filtering algorithm is a physics-based point cloud data processing method. By simulating the process of a rigid cloth conforming to the terrain under gravity, it achieves accurate extraction of ground point cloud data. Its advantages include good filtering effect, simple parameter settings, and strong terrain adaptability.

[0069] (1) Detailed explanation of CSF filtering algorithm.

[0070] Cloth simulation, also known as cloth modeling, is a term in computer graphics used to simulate cloth in a computer program. The cloth simulation utilizes a mass-spring model, dividing the cloth into a lattice of particles of equal size and mass connected by lines. Each proton can store certain information. The motion between the protons follows Newton's second law and Hooke's law, and their spatial positions determine the shape and location of the cloth.

[0071] When cloth simulation is applied to point cloud data filtering methods, the following assumptions are made: First, proton motion occurs in the vertical direction; second, when a proton falls to the ground, it is immobile, and the collision force is negligible; finally, the motion of the proton under gravity and internal forces is decomposed into two discrete processes. In the improved cloth simulation filtering algorithm, the displacement of the cloth proton under gravity is first calculated, and then the proton position is corrected based on the internal forces.

[0072] The specific implementation process of the CSF algorithm to separate ground points from non-ground points is as follows: ① Mesh attribute generation: Determine the grid plane size and grid height for the cloth simulation, and define the space ratio F: F= ×100%; In the above formula, n is the number of point clouds within the grid, H is the grid height, and h is the number of point clouds within the grid. i h is the elevation value of the i-th point within the grid. minF represents the elevation value of the second lowest point within the grid. F reflects the spatial differences in point cloud data. Different semantic attributes of the grid can be defined based on F, and the terrain factors and slope information of different grids can be given in subsequent calculations based on the attribute information.

[0073] ② Point cloud data inversion: The point cloud data is inverted to prepare for subsequent cloth simulation. For example... Figure 3 As shown.

[0074] ③ Scene Classification: Based on indicators such as terrain slope, surface curvature, and undulation, the terrain is divided into different categories, and appropriate thresholds are set. It is usually divided into flat areas (slope less than 5°), hilly areas (slope 5°~10°), mountains (slope 10°~25°), and high mountains (slope greater than 25°).

[0075] ④ Initialize cloth simulation: Determine the number of protons based on the user-defined grid resolution (GR). The starting position of the cloth should be set at a height higher than the terrain data.

[0076] ⑤ Projection and Intersection Point Determination: Project the fabric protons and all point clouds onto a horizontal plane. Find the corresponding point cloud (Corresponding point, CP) for each grid proton in the plane and determine its intersection height value (IHV).

[0077] ⑥ Proton Displacement under Gravity: For each proton in the grid, calculate the displacement of each proton under gravity (if the proton can move). Compare the current height value (Currentheightvalue, CHV) of the proton with its corresponding IHV. If CHV is not greater than IHV, move the proton to height IHV and set it to an immovable state.

[0078] ⑦ Slope Information Iteration and Region Construction: The slope information is iterated using grid protons to make the slope information more accurate. Protons with the same slope information are grouped into the same region.

[0079] ⑧ Proton displacement under the influence of internal forces: Calculate the displacement of each proton under the influence of the internal forces of neighboring protons.

[0080] ⑨ Iteration termination condition check: Repeat steps ④-⑤ to check whether the maximum height variation (M-HV) of all protons is very small or exceeds the set maximum number of iterations. If either condition is met, the simulation process is terminated.

[0081] ⑩ Distance Calculation and Classification: Calculate the distance between each point cloud data point and the cloth-simulated protons. Based on a set threshold, point cloud data points with distances less than the threshold are classified as ground points (Bare earth, BE), and other points are classified as non-ground points (Objects, OBJ). Output the separated ground point cloud data, which consists of surface regions with slope information. Through the above steps, the CSF algorithm can effectively separate ground points and non-ground points in point clouds in complex terrain areas.

[0082] (2) Application of CSF filtering algorithm in this study.

[0083] The source code for the CSF filtering algorithm can be downloaded from GitHub. This algorithm has also been widely integrated into software such as CloudCompare and EasyPoint. Open the CSF source code using MATLAB 2023a. Beforehand, download the ComputerVisionToolbox toolkit and convert the point cloud data type to .ply format in CloudCompare to meet the input format requirements of the code.

[0084] The core parameters for the CSF algorithm are as follows: ① Only terrain details larger than this value will be preserved; those smaller will be ignored. Furthermore, a larger value will filter out more terrain features. For steep mountainous terrain, this value should be set to 1.

[0085] ② The maximum number of iterations for fabric drop simulation calculations is typically 200, which is sufficient for most scenarios. In this study, to improve simulation accuracy, this parameter is set to 500.

[0086] ③ Points with a classification threshold less than the cloth model value will be classified as ground points. This value is set to 0.1.

[0087] After adjusting the parameters and setting the correct file path, the program will output a visual result and generate one ground point cloud file and one non-ground point cloud file. The operations in the second part will all be based on the output ground point cloud (Ground.las).

[0088] 2. Point cloud fitting surface based on CloudCompare.

[0089] (1) Scheme 1: Point cloud interpolation and DEM generation.

