Unmanned aerial vehicle system for removing accumulated snow on roof solar photovoltaic array

Abstract

The invention discloses an unmanned aerial vehicle system for removing accumulated snow on a roof solar photovoltaic array, and relates to the technical field of new energy equipment, and the system comprises a multi-source sensor network module which carries out the comprehensive scanning of a photovoltaic array coverage area through the deployment of a multi-source sensor network, obtains the original data information of accumulated snow coverage, and transmits the original data information to an unmanned aerial vehicle. The data comprises accumulated snow thickness and preliminary distribution characteristics of different areas, and obtaining a preliminary accumulated snow coverage data set; according to the unmanned aerial vehicle system for removing the accumulated snow on the roof solar photovoltaic array, the problem of removing the accumulated snow on the photovoltaic array is effectively solved, the removing efficiency and precision are improved, the labor cost is reduced, and a guarantee is provided for stable operation of a photovoltaic power generation system.

Description

Technical Field

[0001] The present invention relates to the technical field of new energy equipment, and in particular to a drone system for clearing snow from a rooftop solar photovoltaic array. Background Art

[0002] Snow accumulation has a crucial impact on the operational efficiency of large-scale photovoltaic arrays, especially in cold regions. Snow cover not only reduces power generation efficiency but can also cause long-term damage to equipment. Therefore, research on efficient snow removal has become a key area for ensuring the stable operation of the energy industry. Research in this area is directly related to the sustainable use of new energy equipment and plays a significant role in improving energy security and economic efficiency.

[0003] However, current snow removal methods rely heavily on manual inspections or simple mechanical equipment, resulting in inefficiencies and limited coverage. This is particularly true in large, complex terrain scenarios, where traditional methods struggle to achieve comprehensive and accurate snow management. These limitations make snow removal often time-consuming and labor-intensive, with limited effectiveness. Against this backdrop, the field faces significant technical challenges. The primary challenge is accurately measuring the distribution of snow in different areas. Snow accumulation often has uneven distributions in terms of size and density. Without accurate data, effective snow removal strategies are difficult to formulate. This challenge further complicates the process of converting measured data into three-dimensional distribution maps to visually visualize the actual snow accumulation. Only by successfully visualizing this data can a reliable basis be provided for subsequent route planning. These two technical factors are closely intertwined. Inadequate measurement accuracy directly impacts the accuracy of the three-dimensional maps, limiting the systematic and efficient nature of snow removal operations. Summary of the Invention

[0004] The purpose of the present invention is to provide a drone system for clearing snow from rooftop solar photovoltaic arrays. By accurately measuring the span and distribution density of snow and generating an accurate three-dimensional distribution map, it provides data support for the drone to formulate the optimal clearing path and operation sequence.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a drone system for clearing snow from a rooftop solar photovoltaic array, the system comprising:

[0006] The multi-source sensor network module deploys a multi-source sensor network to conduct a comprehensive scan of the photovoltaic array coverage area to obtain raw data information on snow cover, including the thickness and preliminary distribution characteristics of snow in different areas, and obtain a preliminary snow cover dataset;

[0007] The path planning module determines a preliminary clearing operation path plan based on the preliminary snow cover dataset. Based on the preliminary clearing operation path plan, it dynamically adjusts the path based on terrain obstacles and snow distribution density. If snow areas with density higher than a preset threshold are detected in the path, the flight altitude and operation sequence of the UAV are preferentially adjusted to obtain the optimized operation path.

[0008] The operation execution and feedback collection module uses the optimized operation path to obtain real-time environmental feedback data from the drone during the clearing operation. It analyzes the residual snow in the feedback data and determines whether a second clearing operation is required in the area.

[0009] The secondary clearing path planning module analyzes the results of real-time environmental feedback data. If it determines that the residual snow in a local area exceeds a preset threshold, it will re-plan the local path in the area and determine the order of secondary clearing operations;

[0010] The control system module transmits the adjusted path information to the UAV control system through the secondary clearing operation sequence, performs precise operations on the residual snow in the local area, and obtains the final clearing operation completion data.

[0011] Preferably, determining a preliminary clearing operation path plan based on the preliminary snow cover dataset includes:

[0012] Based on the preliminary snow cover dataset, image processing technology is used to denoise and extract features from the data, and classification processing is performed on the snow distribution characteristics of different regions to determine a processed snow distribution feature dataset.

[0013] Preferably, determining a preliminary clearing operation path plan based on the preliminary snow cover dataset further comprises:

[0014] By inputting the processed snow distribution feature dataset into a pre-established 3D modeling module, spatial mapping is performed on the span and density information of the snow distribution to generate the corresponding 3D visualization graphics of the snow distribution.

[0015] Preferably, determining a preliminary clearing operation path plan based on the preliminary snow cover dataset further comprises:

[0016] Based on the generated three-dimensional visualization of snow distribution, the features of areas with snow density higher than the preset threshold and uneven distribution are extracted, and priority marking is performed on the characteristic areas to obtain a snow distribution graphic with priority labels.

[0017] Preferably, determining a preliminary clearing operation path plan based on the preliminary snow cover dataset further comprises:

[0018] By using snow distribution graphics with priority labels and combining them with environmental data of complex terrain, a path optimization algorithm is used to plan the drone's operation path and determine the preliminary clearing operation path plan.

[0019] Preferably, the specific formula for dynamically adjusting the terrain obstacles and snow distribution density in the path according to the preliminary clearing operation path plan is:

[0020] ;

[0021] in, represents the total path cost, Indicates the The flight length of the segment path, Indicates the The obstacle density on the segment path, Indicates the The thickness of snow on the path, Represents the path segment index, Represents the path segment index, represents the weighting coefficient of the path length, represents the weighting coefficient of obstacle density, Indicates the weighting factor of snow thickness.

[0022] Preferably, the weighted coefficient of the path length The specific formula is;

[0023] = ;

[0024] in, represents the weighting coefficient of the path length, Indicates the current remaining battery power of the drone. Indicates the upper limit of the drone battery capacity. Indicates the average wind speed in the current area. Indicates the maximum safe wind speed of the drone. Represents the base of natural logarithms.

[0025] Preferably, the weighted coefficient of the obstacle density is The specific formula is:

[0026] = ;

[0027] in, represents the weighting coefficient of obstacle density, Represents the total projected area of ​​obstacles on the current path segment, represents the total traversable area of ​​the path segment, Indicates the shortest distance between the centerline of the path segment and the nearest obstacle, Indicates the safe flight control radius of the drone;

[0028] described =1- - ;

[0029] in, represents the weighting coefficient of the path length, represents the weighting coefficient of obstacle density, Indicates the weighting factor of snow thickness.

[0030] Preferably, the multi-source sensor network module includes a lidar unit, an infrared thermal imaging sensor unit and a high-definition camera unit. The lidar unit is used to scan the three-dimensional contour of the photovoltaic panel and the thickness of snow. The infrared thermal imaging sensor is used to identify areas with abnormal temperature differences. The high-definition camera is used to obtain visible light images for image recognition and auxiliary verification.

[0031] Preferably, the path planning module includes an obstacle identification processing unit and a dynamic cost weight evaluation unit. The obstacle identification processing unit is used to analyze the position, size and spatial distribution of obstacles above and around the photovoltaic array. The dynamic cost weight evaluation unit dynamically calculates the comprehensive cost of the path segment based on the current remaining power of the drone, the wind speed level in the flight area and the density of snow removal tasks covered by the unit path.