[0090] In the process of 3D terrain reconstruction, the original plan was to use the "Point Cloud to Raster" function in ArcGIS 10.8 to interpolate missing point cloud data using inverse distance weighted interpolation and generate a DEM product. Finally, GlobalMapper and 3ds Max software would be used to convert the DEM product into a .obj format 3D terrain image. This method can achieve high-resolution 3D terrain modeling. However, because the areas covered by the original data are not regular, the method of directly interpolating from point clouds to generate the DEM resulted in incorrect terrain information being interpolated in areas where there was originally no terrain information. For the study area, this method proved infeasible.

[0091] (2) Scheme 2: The point cloud is fitted to a triangular mesh structure terrain (Mesh) surface.

[0092] Load the separated ground points into Cloud Compare, click

Edit

Mesh

Delaunay2.5D (XYbestfittingplane)

OK

[0093] The triangulation process is based on the spatial region growing technique of Delaunay triangulation. First, the algorithm constructs initial seed triangles from the original point cloud. This process does not directly select ready-made triangles, but rather dynamically generates them by analyzing the local point cloud's geometric features: after selecting a core point in the point cloud, it constructs initial triangles that conform to the Delaunay criterion by combining the spatial distribution of its neighboring points and the consistency of their normal vectors, using these as the starting point for surface growth. Then, the algorithm enters an iterative expansion phase, continuously detecting the candidate point set at the growth front and using the Delaunay empty circle criterion (for 2D cases) or the empty sphere criterion (for 3D cases) to select the optimal connection method, ensuring that the optimal topology is maintained with each expansion.

[0094] In the practical application of 3D point cloud processing, the algorithm employs a dimensionality reduction mapping strategy to ensure computational efficiency. By projecting the 3D spatial point cloud onto the best-fitting plane (XY plane) fitted by the least squares method, Delaunay triangulation is first performed in the 2D domain, and then accurate 3D topological connections are reconstructed based on the spatial adjacency relationships of the original 3D points. This method not only inherits the mathematical completeness of Delaunay triangulation but also introduces an adaptive density threshold and a dynamic boundary detection mechanism, enabling the algorithm to intelligently handle real-world situations such as uneven point cloud density and missing data, ultimately outputting a complete triangular mesh surface with geometric consistency.

[0095] 3. Point cloud coordinate transformation and normalization.

[0096] Open the Mesh.obj file, the terrain reconstructed using the above process, in the computer's built-in 3D file viewer. You'll find that the terrain is now vertical. This is because the coordinate systems of CloudCompare and the .obj file are different (Note: Scene elements in LESS are all .obj format files, therefore the coordinate system used by the .obj file is also the coordinate system used by the 3Dviewer in LESS). To ensure that the terrain reconstructed by CloudCompare can be correctly loaded in LESS, coordinate transformation and normalization are necessary.

[0097] Testing showed that the terrain Mesh.obj file reconstructed from the processed point cloud could be displayed correctly in the 3D file viewer, and its center fell on the origin of the LESS coordinate system. Similarly, to ensure that the subsequently reconstructed photovoltaic module model could be correctly loaded in LESS, the non-ground point cloud also needed to undergo point cloud coordinate and normalization processing in advance.

[0098] In another embodiment, photovoltaic module reconstruction can be divided into two main steps: photovoltaic point cloud segmentation and point cloud fitting of the photovoltaic panel. Since the point cloud in the study area reaches millions of elements, manual segmentation is time-consuming and difficult. Therefore, attempts were made to use deep learning networks such as PointNet++ and RandLANet to solve the photovoltaic module segmentation problem. However, due to limited training data and unclear shape features of the photovoltaic modules, the classification results were poor. To ensure the best segmentation effect, the team ultimately adopted a manual segmentation method, segmenting and saving a total of 891 complete imaged photovoltaic panel.txt point cloud files from the scene.

[0099] 1. Exploration of photovoltaic module segmentation based on point cloud deep learning network.

[0100] Traditional point cloud classification methods rely on manual human intervention, which is time-consuming and labor-intensive when dealing with large amounts of data. In recent years, with the rapid development of deep learning technology, many innovative methods have emerged in the field of point cloud segmentation. Deep learning-based point cloud segmentation methods typically employ a supervised learning paradigm: first, a manually labeled point cloud training dataset is constructed; then, the model parameters are adjusted through algorithm optimization, enabling it to automatically learn the feature representations and segmentation rules of the point cloud data. Essentially, these methods map the raw point cloud data to a segmentation label space by constructing a complex nonlinear mapping function (i.e., a "black box" model). For example... Figure 5 As shown, the training process employs optimization algorithms such as gradient descent, continuously adjusting model parameters to achieve the best fit by minimizing the error function between the predicted result and the true label.

[0101] In the practical application stage (reasoning process), such as Figure 6As shown, simply inputting the point cloud to be segmented into the trained model yields the predicted segmentation result. The differences between different models mainly lie in their network structure design. For the photovoltaic module segmentation task, this invention selects the classic point cloud segmentation algorithm PointNet++. This algorithm is a deep learning model improved upon PointNet, effectively addressing the shortcomings of the original PointNet in local feature extraction by introducing a hierarchical learning mechanism. The PointNet series of networks pioneered the processing of raw point clouds, providing valuable reference for this research.

[0102] (1) Detailed explanation of PointNet++ algorithm.

[0103] The specific process of implementing point cloud classification and segmentation using the PointNet++ algorithm is as follows: ① Sampling. PointNet++ uses the FarthestPointSampling (FPS) algorithm for sampling. This algorithm first randomly selects a point as the starting point, then calculates the distance from all other points to the starting point, and selects the point with the farthest distance as the next sampling point. Next, it continues to calculate the minimum distance from the remaining points to the selected sampling points, and selects the next point with the farthest distance. This process is iterated until the required number of sampling points are selected.

[0104] ② Grouping. After sampling the keypoints, PointNet++ uses either the Ballquery grouping algorithm or the KNN algorithm to divide the point cloud into multiple local regions. The Ballquery grouping algorithm, given a center point and a radius, includes all points within the radius as points in that local region. This grouping method ensures a fixed region size, making it more suitable for tasks requiring local features. The KNN algorithm, on the other hand, selects the K points closest to the center point as points in the local region.

[0105] ③ Feature Extraction. PointNet++ uses PointNet as a local feature learner to extract features within each local region. PointNet can handle unordered point cloud data and extract features from local regions. Then, PointNet++ aggregates these local features to obtain a higher-level feature representation.

[0106] ④ Recursive processing. PointNet++ constructs a hierarchical representation of point clouds by recursively using the sampling, grouping, and feature extraction processes described above. This recursive processing allows PointNet++ to progressively abstract local and global features of point cloud data, thereby improving the accuracy and robustness of feature extraction.

[0107] ⑤ Classification and Segmentation. After obtaining the global features, PointNet++ can input them into a classifier or segmenter for classification or segmentation tasks. For classification problems, the global features are used directly for classification; for segmentation problems, the high-dimensional points need to be obtained with the same number of points as the low-dimensional points through methods such as inverse distance interpolation, then feature fusion and extraction are performed, and finally segmentation is performed.

[0108] PointNet++ makes significant innovations in local feature extraction compared to PointNet. PointNet only uses max pooling to extract global information, which makes it difficult to capture local features. PointNet++ introduces a hierarchical sampling strategy, selecting representative points as center points at each sampling layer, and then further sampling only in local regions of these points, constructing a hierarchical point cloud data structure.

[0109] (2) PointNet++ algorithm practice.

[0110] This study used Cloud Compare to divide the original point cloud file of a single scene into 6 blocks (scenes), labeled each of the 6 scenes, and saved them in S3DIS dataset format. Using this dataset, the PointNet++ network was trained for 32 epochs (an epoch refers to training the model sequentially using all the data in the training set, also known as "generation training"). The training results showed a mIoU (mean Intersection over Union) of only 0.16 on the validation set, indicating poor performance.

[0111] Depend on Figure 7 The visualization results show that the model predicted most of the point cloud data as background. Considering the large amount of point cloud data in our application scenario (millions of data points), we tried using a more advanced point cloud segmentation network, RandlaNet, which is suitable for large-scale scenarios.

[0112] (3) Detailed explanation of RandLA-Net algorithm.

[0113] RandLA-Net is a groundbreaking method for processing large-scale point clouds. This network innovatively employs a random downsampling strategy, overcoming the efficiency bottleneck of traditional point cloud networks that rely on computationally intensive sampling (such as farthest point sampling FPS). It is more than 200 times faster than existing methods when processing millions of point clouds. Its core contribution lies in designing a local feature aggregation module that includes local spatial encoding, attention pooling, and expanded residual blocks. While maintaining real-time processing capabilities, it achieved state-of-the-art accuracy on benchmark datasets such as Semantic3D and S3DIS at the time.

[0114] The specific process of implementing point cloud classification and segmentation using the RandLA-Net algorithm is as follows: ① Random Sampling: RandlaNet employs a random sampling strategy to reduce point cloud density and computational cost. Although random sampling may lose some useful information, the network mitigates this problem by extracting and preserving key information through local spatial encoding and attention-based pooling.

[0115] ② Local Feature Integration: Local feature integration is a core part of RandaNet, comprising two key steps: local spatial encoding and attention pooling. Local spatial encoding creates a local coordinate system around each point, enhancing the network's understanding of the point cloud geometry. Attention pooling dynamically weights features according to the importance of each point, allowing the network to focus on points more relevant to the task, thereby improving the accuracy of feature extraction.

[0116] ③ The dilated residual module in RandlaNet increases the network's receptive field, further enhancing its ability to understand point cloud data. This module introduces convolutional layers with different dilation rates, enabling the network to capture information at different scales.

[0117] ④ Global Feature Learning and Classification: After integrating local features, RandaNet aggregates these local features to obtain global contextual information. Finally, classification is performed using fully connected layers and a softmax function to assign the correct semantic label to each point.

[0118] RandlaNet achieves efficient and accurate semantic segmentation of point cloud data through steps such as random sampling, local feature integration, extended residual modules, and global feature learning and classification.

[0119] (4) Practice of RandlaNet algorithm.

[0120] We continued to train using the manually annotated S3DIS dataset. The training results show that the model's mIoU on the validation set is slightly better than PointNet++, at 0.198, but it is still low, and the segmentation results are still not ideal.