[0032] It can be seen from the above technical solution that the present invention has the following beneficial effects:

[0033] This unmanned aerial vehicle (UAV) system for clearing snow from rooftop solar photovoltaic arrays deploys a multi-source sensor network to acquire snow cover data, uses image processing technology for feature extraction and classification, and generates a three-dimensional visualization of snow distribution. Incorporating data from complex terrain environments, the system employs a path optimization algorithm to plan the UAV's clearing operation path and dynamically adjusts the operation sequence based on real-time feedback. For residual snow, the system can also perform localized secondary clearing for precise operation. This method effectively addresses the challenges of snow removal from photovoltaic arrays, improving clearing efficiency and accuracy, reducing labor costs, and ensuring the stable operation of photovoltaic power generation systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 This is a system connection diagram of the present invention. DETAILED DESCRIPTION

[0035] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0036] like Figure 1 As shown, the present invention provides a technical solution: a drone system for clearing snow from a rooftop solar photovoltaic array, the system comprising:

[0037] The multi-source sensor network module deploys a multi-source sensor network to conduct a comprehensive scan of the photovoltaic array coverage area to obtain raw data information on snow cover, including the thickness and preliminary distribution characteristics of snow in different areas, and obtain a preliminary snow cover dataset;

[0038] The path planning module determines a preliminary clearing operation path plan based on the preliminary snow cover dataset. Based on the preliminary clearing operation path plan, it dynamically adjusts the path based on terrain obstacles and snow distribution density. If snow areas with density higher than a preset threshold are detected in the path, the flight altitude and operation sequence of the UAV are preferentially adjusted to obtain the optimized operation path.

[0039] The operation execution and feedback collection module uses the optimized operation path to obtain real-time environmental feedback data from the drone during the clearing operation. It analyzes the residual snow in the feedback data and determines whether a second clearing operation is required in the area.

[0040] The secondary clearing path planning module analyzes the results of real-time environmental feedback data. If it determines that the residual snow in a local area exceeds a preset threshold, it will re-plan the local path in the area and determine the order of secondary clearing operations;

[0041] The control system module transmits the adjusted path information to the UAV control system through the secondary clearing operation sequence, performs precise operations on the residual snow in the local area, and obtains the final clearing operation completion data.

[0042] This system systematically addresses the issue of winter snow accumulation affecting the power generation efficiency of rooftop photovoltaic arrays by building a multi-module collaborative structure. First, the multi-source sensor network module deploys thermal infrared sensors, lidar, high-definition cameras, and ambient temperature and humidity sensors above or around the photovoltaic array to achieve multi-dimensional and multi-angle sensing of the surface conditions of the photovoltaic panels. Thermal infrared sensors are used to determine surface temperature differences and the presence of ice and snow. Lidar is used to obtain high-precision snow thickness and morphological characteristics. High-definition images are used for computer vision to identify the extent of snow cover. Data fusion algorithms on edge devices uniformly convert data from different types of sensors into a two-dimensional snow distribution map and a three-dimensional snow thickness model, forming a preliminary snow cover dataset.

[0043] Based on the aforementioned dataset, the path planning module first performs obstacle modeling and PV panel edge extraction, then compares geographic information system (GIS) data with a map for positioning. An initial path is then generated using the A* path algorithm, which optimizes a cost function. This cost function combines snow thickness, terrain slope, and an energy consumption model to dynamically avoid obstacles and prioritize coverage of high-density snow areas. The path planning process also determines in real time whether any areas within the path have snow density exceeding a preset upper limit for clearing efficiency. If so, the clearing sequence is prioritized and the drone's altitude is adjusted to the optimal clearing distance to maximize the impact of the jet or brushing efficiency.

[0044] The job execution and feedback acquisition module collects high-frequency image streams and temperature data in real time while the drone is performing its job. The edge computing module then performs fast image comparison and residue identification algorithms, such as using lightweight object detection networks like YOLOv5 to identify uncleared areas. Residual areas are dynamically analyzed based on before-and-after snow removal difference maps. If the residual snow area or thickness exceeds a set threshold, the system invokes the secondary clearing path planning module to regenerate and optimize local paths based on the residual distribution and determine clearing priorities.

[0045] The control system module connects to the drone's flight control system via high-speed wireless communication, issuing real-time commands for altitude, speed, and operation methods (such as heating, air jets, and sweeping). It also receives feedback to adjust the path and actions in real time. Upon completion, the system records the entire cleaning operation path, time, energy consumption, and feedback data, creating a complete operation log for subsequent optimization and tracking.

[0046] This system aims to solve the problem of reduced power generation efficiency of rooftop photovoltaic arrays caused by snow cover. It adopts a modular intelligent drone system architecture and consists of five main functional modules. The modules are coordinated and linked to achieve accurate identification, dynamic path planning, operation feedback control and residual snow removal.

[0047] First, during the data collection phase, the multi-source sensor network module deploys thermal imaging sensors, lidar, high-resolution cameras, and temperature and humidity monitors to comprehensively scan the entire coverage area of ​​the photovoltaic array. Thermal imaging sensors detect surface temperature differences and identify low-temperature areas, indicating possible snow accumulation. Lidar scans the photovoltaic panel surface using high-frequency points, and by comparing this with an initially established snow-free height model, it infers the current snow depth. High-definition cameras capture visual images and extract the boundaries of snow-covered areas through image grayscale changes and edge recognition algorithms. The data collected by these various sensors is then fused within the system to form a two-dimensional snow distribution map and a three-dimensional snow thickness model, which serve as input for the subsequent path planning module.

[0048] During the path planning stage, the system first extracts the outline of the photovoltaic array based on the snow cover data, and identifies inaccessible areas such as roof obstacles, structure edges, protrusions, and generates a basic geometric model. Then, based on the known take-off point, clearing area range, snow thickness and density information, the system uses a heuristic search algorithm to perform preliminary operation path planning. During the path planning process, the system will give priority to clearing areas where the snow thickness is greater than a certain threshold. This threshold is usually set to two centimeters, which is derived from the minimum obstruction impact on power generation efficiency standard provided by photovoltaic panel manufacturers. At the same time, if the snow thickness in some areas is significantly higher than other areas, the system will arrange for drones to prioritize these areas and automatically adjust the flight altitude during the path to keep the clearing equipment at the optimal cleaning distance to improve efficiency and reduce energy consumption.

[0049] During the operation execution phase, the drone performs snow removal tasks according to the optimized path, using an integrated jet device, heating device or mechanical brush sweeping device to clear snow. After completing each operation path, the drone automatically hovers and re-activates the sensor to collect images and temperature data after the clearance. The system compares the images before and after the operation, and by analyzing the proportion of white areas remaining in the images and the temperature difference changes, it determines whether there is any residual snow in the area. If the system determines that there is still significant residual snow in a certain area, and the residual area exceeds 10% of the total area of ​​the area, it is considered that the initial clearance did not achieve the expected results and a second operation is required.

[0050] Next, the system activates the secondary clearing path planning module. This module remaps and optimizes paths only for the residual areas. By comparing the snow thickness and density of the residual distribution in each residual area, the system determines a new clearing priority and reroutes the drones based on the severity of the residual snow. During this process, the system not only optimizes the clearing order but also fine-tunes the flight altitude, operating speed, and clearing method based on local environmental changes, such as switching the jet to heating mode to improve clearing effectiveness.