[0121] (5) Try other segmentation methods.

[0122] Considering that the aforementioned point cloud segmentation networks primarily learn the structured information and features of point clouds, while actual aerial photographs of photovoltaic panel point clouds lack obvious structural features (blending seamlessly with the ground and indistinguishable from other objects), yet possess rich color information, this paper attempts to segment point clouds by converting them into pseudo-images. The basic idea is to project the point cloud onto the XY plane to generate a pseudo-image, perform instance segmentation on the pseudo-image, and then segment the original point cloud according to the labels of the pixel grid mapped onto the pseudo-image to obtain the point cloud segmentation result.

[0123] The process of converting point cloud into pseudo image is as follows: obtain the projection length of point cloud on X and Y axes (885.269m and 901.961m respectively), construct voxel grid on XY plane (voxel size is set as needed), count the point with the largest Z coordinate in each voxel, and use the RGB information of the point to fill the pseudo image.

[0124] The pseudo-images generated under 0.5m voxel grid and 1m voxel grid are as follows: Figure 8 As shown in the pseudo-image results, when the voxel grid is set to a small size, the image is clearer, but the point cloud density is sparser. When the voxel grid is set to a large size, the point cloud density is denser, the image is clearer, but the photovoltaic panel boundary is unclear, and image segmentation cannot be supported.

[0125] Visually, photovoltaic modules exhibit a regular rectangular geometric shape, a high-reflectivity bluish-gray tone, and a typical spatial distribution pattern of rows and columns. To ensure optimal segmentation results, a traditional manual segmentation method was ultimately adopted, such as... Figure 9 As shown, a total of 891 complete photovoltaic panel point cloud files were segmented and saved from the scene.

[0126] In another embodiment, the segmented photovoltaic module point cloud is fitted to a photovoltaic panel. First, the DBSCAN algorithm is used to perform cluster analysis on the point cloud of a single photovoltaic module, identifying and removing discrete noise points caused by adhesion to the terrain, and retaining the cluster with the most points as the aggregated point cloud of that photovoltaic module. Then, principal component analysis is performed on the aggregated point cloud of the photovoltaic module to determine its three principal component directions; the first two principal component directions are selected to form the plane of the photovoltaic panel, ignoring the third principal component with the smallest change in the thickness direction; the boundary of the point cloud in the principal component coordinate system is calculated to generate a rectangular plane ignoring thickness, and then the plane is transformed back to the original coordinate system through inverse transformation, thereby completing the fitting of the simplified photovoltaic panel model.

[0127] For example, the point cloud of photovoltaic modules directly segmented from the scene contains many discrete points at the bottom due to adhesion to the terrain. To eliminate discrete points and obtain a aggregated point cloud of photovoltaic modules, the spatial clustering algorithm DBSCAN is used to process the photovoltaic module point cloud one by one, retaining the largest cluster. Based on the point cloud clustered by DBSCAN, the principal plane orientation of the photovoltaic modules is determined by PCA (Principal Component Analysis), the point cloud boundary is extracted to generate a two-dimensional rectangular plane, and a simplified photovoltaic panel .obj format file ignoring thickness is output, thus completing the three-dimensional reconstruction of the photovoltaic modules.

[0128] 1. Obtain the aggregated point cloud of photovoltaic modules using the DBSCAN algorithm.

[0129] In photovoltaic power plants, photovoltaic modules are still 1 to 2 meters above the ground. Therefore, in the original point cloud file, the photovoltaic modules appear to be "floating" on the terrain surface. This facilitates the separation of ground points and non-ground points, but it also results in too many discrete points at the bottom of the segmented photovoltaic module point cloud that are stuck to the terrain, making it impossible to correctly fit into a simplified photovoltaic panel in .obj format.

[0130] To eliminate discrete points and obtain aggregated point clouds of photovoltaic modules, the spatial clustering algorithm DBSCAN is used to process the point cloud files of individual photovoltaic modules one by one, and the largest cluster is retained.

[0131] (1) Detailed explanation of DBSCAN algorithm.

[0132] DBSCAN, proposed by Ester et al., is a density-based spatial clustering algorithm that can automatically discover clusters of arbitrary shapes and effectively identify noise points, making it suitable for point cloud clustering. DBSCAN requires manual specification of two parameters: the scan radius ε (epsilon) and the minimum number of contained points (minPts). The algorithm principle is as follows: Figure 10 As shown.

[0133] The specific process of implementing point cloud clustering using the DBSCAN algorithm is as follows: ① Input a point cloud sample and mark all points as unvisited; ② Randomly select an unvisited point P and mark P as visited; ③ If there are at least minPts points within the ε range of P, create a new cluster C, add P to C, and denote the set of points within the same range as P as Q; ④ Iterate through each point in Q. If the point is unvisited, mark it as visited. If there are at least minPts points within the ε range of the point, add these points to N. If the point is not yet a member of any cluster, add the point to C. ⑤ Output C; ⑥ Select other points marked as unvisited and repeat steps ③-⑤ until all points in the sample are marked as visited.

[0134] (2) DBSCAN algorithm practice.

[0135] Tests showed that setting the scanning radius ε and the minimum number of points minPts to 0.45 and 2 respectively can meet the clustering requirements of most photovoltaic module point clouds, and the clustering effect is good, which can effectively improve the accuracy of subsequent photovoltaic panel fitting.

[0136] 2. Simplify the photovoltaic panel by fitting the point cloud of the photovoltaic module.

[0137] In this invention, point cloud data is fitted to a photovoltaic panel plane using PCA (Principal Component Analysis), which is suitable for geometric simplification of thin-plate photovoltaic panels. In this embodiment, the basic scene element unit in the LESS model is an Object, and each Object represents one or a group of objects. Objects are represented by triangular facets and can be one or more imported .obj files. After obtaining the aggregated photovoltaic module point cloud via the DBSCAN algorithm, the photovoltaic module point cloud needs to be fitted to Objects that can be placed in the LESS scene, i.e., simplified photovoltaic panels in .obj format.

[0138] The above steps first perform PCA dimensionality reduction on the point cloud to find three principal component directions (i.e., directions with the greatest data variation). Since photovoltaic panels typically have a thin plate-like structure, with their length and width much greater than their thickness, the first two principal components are selected to form the planar directions of the photovoltaic panel, while the third principal component direction with the least variation is ignored, thus achieving planarization of the photovoltaic panel. Then, the program calculates the boundary extrema of the point cloud in the PCA coordinate system, generates a four-vertex rectangular plane, and then inversely transforms it back to the original coordinate system, thereby obtaining the photovoltaic panel plane that best matches the point cloud distribution. This plane ignores thickness (Z-axis dimension is set to 0), retaining only the length and width dimensions, and finally outputs an .obj model and visualization results.

[0139] S103. Import the triangular mesh terrain model and the simplified photovoltaic panel model into the three-dimensional radiation transfer model LESS to construct a three-dimensional scene of the photovoltaic power station. Set the spectral attributes of the scene elements, adjust the solar zenith angle and azimuth angle parameters, and output the effective radiation value of each photovoltaic module.

[0140] In this step, the 3D radiative transfer model based on the real structure can well take into account sensor observations and the direction of solar incidence, direct and diffuse irradiance on the Earth's surface, as well as the complex 3D structure of the scene, and can accurately describe spatial heterogeneity. The 3D radiative transfer model LESS can accurately construct a realistic 3D scene in the form of triangular patches, and fully preserves the multi-scale of the 3D scene in the form of a virtual laboratory. Combined with the adjustment of the solar zenith angle and azimuth angle parameters in the model, more accurate and efficient simulation of solar radiation processes can be achieved.

[0141] 1. Import the 3D model of the photovoltaic power station into LESS.

[0142] Terrain parameters are set in the

Terrain

Terrain

Mesh

[0143] Therefore, in this study, both the terrain and the simplified photovoltaic panels were imported as scene elements, and appropriate coordinate offsets were set to ensure that the entire terrain was located above the reference grid plane.

[0144] Scene elements are added in the

Objects

[0145] The spectral properties of the terrain are set to the soil spectrum built into LESS; the spectral properties of the photovoltaic modules are set with reference to the physical characteristics of the photovoltaic modules and existing data, with the front reflectance set to 0.01, the back reflectance to 0.80, and the transmittance to 0 (Note: the specific spectral properties to be added need to be set in the

OpticalProperty

[0146] like Figure 11 As shown: The final model loaded into LESS is a simplified photovoltaic power station model. Compared to the original point cloud, the simplification is reflected in the omission of sparse vegetation and houses around the power station, retaining only the complex terrain of the triangular mesh structure and the simplified photovoltaic modules (thickness not considered). Before officially starting the downlink radiation simulation, the coordinates of scene elements (including terrain mesh and simplified photovoltaic panels) need to be adjusted to ensure that the scene in the [3Dviewer] is located above the reference mesh plane, and that the relative positions of the terrain and photovoltaic modules remain unchanged, so that they can correctly receive simulated photons.

[0147] 2. Sensor parameter settings.

[0148] Sensor parameters are set in the [Sensor] section. The downlink radiation simulation uses the PhotonTracing mode in LESS. This mode does not generate remote sensing images, so there is no need to set parameters such as image width and height; however, it is necessary to set [IlluminationResolution], i.e., the incident photon density. The smaller the Illumination value, the denser the photons, and the more accurate the results, but the slower the calculation speed; in this study, this parameter is set to 0.02.

[0149] As shown, the wavelength range of light absorbed by photovoltaic power generation is approximately 300~1200nm. Therefore, the wavelength range is set to 300 to 1200nm during simulation. Referring to existing cases, the number of wavelengths in

SpectralBands

[0150] The LESS model primarily serves vegetation remote sensing research; therefore, the simulation of the radiation received by photovoltaic modules in LESS is based on layered grouping and APAR simulation. APAR (Absorbed Photosynthetically Active Radiation) originally refers to PAR (Photosynthetically Active Radiation) absorbed by the canopy of green plants. In this study, it is equivalent to treating a single photovoltaic module as a special "vegetation component," simulating the solar radiation absorbed by each component (i.e., a single photovoltaic panel at different locations). Selecting

Products

fPAR

Layer definition

[0151] To fully cover the entire scene with significant terrain elevation differences and to differentiate the solar radiation received by photovoltaic modules at different locations, the scene layer in

LayerDefinition

[0152] 3. Lighting and atmospheric correction parameter settings.