[0051] The control system module runs the entire execution process. This module wirelessly transmits real-time generated paths and operational instructions to the drone's flight control system, orchestrating the drone's movement in three-dimensional space, starting and stopping operational equipment, and adjusting flight attitude. The control system also receives real-time sensor feedback to monitor flight path deviations and anomalies such as wind speed disturbances and ice and snow shedding, instantly adjusting instructions if errors occur. After completing all clearing tasks, the control system collects path data, energy consumption data, and feedback results, generating a clearing log for back-end archiving and optimization analysis.

[0052] One of the system's key technologies is a dynamic feedback closed-loop control strategy. This strategy uses feedback images and environmental data after each operation to continuously learn and optimize paths, ensuring high responsiveness and adaptability throughout the entire operation process. The snow removal module in the drone system also supports different operating parameter configurations. For example, in jet mode, the jet pressure is adjustable to accommodate varying snow densities; the heating mode supports temperature control to prevent thermal damage to the photovoltaic panels; and the brushing device supports speed adjustment to match varying snow hardness. All configuration parameters can be dynamically set by the control system based on real-time feedback, ensuring optimal snow removal for each photovoltaic panel.

[0053] The entire system, from perception, analysis, decision-making to execution, is based on quantifiable parameters and known algorithms. It has a complete technical closed loop and high feasibility, and can achieve stable and effective snow removal operations under different roof structures, different weather conditions and different photovoltaic module layouts.

[0054] During system initialization, the multi-source sensor network module initiates data collection. The LiDAR uses high-frequency point cloud scanning to obtain the three-dimensional spatial position of each coordinate point on the PV array surface, combining timestamps and coordinate systems to generate a height model. A pre-installed reference map of the PV array height in a "snow-free state" is stored permanently within the system, acquired during initial deployment through flight scan recordings. During each run, the system compares the current LiDAR-measured point height with the reference height. If the surface exceeds two centimeters above the reference level, the system deems snow accumulation to be present. Furthermore, the system incorporates image recognition algorithms to process camera image frames, employing edge detection and region growing algorithms to identify bright areas in the image—areas typically caused by snow reflections. The snow area outlines obtained by image recognition are then matched and fused with the LiDAR-generated height difference map, using a confidence-weighted average to generate a unified snow map. Simultaneously, the thermal imaging module detects areas with temperatures below 0°C and, by overlaying the image with the thermal map, further eliminates misidentified areas, such as false snow caused by reflections from the PV panel edges or shading. Finally, the system generates a grid map with a resolution of 0.1 meters. Each grid cell records the snow thickness, coverage status and confidence level of the corresponding area, providing high-precision input for the path planning module.

[0055] After receiving the snow distribution map, the path planning module first analyzes the roof structure and photovoltaic array layout to construct a binary map of traversable and obstructed areas. The map is then divided into several processing units, each with a corresponding "operation cost" that takes into account snow depth, estimated regional energy consumption, path length, and the degree of interference from adjacent obstacles. The system specifies the starting point as the drone's initial position, and the target point as the traversal set of all heavily snowed areas. Path search is implemented using a modified A* algorithm. During each node expansion, areas with high clearance priority (i.e., thicker or denser snow) are assigned lower path costs to guide the search path toward key areas. Furthermore, the system penalizes changes in turning angles to avoid increased energy consumption and decreased efficiency due to frequent turns. After the search is complete, a set of preliminary path points is generated. The system then applies a B-spline fitting algorithm to smooth the path, ensuring that it complies with the drone's dynamic constraints, such as the minimum turning radius and maximum climb rate. Furthermore, based on the snow cover conditions in each path segment, the system automatically matches the clearing mode (such as heating, air jets, or brushing) and configures the operation parameters, ensuring software and hardware adaptation of the planning results.

[0056] After completing the initial round of operations, the system immediately triggers a post-operation evaluation process. By comparing pre- and post-operation images in a temporal sequence, the system uses image difference and grayscale change analysis to detect areas of residual snow. If a region's brightness, edge texture, or infrared temperature difference does not change significantly, the system labels it as a "possible residual snow area." After all such areas are included in the candidate set, the system further merges regions, removing isolated, misjudged points and retaining only those with continuity and an area exceeding a minimum threshold (e.g., 0.5 square meters). If the total area of ​​residual snow exceeds 10% of the PV array's active area, the system automatically enters a secondary clearance path planning process. During this process, the system constructs a new high-density cost graph only for the residual snow areas and applies a localized path search algorithm (such as a heuristic greedy search) to plan a coverage path. Prioritizing areas with significant residual snow creates a localized secondary operation path. Simultaneously, the system adjusts parameters based on feedback from the initial operation, such as increasing jet power, extending heating time, or reducing flight speed, to improve secondary clearance effectiveness.

[0057] The control system module is responsible for converting the output of the path planning module into low-level instructions executable by the flight control system. This module first parses each path segment, extracting information such as the flight start and end points, terrain characteristics of the area, and operational requirements. This information is then matched against the UAV platform's performance parameters (such as maximum speed, permitted altitude, and payload) to generate a time-stamped list of flight maneuvers. The control system transmits these instructions to the flight control system in real time via wireless communication interfaces (such as 5G modules, LoRa, or Wi-Fi links). Simultaneously, the control system receives feedback from the UAV, including sensor data, operational status, and GPS trajectory data, from the ground station. If issues such as path deviation, attitude anomalies, or external wind disturbances occur during flight, the system automatically adjusts the flight attitude or route without interrupting operations. For operational modules (such as sweeping, air jets, or heating devices), the control system also simultaneously issues control commands for start / stop, power adjustment, and frequency setting, enabling dynamic configuration and coordinated control of physical operational actions. The control system also records a full operational log, including key metrics such as path execution, anomaly flags, and energy consumption, to support subsequent optimization analysis and training data accumulation.

[0058] Compared to traditional manual or timed automated systems, this drone system achieves intelligent, automated, and refined snow removal operations. On the one hand, multi-source sensor data fusion enables comprehensive perception of snow accumulation, improving path planning accuracy. On the other hand, a feedback-based path adjustment mechanism enables closed-loop control of operations, significantly reducing the rate of snow omission. A secondary clearing mechanism ensures thorough clearing of complex areas, and the precise command transmission of the control module further enhances the system's response speed and operational efficiency. Overall, the system improves the photovoltaic array's power generation efficiency and operational safety in winter, reducing maintenance costs and labor risks.

[0059] Based on the preliminary snow cover dataset, a preliminary clearing operation path plan is determined, including using image processing technology to denoise and extract features from the data based on the preliminary snow cover dataset, classifying the snow distribution characteristics of different areas, and determining a processed snow distribution feature dataset.

[0060] After acquiring the preliminary snow cover dataset, the system needs to perform a series of standardized image processing and structured analysis on the raw image data to further improve the effectiveness and pertinence of the clearing path planning. First, in the image preprocessing stage, the system performs a noise suppression operation by sorting all the grayscale values ​​in a three-by-three neighborhood around each pixel and selecting the value in the middle as the new value of the pixel. This method is called median filtering. This process can effectively remove bright or dark spots in the image caused by sensor interference, while preserving the image edge structure, which helps the accuracy of subsequent boundary detection. In some images, in order to further suppress background changes and retain structural details, the system will use bilateral filtering. This method combines the physical distance between pixels with the grayscale value difference to avoid over-blurring the edges.