[0153] Illumination and atmospheric parameters are set in the

Illumination & Atmosphere

Illumination

[0154]

Atmosphere

[0155] This step also explains the basic principles of LESS model performing downlink radiation simulations in complex terrain areas.

[0156] LESS is a radiative transfer model based on ray tracing of a realistic 3D structure. It simulates the transmission process of incident light in a scene (absorption, reflection, and transmission) and outputs corresponding simulation data (such as reflectivity, albedo, fPAR, etc.).

[0157] In the LESS model, forward photon tracing and backward ray tracing algorithms are implemented based on the originating position of the light rays. Forward photon tracing emits photons from the light source and simulates the propagation process of the photons throughout the scene. By recording the photon energy, the downlink radiation can be simulated. All radiation sources in LESS are directly input parameters, which can also be derived from actual measurements of solar downlink radiation.

[0158] In forward photon tracing mode, photons originate from the light source and perform intersection operations with elements in the scene. The basic principle of downlink radiation simulation in LESS is as follows: Figure 13 As shown, the entire simulation scene is divided into basic grids, each grid representing a pixel. A photon is simulated reflecting from a starting point P1 and intersecting the scene at P2. There are three cases: If P1 and P2 belong to different pixels, the upward radiation is recorded at the pixel where P1 is located (the outgoing pixel), and the downward radiation is recorded at the pixel where P2 is located (the incoming pixel); if P1 and P2 belong to the same pixel, it is considered internal scattering and is not recorded; if pixel P2 does not exist, meaning the photon left the scene after departing from P1, only the upward radiation at the pixel where P1 is located (the outgoing pixel) is recorded.

[0159] In radiation simulations, light sources typically considered include direct sunlight and diffused skylight. Direct sunlight is treated as parallel light, while diffused skylight is treated as isotropic incident light. For direct sunlight, LESS represents it as a disk that, when projected, covers the entire scene. The size and position of the disk are determined by the scene's circumsphere to ensure complete coverage of the scene by incident radiation at any angle of solar incidence. For diffused light, the photons originate on a hemisphere surrounding the entire scene, with a radius several times larger than the circumsphere.

[0160] During forward photon tracing, the APAR of the canopy's layered groups can be calculated based on the photon collision point and the local optical properties of scene elements. When a photon collides with elements in the scene (such as leaves or simplified photovoltaic panels), the amount of energy absorbed by the leaves can be calculated based on the absorptivity. Under the Lambert assumption, the absorptivity of the photovoltaic panel can be expressed as 1-τ λ -P λ , where τ λ P represents transmittance. λ This represents the absorption rate.

[0161] At each collision point, the absorbed energy of the photon is recorded as (1-τ). λ -P λ )·P Q-1 (λ). Based on the location of the photon collision point and the type of colliding element, the absorption energy of the canopy can be recorded in different layers and groups.

[0162] This invention aims to simulate the solar radiation received by a photovoltaic power station throughout the day on the summer solstice (June 21) and winter solstice (December 21) in 2024. Specifically, the simulation is performed once at the top of the hour and half-hour, and the daily variation curve of radiation is plotted. Finally, the total amount of radiation received by each photovoltaic module in a day is calculated by integrating the curves.

[0163] Visualization of radiation simulation results includes both scene and data visualization. In LESS, downlink radiation simulation values ​​accurate to a single simplified photovoltaic panel can be displayed in a 3D scene with layered color schemes. Daily radiation variation curves plotted based on simulation data can be used to calculate the daily power generation of a single photovoltaic module; comparing daily radiation variation curves of photovoltaic modules at different locations allows for in-depth exploration of the impact of terrain shading, photovoltaic module orientation, and location on power generation.

[0164] The radiation simulation results can be visualized using 3Dviewer, and the visualization effect is as follows: Figure 14-17 As shown. Since the amount of radiation received by the underlying surface is not considered (the radiation received by the ground far exceeds that of a single photovoltaic panel, affecting the layered coloring results), the terrain in the figure is shown in black. In each figure, the layered coloring is based only on the results of a single simulation, reflecting the relative magnitude of solar radiation received by each photovoltaic module in the figure.

[0165] S104. Based on the effective radiation value, plot the daily variation curve of radiation for each photovoltaic module, calculate the total daily radiation received by each photovoltaic module by integration, and then obtain the light resources of the photovoltaic power station in the target area.

[0166] In this step, the differences in the time of radiation peak occurrence, peak size, and effective light reception duration can be determined by comparing the daily variation curves of radiation of photovoltaic modules at different locations. These differences can then be correlated with the terrain undulations and relative positions between modules in the three-dimensional scene to determine the impact of terrain shading on the light reception efficiency of photovoltaic modules.

[0167] Based on multiple instantaneous simulated radiation values ​​taken every half hour, a daily variation curve of the solar radiation received by a single photovoltaic module is plotted, such as... Figure 18 As shown in the figure. By solving for the area enclosed by the curve and the coordinate axes, the total amount of radiation received by each photovoltaic module in a day can be calculated.