[0061] Next, the image feature extraction process begins. The system uses an image gradient algorithm to analyze the horizontal and vertical grayscale changes of each pixel in the image, and then determines the area with the most dramatic grayscale changes in the image, which is also the image edge. The above operation helps the system determine the boundary shape and range of the snow-covered area. Then, using edge connection analysis technology, closed figures composed of closed boundary lines are identified as candidate snow-covered areas. For each candidate area, the system extracts its average brightness, contour area, and boundary morphological parameters, such as boundary tortuosity or the presence of sudden angles.

[0062] After completing the candidate area construction and feature extraction, the system needs to classify the snow-covered areas. This step uses a cluster analysis method to divide all candidate areas into three categories: light, moderate and heavy under the three categories set by the system. The classification is based on the combined similarity of features such as regional brightness, area and boundary complexity. In order to improve the accuracy of the classification results, the system performs dimensionality reduction on all regional features before classification, compressing multi-dimensional attributes into core information dimensions to facilitate more efficient clustering operations. After the classification is completed, the system assigns a clearing priority level to each type of area. For example, the heavily snowed area has the highest priority, which is used to guide the subsequent operation sequence.

[0063] In addition, the system sets a threshold judgment mechanism at specific key processing points to trigger strategy switching. For example, the number of pixels in the image with a grayscale value greater than 200 (based on the grayscale range of 0 to 255 for 8-bit images) is counted. When this number exceeds 15% of the total number of pixels in the image, the system determines that the current image is in a high-density snow field state. In this case, a more sophisticated edge detection algorithm and area refinement strategy will be automatically enabled. The determination of this percentage threshold is based on the analysis of a large number of actual scene image samples. It is adjusted by comparing the proportion of high-reflectivity areas with the consistency of manually marked snow areas, and finally the optimal boundary value is selected that can accurately distinguish heavy snow areas in most typical environments.

[0064] Finally, the system labels each image region as a data unit with a clear thickness grade, boundary delineation, area statistics, and classification label. All units are integrated to generate a standardized snow distribution feature dataset, which serves as direct input to the path planning module. During subsequent path generation, the system uses this dataset to perform area sorting, operational strategy matching, and path point weight calculations, achieving seamless integration and automatic optimization from data processing to path execution.

[0065] By introducing image processing and feature extraction, preliminary snow accumulation data is denoised and classified, enabling the subsequent path planning module to perform strategy optimization based on the actual, clearly structured snow layer distribution. Compared to traditional path planning methods based on raw thickness data, this solution can significantly improve the rationality and efficiency of operation paths when dealing with complex snow accumulation patterns and areas with discontinuous boundaries, effectively reducing repeated clearing and path blind spots. The classification results can also support subsequent resource allocation, such as adjusting operation time and clearing modes in different areas, thereby further improving the overall system operation efficiency and energy utilization.

[0066] Based on the preliminary snow cover dataset, determining the preliminary clearing operation path plan also includes inputting the processed snow distribution characteristic dataset into a pre-established 3D modeling module to spatially map the span and density information of the snow distribution to generate a corresponding 3D visualization graph of the snow distribution.

[0067] After acquiring and classifying the snow distribution characteristics of the rooftop photovoltaic array, the system further inputs this structured data into the 3D modeling module, generating a visual representation of the spatial morphology of the snow accumulation to assist in subsequent path planning, key area identification, and operational strategy optimization. The 3D modeling module first utilizes an established basic model of the rooftop photovoltaic structure. This model is generated using 3D point cloud data acquired by the drone's initial deployment, using a LiDAR scan. A surface reconstruction algorithm then creates a triangular mesh structure, accurately representing the 3D geometry of the rooftop profile, slope angle, photovoltaic panel arrangement, and surrounding obstructing structures.

[0068] The system converts the center coordinates and boundary contours of each snow-covered area extracted from the processed two-dimensional image into spatial coordinates corresponding to the three-dimensional model through proportional conversion. This conversion is based on the mapping relationship established between the resolution of the sensor at the time of acquisition and the actual size of the surveying area, ensuring that the image data can accurately fall on the correct position on the surface of the three-dimensional model. For each successfully mapped three-dimensional grid unit, the system increases the vertical height value of the grid according to the snow thickness value of the corresponding area. For example, when the original height of a grid is the roof base height and the snow thickness value is 5 cm, the system will "raise" the grid surface by 5 cm in the three-dimensional visualization to reflect the physical occupancy of the actual snow layer.

[0069] If the snow depth in certain areas exceeds a preset threshold (e.g., 15 cm), the system will not continue with linear stretching. Instead, it will use a warning texture icon (e.g., a red outlined block) to avoid distortion. This thickness threshold was determined based on experimental findings: when the snow depth exceeds 15 cm, vertical changes in the graphics are no longer sufficient to convey the difficulty of clearing, and are actually detrimental to overall visual interpretation. Therefore, using graphic symbols to highlight operational risks is more practical and instructive.

[0070] In terms of expressing snow density, the system assigns a density score to each area, which is derived from the regional grayscale, temperature difference and structural feature assessment during the image analysis phase. The system divides the density score into three levels and sets three color codes accordingly. Low-density areas are presented in light blue and given a higher degree of transparency; medium-density areas use medium blue and reduce transparency; and high-density areas are presented in dark blue or even red and are completely opaque. This color and transparency mapping relationship is a linear relationship. The system supports users to adjust the color threshold and gradient curve through the interface to adapt to different display needs and visual preferences.

[0071] In order to further evaluate the difficulty of the operation and the complexity of the path, the system calculates the maximum spatial span of the snow-covered area. The specific approach is to perform geometric envelope processing on the boundary points of each snow-covered area, that is, to find the minimum closed figure containing all the boundary points, and then measure the actual spatial distance between the two farthest points in the figure as the maximum span value of the area. If the span value exceeds the continuous coverage threshold set by the system (such as 2 meters), the system will mark the area as a "high continuous snow barrier area" for priority processing during path planning. This threshold is set based on the drone's spray range and the width of the flight stability zone. Taking the effective width of the drone's spraying as 0.5 meters and the flight safety buffer distance of 1 meter as an example, continuous coverage of more than 2 meters will significantly increase the path overlap rate and resource consumption, so it should be identified as a key clearing area.

[0072] After modeling is complete, the system renders the 3D graph into an interactive view. Users can use a mouse or touchscreen to zoom in, rotate, and zoom in on specific areas, even slicing them to view the density gradient and thickness overlay effects of specific regions. This 3D graph is used not only for manual pre-assessment but also as input to the path planning algorithm, which calculates the shortest coverage path, obstacle avoidance paths, and multi-machine collaborative clearing zones during the path simulation phase.

[0073] Finally, the module realized the reconstruction from two-dimensional image snow data into a real three-dimensional space scene, and has the functions of intuitively presenting thickness expression, density expression and distribution pattern. It is an indispensable basic module for realizing high-precision system pre-operation state perception, path decision optimization and risk area prompts.

[0074] By introducing a 3D snow distribution modeling and visualization module, the system provides intuitive and accurate spatial information for path planning. This effectively avoids the limitations of 2D data in complex roof structures and enhances data comprehension and spatial judgment. Especially in situations with sloping roofs, irregular layouts, and obstructing structures, 3D graphics accurately reflect actual distribution and potential risk areas, assisting with obstacle avoidance, routing, and dynamic adjustment of clearing strategies, further improving the overall system's safety, accuracy, and efficiency.

[0075] Based on the preliminary snow cover dataset, determining the preliminary clearing operation path plan also includes extracting the features of areas with snow density higher than a preset threshold and uneven distribution based on the generated three-dimensional visualization of snow distribution, prioritizing the characteristic areas, and obtaining a snow distribution graph with priority labels.

[0076] After generating a 3D visualization of snow distribution, the system further analyzes all modeled snow areas to identify those requiring priority removal. To achieve this, the system uses two key criteria: whether the snow density within an area exceeds a pre-set risk threshold, and whether the spatial uniformity of the snow distribution within that area is abnormal.

[0077] In the first step, the system will traverse each three-dimensional grid unit in turn and judge its corresponding snow density score. The snow density score is a continuous value between 0 and 1, which is derived from a comprehensive evaluation of factors such as image grayscale, thermal infrared temperature difference, regional brightness distribution and boundary characteristics, representing the compactness of the snow layer per unit area of ​​the grid. The density threshold set by the system is 0.7, that is, when the score of a certain grid unit is higher than 0.7, the area is classified as a "high-density snow area". This value comes from a large amount of statistical analysis of experimental data. Specifically: when the snow density score exceeds 0.7, the average remaining snow area after a single clearing exceeds 20% of the original area, and the clearing efficiency is significantly reduced, so priority identification and arrangement are required.

[0078] In the second step, the system analyzes the spatial uniformity of snow distribution. The method is: with each target grid as the center, a local analysis window is constructed consisting of the grid and its 25 neighboring grids above, below, left, right, front and back. Within this window, the system counts the density scores of all grid cells and calculates the average and dispersion of these scores. If the dispersion of the density score, that is, the amplitude of the score fluctuation is greater than 0.08, the system will identify the area as an "uneven snow distribution area". This threshold is also obtained from the sample training set and is summarized by comparing the difference distribution of snow layer heat maps of a large number of complex roofs to ensure that the accumulation of snow in mutation areas or edge areas can be effectively identified, such as ridges, air duct outlets, and next to sunshade structures.

[0079] Next, the system performs a spatial connectivity analysis on the identified high-density and uneven areas, merging cells with the same type of label in adjacent grids into a single, integrated region. The system uses an eight-way adjacency rule to determine whether grids are connected. If so, they are grouped together under the same region identifier. Each merged region is assigned a unique number by the system, and a structured data object is constructed containing the region number, total area, average density value, and boundary shape information.

[0080] The system then assigns priority scores to these areas. The scoring method is based on two main indicators: one is the average snow density value of the area. The higher the value, the higher the score; the other is the spatial position weight of the area in the entire photovoltaic array. If the area is in a key area such as the central main channel, close to the junction box or boundary opening, the system will give additional weight bonus. For example, high-density snow blocks near the center of the array are more likely to cause the failure of the entire group of panels, and their priority is naturally higher than that of the edge areas. The final score is used to determine the clearance level for each area. The system usually sets three levels of clearance priority: Level 1 is a high-risk area that needs to be cleared first; Level 2 is a less important area that can be arranged for immediate processing after the first round; Level 3 is a low-risk area that is cleared when resources permit.

[0081] The system labels each area's priority within the original 3D graphic through color coding, transparency changes, or border styles. It also generates a list of all labeled areas for use by the path planning module, scheduling module, and user interface. The resulting prioritized 3D graphic not only guides the optimization of drone clearance sequences during path generation but also provides a foundation for task allocation among multiple drones, supporting the implementation of collaborative clearance strategies.

[0082] By accurately extracting high-density snow areas and unevenly distributed areas within the 3D graphics and implementing priority labeling, the system can more effectively allocate drone operational resources, achieving a shift from "uniform operation" to "tiered, differentiated operation." This approach significantly improves overall operational efficiency, reducing duplicate coverage of non-critical areas, thereby reducing energy consumption, extending operation time, and improving the completeness of clearing and power generation efficiency of the photovoltaic array. Priority labeling also facilitates task coordination among multiple drones, enhancing the overall intelligent scheduling of the system.

[0083] Based on the preliminary snow cover dataset, determining the preliminary clearing operation path plan also includes planning the UAV's operation path through a snow distribution graph with priority labels, combined with environmental data of complex terrain, and using a path optimization algorithm to determine the preliminary clearing operation path plan.

[0084] After the system completes the construction of the snow distribution graph with priority labels, it enters the path optimization stage. The system will comprehensively consider the clearing priority of the snow area in the current three-dimensional graph, the actual terrain environment of the roof, the distribution of obstacles and the operating capabilities of the drone platform, and generate a preliminary clearing operation path through the path optimization algorithm.

[0085] First, the system sorts each snow zone according to its priority tag. Clearance is divided into three levels: Level 1 for high-risk areas that must be cleared first, Level 2 for medium-density areas, and Level 3 for general areas. Based on the tag, the system prioritizes Level 1 areas for the start of the route, ensuring that operations prioritize snow areas that most impact PV power generation efficiency.

[0086] Next, the system loads the rooftop environmental data, including the layout of the photovoltaic panels, the roof's slope angle, the height and structure of the mounting brackets, wind direction and speed distribution, the location of adjacent edges, and the coordinates of obstacles such as chimneys, vents, and skylights. This data is typically derived from a building information model generated by drone 3D scanning during initial deployment, or imported from construction drawings and a GIS (Geographic Information System).

[0087] The core of this path adjustment mechanism is the introduction of a path cost function, which dynamically determines the quality of each path segment and guides the path planning algorithm toward the path with the lowest cost. The system breaks down the initial path into several segments, each of which corresponds to a set of parameters: the first is the flight length, which is the three-dimensional distance between the drone's starting point and the end point of the path segment; the second is the obstacle density, which represents the number or percentage of obstacles (such as exhaust vents, brackets, and obstructions) per unit area within the path segment; and the third is the snow depth, which is the average thickness of the snow layer over the area covered by the path segment.

[0088] To achieve optimized calculations, the system can call an improved A-star algorithm, whose core mechanism is to give priority to those paths among all candidate paths that are currently known to have lower costs and are expected to have lower total costs after reaching the target. Each time the system evaluates the path by extending the nodes in four or eight directions from the current position, it calculates the sum of the energy consumed by the current path and the energy consumed by the expected remaining paths at each step, thereby selecting the optimal direction to move forward. Another optional method is the ant colony optimization algorithm. In the process of multiple "virtual ants" exploring paths at the same time, each time a path with a better cost is discovered, the path will be assigned a higher "pheromone value". The more pheromones a path has, the more likely it is to be selected in subsequent searches. As the number of times the algorithm runs increases, one or several approximately optimal paths will eventually be gathered.

[0089] When the system generates a path, it also adds execution parameters for each segment, including starting coordinates, target coordinates, flight altitude, recommended speed, operating mode, and estimated energy consumption. If a segment exceeds 15 meters in length or its continuous turning angle exceeds 45 degrees, the system automatically inserts relay nodes in the middle to smooth the path and prevent the drone from becoming unstable or missing the target area.

[0090] The system also incorporates a pre-execution verification mechanism. If the total estimated energy consumption of the generated route exceeds 80% of the battery capacity, the system will prompt a return or deployment of a backup drone to continue the mission. This threshold is determined by taking into account the platform battery capacity (e.g., 6000 mAh), average per-minute power consumption (approximately 30 Wh), and the unforeseen windage losses and power consumption ratios associated with attitude corrections during actual flight. The system reserves 20% of its power for emergency return-to-home, providing redundancy for drone operations.