[0168] Figure 18This paper presents the daily solar radiation received curves of individual photovoltaic (PV) modules at five typical locations within a 3D model of a photovoltaic power plant. By combining the 3D simulation visualization results with the observed curves, a brief analysis can be made: First, due to the undulating terrain and varying orientations of the mountainous terrain, the peak solar radiation received by PV modules at different locations differs. For example, PV panel 006138 (Group 6, No. 138) reached its radiation peak at 11:00 AM. Second, micro-topography significantly affects the orientation and tilt angle of PV modules, thus greatly influencing the degree of downhill radiation received by the modules. Especially when PV modules are installed facing east or west, they suffer greater solar radiation loss. Third, the high reflectivity of the PV module backsheet significantly impacts the radiation received by surrounding PV modules. For example, PV panels located on protruding terrain have difficulty receiving reflected light from the backsheets of other PV modules, resulting in a slight reduction in their total radiation. Furthermore, compared to densely distributed PV panels, sparsely distributed PV panels generally receive a lower average level of solar radiation.

[0169] Table 1. Estimated total solar radiation received on the summer solstice (area integral, unit: kWh) This invention, through a systematic technical approach, constructs a complete 3D scene including terrain and simplified photovoltaic panels (arrays), and innovatively applies 3D radiative transfer theory to solar radiation simulation down to the component level, thus successfully achieving 3D modeling and refined downlink solar radiation simulation of a typical mountain photovoltaic power station. The research not only verifies the applicability of the LESS 3D radiative transfer model in accurate estimation of downlink radiation in photovoltaic power stations, but also establishes a set of practical and scalable technical processes and methodologies.

[0170] Using the above method, aerial photographs of the target area are acquired and a 3D point cloud is reconstructed based on the structure-reconstruction-motion algorithm. Then, 3D reconstructions of the terrain and photovoltaic (PV) modules are performed separately. A triangular mesh terrain model is generated using a cloth simulation filtering algorithm, and a simplified PV panel model is fitted using spatial clustering and principal component analysis. Both are then imported into the LESS 3D radiative transfer model to construct a 3D scene of the PV power station and set spectral attributes. Solar parameters are dynamically adjusted in the model to simulate radiative transfer, outputting the absorptive photosynthetically active radiation (SEPA) value at the PV module level. Based on multiple simulation results, the daily variation curve of radiation received by a single PV module is plotted, and the total daily received radiation is calculated by integration. This method can more accurately depict the complex terrain and spatial distribution characteristics of mountainous PV power stations, dynamically simulate the radiative transfer process under different solar angles, and effectively quantify the impact of terrain shading on the light-receiving efficiency of PV modules, thereby significantly improving the estimation accuracy of solar irradiance and power generation of mountainous PV power stations.

[0171] Secondly, the present invention also provides a photovoltaic power plant light resource estimation device, such as... Figure 19 As shown, it includes: The acquisition module 201 is used to acquire aerial images of photovoltaic power plants in the target area and reconstruct three-dimensional point clouds based on the structure-reconstruction-motion algorithm.

[0172] The reconstruction module 202 is used to separate ground points and non-ground points based on the reconstructed 3D point cloud using a cloth simulation filtering algorithm, generate a triangular mesh structure terrain model based on the ground point cloud, segment the photovoltaic module point cloud from the non-ground point cloud, process the photovoltaic module point cloud using a spatial clustering algorithm to obtain an aggregated point cloud, and fit the aggregated point cloud into a simplified photovoltaic panel model based on principal component analysis.

[0173] The simulation module 203 is used to import the triangular mesh structure terrain model and the simplified photovoltaic panel model into the three-dimensional radiation transfer model LESS to construct a three-dimensional scene of the photovoltaic power station, set the spectral attributes of scene elements, adjust the solar zenith angle and azimuth angle parameters, and output the effective radiation value of a single photovoltaic module.

[0174] Analysis module 204 is used to plot the daily variation curve of radiation of a single photovoltaic module based on the effective radiation value, calculate the total daily received radiation by integration, and thus obtain the light resources of the photovoltaic power station in the target area.

[0175] The present invention also provides a computer-readable storage medium storing a computer program that can be used to execute the above-described... Figure 1 The steps of the provided method for estimating solar resources for photovoltaic power plants.

[0176] This invention also provides a computer device. At the hardware level, the computer device includes a processor, an internal bus, a network interface, memory, and non-volatile memory, and may also include other hardware required for various operations. The processor reads the corresponding computer program from the non-volatile memory into memory and then executes it to achieve the above-mentioned functions. Figure 1 The steps of the provided method for estimating solar resources for photovoltaic power plants.