[0091] In summary, the path optimization module integrates the full amount of information on priority distribution maps, roof terrain data, and drone operating capabilities to build a multi-objective cost analysis mechanism. With the help of heuristic search or evolutionary optimization methods, it generates a preliminary clearing operation path with high efficiency, high safety, and high responsiveness, ensuring the system's deployability and dynamic adaptability to meet the needs of snow removal operations in complex and changing environments.

[0092] This path optimization mechanism overcomes the limitations of traditional static path design, which is unable to adapt to complex terrain and adjust clearance priorities in real time. By introducing a priority-driven, multi-objective path cost function, and environmental perception strategies, the system achieves highly intelligent path sorting and optimization, improving the efficiency, stability, and safety of drone operations. This mechanism significantly reduces redundant paths and flight risks, optimizing overall energy consumption and clearance completion time, particularly in environments with complex photovoltaic array layouts, irregular roof structures, and significant wind pressure distribution.

[0093] According to the preliminary clearing operation path plan, the specific formula for dynamic adjustment based on terrain obstacles and snow distribution density in the path is:

[0094] ;

[0095] in, represents the total path cost, Indicates the The flight length of the segment path, Indicates the The obstacle density on the segment path, Indicates the The thickness of snow on the path, Represents the path segment index, Represents the path segment index, represents the weighting coefficient of the path length, represents the weighting coefficient of obstacle density, Indicates the weighting factor of snow thickness.

[0096] The core of this path adjustment mechanism is the introduction of a path cost function, which dynamically determines the quality of each path segment and guides the path planning algorithm toward the lowest-cost path. The system breaks down the initial path into several segments, each of which corresponds to a set of parameters: the first is flight length, which is the three-dimensional distance between the drone's starting point and the end point of the path segment; the second is obstacle density, which represents the number or percentage of obstacles (such as exhaust vents, brackets, and obstructions) per unit area within the path segment; and the third is snow depth, which is the average snow thickness over the area covered by the path segment. When calculating each path segment, the system assigns weighted coefficients to these three parameters. The weighting factor for path length controls the system's trade-off between prioritizing flight time and energy consumption, and is typically set between 0.5 and 1.0. The weighting factor for obstacle density can be adjusted based on the complexity of the roof structure and is typically set to 1.0 or higher to increase the priority of the obstacle avoidance path in the algorithm. The weighting factor for snow thickness is adjusted based on the intensity of the operation. If the snow is thicker than 5 cm and is hardened, the system may set the weighting factor to 1.5 or even 2.0, indicating that the energy consumption or operation time for this segment is significantly higher than in normal areas. The algorithm operates as follows: During system initialization, the initially generated path segments are sequentially numbered. For each segment, flight distance, obstacle density, and snow thickness are evaluated. These three metrics are multiplied by their weighting factors and summed to obtain the overall cost of the segment. The system then sums the costs of all path segments to form a total path cost. A path optimization algorithm (such as A* or the ant colony algorithm) compares the total costs of all candidate paths and prioritizes the path with the lowest total cost as the final execution path.

[0097] The introduction of a segment-level cost function and the weighted calculation of three key environmental factors significantly improve the adaptability of path planning and the stability of actual execution. By dynamically identifying high-obstacle, high-snow, and high-risk areas, the system effectively avoids redundant paths, reduces operational energy consumption, and improves the efficiency and safety of drone clearing in complex scenarios. This mechanism also supports multi-scenario adaptive adjustment, a key expansion capability for intelligent autonomous flight systems.

[0098] The total path cost represents the combined cost of executing the entire path. It's the sum of the costs of all path segments and is used to measure the overall performance of different paths in terms of safety, energy consumption, and clearing efficiency. Each path is divided into several segments, each of which corresponds to three key parameters: flight length, obstacle density, and snow depth. Flight length represents the three-dimensional distance from the starting point to the end point of the i-th path segment, measured in meters and typically ranging from 2 to 10 meters. It is a direct indicator of battery consumption and flight time. Obstacle density represents the number or structural complexity of obstacles per unit area of ​​the i-th path segment, typically ranging from 0 to 5. Obstacles include brackets, edge devices, vents, and other obstacles. This parameter can be obtained through lidar or image recognition models and is used to measure the complexity and safety risk of the flight path. Snow depth represents the average snow depth in the i-th path segment, measured in centimeters, typically ranging from 0 to 15 centimeters. It is calculated through sensor fusion and is an important indicator for evaluating the duration of the clearing task and the choice of snow removal mode (such as heating or sweeping). The system assigns weighting coefficients to each of these three parameters. The weighting coefficient for path length is typically set between 0.5 and 1.0, representing a balance between flight efficiency and energy conservation. The weighting coefficient for obstacle density is typically set between 1.0 and 1.5, with higher values ​​set for complex rooftop structures and dense obstacles to prioritize obstacle avoidance. The weighting coefficient for snow depth is typically set between 1.0 and 2.0, with higher coefficients applied for thicker, denser, and more challenging snow removal areas, increasing the algorithm's tendency to avoid or prioritize heavily snowed areas. Furthermore, the total number of path segment indices depends on the length of the entire path and the granularity of the segmentation, typically with segments every 5 to 10 meters. This helps control path accuracy and algorithm computational complexity. The system traverses all path segments, calculates the weighted cost of each segment, and sums them to obtain the total cost for the path. This is used in the path optimization algorithm to compare the pros and cons of all candidate paths, ultimately selecting the path with the lowest total cost as the optimal path for the clearing operation. This method integrates path structure, environmental complexity and task load factors into a unified evaluation system, which has good engineering feasibility and adaptability.

[0099] During testing, a drone equipped with a lidar and RGB camera was used to conduct actual scans of rooftop areas. The results showed that the flight length of a typical path segment ranged from 3 to 8 meters, with an average power consumption of approximately 0.12 watt-hours per meter during flight. Using a standard battery (6000 mAh, 11.1 V), a single drone could execute a complete operational path with an average length of over 90 meters, meeting the requirements of a single flight mission. Obstacle density was calculated by marking structural areas such as air vents, skylights, and support shadows. The density was approximately 0.8 obstacles per square meter in typical areas, but could reach as high as 2.5 obstacles per square meter on densely populated or older factory rooftops. Comparison revealed that when the obstacle density exceeded 2.0, the obstacle avoidance failure rate in path planning rose to over 15%. Therefore, setting a weighting factor of 1.5 in the simulation system effectively reduced the risk of path collisions. In terms of snow thickness, field observations of snow accumulation patterns during different time periods (such as early morning and afternoon) indicate that when snow thickness is less than 3 cm, the success rate of a single jet clearing operation exceeds 92%. However, in areas with a thickness exceeding 7 cm, the residual snow area after a single operation exceeds 40%. Therefore, setting a thickness weighting coefficient of 2.0 to increase the path penalty weight in thick snow areas is an effective strategy for balancing energy consumption control and clearance efficiency. In simulations, the total path cost model was applied to the same snow distribution map input and compared with three path strategies: unweighted, partially weighted, and three-factor weighted. The three-factor weighted path performed best in avoiding redundant flights, improving obstacle avoidance in high-risk areas, and reducing the number of repeated residual snow clearing operations. It shortened the average operation time by 17.6% and reduced energy consumption per operation by approximately 11.3%.

[0100] Path length weighting factor The specific formula is;

[0101] = ;

[0102] in, represents the weighting coefficient of the path length, Indicates the current remaining battery power of the drone. Indicates the upper limit of the drone battery capacity. Indicates the average wind speed in the current area. Indicates the maximum safe wind speed of the drone. Represents the base of natural logarithms.

[0103] When calculating the total path cost, the system needs to dynamically adjust the weight of the impact of the path length on the overall path cost according to the current operating status of the drone, that is, the path length weighting coefficient The core idea is: when the remaining battery power of the drone approaches the lower limit or the current wind speed approaches or exceeds the maximum safe wind speed threshold set by the flight platform, the importance of path length should be increased to encourage the path planning algorithm to give priority to shorter and safer path segments, avoiding the risk of energy overdraft or flight control instability caused by long-distance flight. To this end, the system will The calculation is constructed as a function with exponential logarithmic characteristics, and the input variables include: the current remaining power of the drone , Maximum battery capacity of the system , Current regional wind speed And the maximum wind speed allowed for drone flight .in and It can be read in real time by the flight control system power management module. The values ​​are obtained by meteorological sensors (such as anemometers) before or during flight. is the system initialization setting value, which is usually determined according to the UAV platform test parameters, such as 8 meters per second. Get the power ratio and Divide by The wind speed relative risk index is obtained, and the two are subtracted as the input of the exponential function, thereby outputting a weighted coefficient value that varies between 0 and 1. When the remaining power is low and the wind speed is high, Approaches 1; and when the power is sufficient and the wind speed is stable, Converging to smaller values, such as between 0.3 and 0.5, indicates that the path length has little impact on the total cost.

[0104] The above formula makes the impact of path length on path planning results environmentally aware and adaptive. Its significant advantages are: on the one hand, under low battery or adverse weather conditions, the system can proactively control flight risks, shortening path segments and avoiding long flight segments, effectively reducing the risk of flight interruption or deviation; on the other hand, under normal conditions, it does not excessively restrict path optimization flexibility, ensuring global path optimality. Experiments have shown that the implementation of this weighting mechanism reduces the in-flight failure rate by approximately 22%, increases the path return rate by 17%, and shortens the average path length by approximately 13%, comprehensively improving the system's operational safety and efficiency.

[0105] Path length weighting factor It is used to reflect the influence of path length in the total path cost function. Its value range is 0 to 1. It usually approaches 1 when the flight risk increases or the energy margin decreases, and tends to be between 0.3 and 0.5 when the operating environment is good and the system power is sufficient. Indicates the current real-time remaining power of the drone, which is monitored by the flight control power module. The unit is usually milliampere-hour or percentage. If it is lower than the system preset safety power threshold (such as 20%), the system will increase the power. weighting to avoid long flight segments; The maximum battery capacity of the drone, generally between 5000 and 8000 mAh, is statically set according to the drone model and is used to normalize the battery risk ratio. Indicates the average wind speed in the current operating area, which can be obtained through an ultrasonic anemometer or environmental sensor. The unit is meters per second, and the common value is 2 to 6 meters per second. When the wind speed exceeds the system's preset safety limit, it will significantly affect flight stability. The maximum wind speed threshold that a drone can withstand is usually between 7 and 10 meters per second. It is determined by the power system and aerodynamic performance of the flight platform and cannot be changed with the mission. In order to unify the modeling of power risk and wind speed risk as logical factors for dynamically adjusting the path weight, the system subtracts the power ratio from the wind speed ratio as the input of the exponential function, and constructs a smoothly changing response curve through the natural logarithm base e (approximately 2.718), thereby outputting a weighting coefficient between 0 and 1. , achieving intelligent dynamic adjustment of path length weights. This modeling approach enables drones to proactively compress flight paths in energy-starved or windy environments, reducing flight risks and the probability of sudden energy consumption increases during operations. It also ensures maximum flexibility in path optimization under safe conditions, improving the overall system's adaptability and mission execution success rate.

[0106] During the experiment, the system collected the remaining power of the drone. The path adjustment behavior is gradually reduced from 100% to 15%, and the path length weighting coefficient is observed under the same target point. The results show that when When it is higher than 80%, Basically maintained between 0.35 and 0.45, the system tends to optimize the overall efficiency of the path; when Reduced to below 40%, It starts to rise rapidly and reaches above 0.85 when the battery is at 25%. At this time, the path algorithm significantly compresses the length of the segment and actively avoids distant targets, showing a clear energy-saving trend. In addition, wind speed data was collected on site. From 1 m / s to 7 m / s, The impact of also shows a stable increasing relationship. When the speed is lower than 3 m / s, the path planning is not affected; when the speed is lower than 3 m / s, the path planning is not affected. Approaching the maximum safe wind speed threshold for drones (9 m / s) 70%, that is, above 6.3 m / s, When the error rate increases to 0.9, the path change rate increases by 26%. The system clearly shows a tendency to move towards safe areas and avoid long-distance clearing points. The regulation mechanism of the system is relatively stable under conditions of sudden wind speed changes or power fluctuations. The model's system mission completion rate increased by 14%, average energy consumption per flight decreased by approximately 11.2%, and mission interruption rate dropped by approximately 22%. These data clearly demonstrate that this exponential modeling approach using path length weighting coefficients accurately reflects the current state of the aircraft and changes in environmental loads, exhibits excellent dynamic response capabilities, engineering adaptability, and deployment effectiveness, and represents a mature control algorithm framework that can be implemented by technical personnel in the relevant field within their usual capabilities.

[0107] Weighting coefficient of obstacle density The specific formula is: = ;

[0108] in, represents the weighting coefficient of obstacle density, Represents the total projected area of ​​obstacles on the current path segment, represents the total traversable area of ​​the path segment, Indicates the shortest distance between the centerline of the path segment and the nearest obstacle, Indicates the safe flight control radius of the drone; =1- - ;in, represents the weighting coefficient of the path length, represents the weighting coefficient of obstacle density, Indicates the weighting factor of snow thickness.

[0109] The system determines the obstacle density weighting coefficient through two core obstacle factors The values ​​of are: first, the obstacle occlusion ratio, which is the ratio between the area blocked by obstacles in the projection area of ​​the path segment and the total pass area of ​​the path segment, reflecting the degree to which the space resources of this segment are compressed by the obstacle structure; second, the obstacle proximity, which is the shortest distance between the center line of the path segment and the nearest obstacle, reflecting whether the operation redundancy space during the UAV flight is sufficient. In this formula, It represents the total projected area of ​​obstacles on the path segment and can be obtained through projection reconstruction or 3D modeling. The total traversable area of ​​the area covered by the path segment, in square meters, is calculated by the system using the navigation layer. The shortest distance from the centerline of the path segment to the surface of the nearest obstacle, in meters, can be measured in real time by a high-precision lidar; The safe flight control radius set in the flight control system, for example, 1.2 meters, is to ensure that the platform has basic maneuvering space. The above two factors respectively construct a logarithmic growth relationship and a square decreasing inhibition term, forming a sensitive response to complex obstacle structures. If a certain path segment has a large occlusion ratio (such as / > 0.5), or the centerline is too close to an obstacle (e.g. Less than ), calculated by the system The value will quickly rise to above 0.8, which means that the obstacle cost of this path segment is significantly higher than that of the ordinary segment. On the contrary, if the obstacles in the area are sparse, If it can be kept below 0.3, the system will regard the section as a priority area. =1- - , so that the total weight of the three cost factors is constant to 1, ensuring the structural balance and logical closure of the total cost function of the path.