[0177] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0178] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0179] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0180] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0181] It should be noted that the above-described specific embodiments enable those skilled in the art to more fully understand the present invention, but do not limit the present invention in any way. Therefore, although the present invention has been described in detail in this specification, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the present invention; and all technical solutions and improvements that do not depart from the spirit and scope of the present invention are covered within the protection scope of the patent of the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A method for estimating solar resources in a photovoltaic power plant, characterized in that, The method includes: Aerial images of photovoltaic power plants in the target area are acquired, and 3D point clouds are reconstructed based on the structure-recovery-motion algorithm. Based on the reconstructed 3D point cloud, a cloth simulation filtering algorithm is used to separate ground points from non-ground points. A triangular mesh structure terrain model is generated based on the ground point cloud. The photovoltaic module point cloud is segmented from the non-ground point cloud. A spatial clustering algorithm is used to process the photovoltaic module point cloud to obtain an aggregated point cloud. The aggregated point cloud is then fitted into a simplified photovoltaic panel model based on principal component analysis. The triangular mesh terrain model and simplified photovoltaic panel model are imported into the three-dimensional radiation transfer model LESS to construct a three-dimensional scene of the photovoltaic power station. The spectral attributes of the scene elements are set, the solar zenith angle and azimuth angle parameters are adjusted, and the effective radiation value of each photovoltaic module is output. Based on the effective radiation value, the daily variation curve of radiation for each photovoltaic module is plotted. The total daily radiation received by each photovoltaic module is calculated by integration, thereby obtaining the light resources of the photovoltaic power station in the target area.

2. The method according to claim 1, characterized in that, After reconstructing the 3D point cloud based on the structure-of-motion (SOG) algorithm, the reconstructed 3D point cloud is preprocessed, including: Denoising of the reconstructed 3D point cloud is performed based on a pre-selected filtering algorithm; The denoised point cloud is manually cropped to remove residual noise points close to the terrain surface; The spatial thinning method is adopted to reduce the point cloud density while preserving the topographic and feature characteristics by setting a minimum point spacing.

3. The method according to claim 1, characterized in that, The method of using a cloth-simulation filtering algorithm to separate ground points from non-ground points includes: Invert the point cloud data and initialize a cloth mesh above the point cloud; The process of cloth protons falling under gravity and displacement caused by the internal forces of neighboring protons was simulated iteratively. In each iteration, the current position of the cloth proton is compared with the elevation value of the inverted point cloud, and when the proton touches the point cloud, it is set to an immovable state. After the iteration terminates, the distance between each point cloud data and the final cloth model is calculated, and the point cloud is classified into ground points and non-ground points based on a preset distance threshold.

4. The method according to claim 1, characterized in that, A triangulation model based on ground point clouds is generated using the Delaunay triangulation algorithm, including: The 3D ground point cloud is projected onto the best plane fitted by the least squares method, and the Delaunay triangulation is completed in the 2D domain. Based on the spatial adjacency relationships of the original 3D points, the 3D triangular mesh topology is reconstructed to generate a triangular mesh terrain model.

5. The method according to claim 1, characterized in that, The process of using spatial clustering algorithms to process photovoltaic module point clouds to obtain aggregated point clouds includes: The DBSCAN algorithm is used to perform cluster analysis on the point cloud of a single photovoltaic module, identify and remove discrete noise points that are attached to the terrain, and retain the cluster with the most points as the aggregated point cloud of the photovoltaic module.

6. The method according to claim 1, characterized in that, The process of fitting the aggregated point cloud into a simplified photovoltaic panel model based on principal component analysis includes: Principal component analysis was performed on the aggregated point cloud of photovoltaic modules to determine the directions of its three principal components; The first two principal components are selected to form the plane of the photovoltaic panel, while the third principal component, which has the smallest change in the thickness direction, is ignored. The boundary of the point cloud in the principal component coordinate system is calculated, a rectangular plane ignoring thickness is generated, and then the plane is transformed back to the original coordinate system through inverse transformation, thereby completing the fitting of the simplified photovoltaic panel model.

7. The method according to claim 1, characterized in that, After obtaining the daily radiation variation curve for each photovoltaic module, the method further includes: By comparing the daily variation curves of radiation of photovoltaic modules at different locations, the differences in the time of radiation peak occurrence, peak size and effective light reception duration were determined, and the correlation was made with the terrain undulations and relative positions between modules in the three-dimensional scene, thereby determining the impact of terrain shading on the light reception efficiency of photovoltaic modules.

8. A photovoltaic power plant solar resource estimation device, characterized in that, The device includes: The acquisition module is used to acquire aerial images of photovoltaic power plants in the target area and reconstruct 3D point clouds based on the structure-reconstruction-motion algorithm. The reconstruction module is used to separate ground points from non-ground points based on the reconstructed 3D point cloud using a cloth simulation filtering algorithm, and generate a triangular mesh structure terrain model based on the ground point cloud; it segments the photovoltaic module point cloud from the non-ground point cloud, processes the photovoltaic module point cloud using a spatial clustering algorithm to obtain an aggregated point cloud, and fits the aggregated point cloud into a simplified photovoltaic panel model based on principal component analysis. The simulation module is used to import the triangular mesh structure terrain model and the simplified photovoltaic panel model into the three-dimensional radiation transfer model LESS to construct a three-dimensional scene of the photovoltaic power station, set the spectral attributes of scene elements, adjust the solar zenith angle and azimuth angle parameters, and output the effective radiation value of a single photovoltaic module. The analysis module is used to plot the daily variation curve of radiation of a single photovoltaic module based on the effective radiation value, calculate the total daily received radiation by integration, and thus obtain the light resources of the photovoltaic power station in the target area.

9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the method described in any one of claims 1 to 7.

10. A computer device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described in any one of claims 1 to 7.

Images