[0110] By introducing the dual functions of logarithmic growth and distance decay, The weighted mechanism enables more nuanced identification of the risk level of path obstacle spaces, making it particularly suitable for densely structured rooftop environments, such as those containing skylights, distributed supports, and ventilators. The dynamic nature of this mechanism automatically increases penalty weights when the path approaches obstacles or when the flight space is critically compressed, preventing path planning from excessively approaching high-risk areas, significantly improving the system's flight safety and path execution stability.

[0111] It is the obstacle density weighting coefficient, which indicates the relative influence of the flight risk caused by obstacles in the path segment in the total path cost. The value range is usually set between 0.2 and 0.9. The larger the value, the more serious the impact of obstacles on the path segment, and the more it should be avoided by the system. The total projected area of ​​all obstacles within the path segment, expressed in square meters. Common values ​​range from 0 to 3 square meters. Obstacles include fixed components such as skylights, support columns, fans, or exhaust pipes. The passable area of ​​the path segment is usually obtained by multiplying the path segment length by the flight safety width. The value is generally between 3 and 10 square meters. and The ratio reflects the proportion of obstacles blocking the space, which is the key basis for measuring the degree of space access restriction. It indicates the shortest distance from the centerline of the path segment to the nearest obstacle in meters. It is obtained by the airborne laser radar or visual ranging module. The typical value is 0.5 to 3 meters. The smaller the value, the more severe the flight space compression. The minimum lateral control space required by the UAV platform in a safe flight state, that is, the safe flight control radius, is usually set to 1.0 to 1.5 meters, which is used to construct the space compression ratio and reflect the flight margin. Lower than The system automatically determines that there is a risk of spatial congestion on this path segment and needs to increase Weight. The system establishes a dynamic weighting mechanism in complex environments through the joint control of the above two dimensions, and avoids indicators from getting out of control through logarithmic growth and square decay. At the same time, to maintain the stability of the total cost function structure, the system sets the snow thickness weighting coefficient w3 to 1 minus the path length weighting coefficient. and obstacle density weighting coefficient The sum of =1- - , so that the three weights are always normalized, which not only ensures the rationality of the cost function calculation, but also achieves a dynamic balance between work intensity, flight safety and structural complexity in path planning, effectively improving the system's path adaptability and clearing efficiency in real scenarios.

[0112] The average obstacle projection area per path segment identified by 3D modeling using a drone-mounted lidar The area of ​​the passable area is 2.3 square meters It is about 5.2 square meters, and the projection shading ratio reaches 0.44. The system automatically calculates the The values ​​are concentrated between 0.68 and 0.82, which are significantly higher than the 0.25 to 0.35 in the open area. In the sample with the smallest obstacle spacing, the path centerline has the smallest distance from the obstacle. The safe flight control radius set by the system is 0.7 meters. is 1.2 meters, at this time / is 0.58, the system calculates The value reached 0.89, and the path optimization module automatically triggered the "low priority" mark on this path and tried to detour, showing good responsiveness. In the comparative simulation of 63 path segments, the system introducing this formula successfully avoided the structural interference path 22 times, the path planning efficiency was improved by about 19%, and the interruption rate was reduced by nearly 15%. In addition, because the system is based on and Adaptive Computing (snow thickness weighting coefficient) ensures that the system automatically reduces the priority given to snow removal in high-obstacle risk areas, thereby enhancing the path safety of obstacle avoidance planning. Extensive sample data demonstrates that all parameters in the formula can be obtained in real time through existing software and hardware architectures. This formula can stably drive the path evaluation system to produce the expected response in actual missions, demonstrating excellent embeddability, sensitivity, and mission adaptability, thus ensuring full real-world feasibility.

[0113] The multi-source sensor network module includes a lidar unit, an infrared thermal imaging sensor unit and a high-definition camera unit. The lidar unit is used to scan the three-dimensional contours of the photovoltaic panel and the thickness of snow. The infrared thermal imaging sensor is used to identify areas with abnormal temperature differences. The high-definition camera is used to obtain visible light images for image recognition and auxiliary verification.

[0114] In this embodiment, the multi-source sensor network module consists of a lidar unit, an infrared thermal imaging sensor unit, and a high-definition camera unit. These three units are connected to the drone's main control system via a unified communication bus (such as CAN or UART), with synchronous control and data processing performed by the perception and scheduling module. The lidar unit utilizes a rotating multi-line lidar device, such as a 16- or 32-line laser array. It operates by periodically emitting laser pulses and receiving reflected echoes. The distance to the target surface is calculated by measuring the time-of-flight difference. Combined with attitude and position data provided by the drone's inertial navigation system (IMU) and GPS module, a SLAM algorithm is used to construct a three-dimensional point cloud model of the photovoltaic surface. Based on this, the system compares the point cloud with a reference panel height benchmark to identify localized protruding snow areas and further estimate the snow thickness. To ensure data validity, the system sets a distance tolerance threshold, typically 3 to 7 centimeters. Points exceeding this height difference are considered snow-covered.

[0115] The infrared thermal imaging sensor unit measures the intensity of infrared radiation from the photovoltaic panel surface to generate a thermal distribution image. Due to differences in thermal conductivity, snow-covered areas are generally cooler than bare areas. The system uses the image's grayscale distribution to calculate the temperature difference range, setting a threshold of 1.5°C to 3°C. When the temperature difference between adjacent areas exceeds this range, the system marks them as potential snow accumulation areas and generates a heat map layer based on the results. In the early morning or when sunlight is insufficient, infrared images can provide supplementary information for primary judgment.

[0116] The high-definition camera unit utilizes a 5-megapixel or higher CMOS module with automatic exposure and low-light compensation capabilities to capture RGB images of the covered area. The system utilizes a trained image segmentation model (such as an improved U-Net or DeepLabV3+) to perform feature extraction and semantic segmentation on the images, identifying snow boundaries, thickness and color gradients, and occlusion structures. This image data can be used to verify the consistency of radar and thermal imaging recognition results and assist in analysis in obscured environments.

[0117] The data from these three sensor types undergoes time alignment, spatial calibration, and coordinate system transformation in the perception fusion module. Ultimately, a weighted strategy is used to generate a multi-source fused snow feature dataset. The fusion results are then fed into the path planning module to generate an optimal operational path, ensuring the drone can accurately avoid obstacles and clear targets based on a high-confidence environmental model. This entire process can be performed once for global perception initialization before flight or iterated in real time during flight, offering a high degree of automation.

[0118] The path planning module includes an obstacle recognition processing unit and a dynamic cost weight evaluation unit. The obstacle recognition processing unit is used to analyze the position, size and spatial distribution of obstacles above and around the photovoltaic array. The dynamic cost weight evaluation unit dynamically calculates the comprehensive cost of the path segment based on the current remaining power of the drone, the wind speed level in the flight area and the density of snow removal tasks covered by the unit path.

[0119] In this embodiment, the path planning module integrates an obstacle recognition processing unit and a dynamic cost weight assessment unit. These two units work together to achieve intelligent path generation and optimization for drones operating in complex photovoltaic rooftop environments. First, the obstacle recognition processing unit receives 3D laser point cloud data, infrared thermal images, and visible light image data from a multi-source sensor network module and simultaneously fuses them into a unified spatial perception map. The system then downsamples the point cloud data using a voxel grid method to construct a sparse 3D model. Cluster analysis is then used to extract dense regions as potential obstacle candidates. A box-fitting algorithm is then applied to each candidate point cluster to calculate its spatial position (X, Y, Z), dimensions (length, width, height), and orientation angle, among other geometric information. Simultaneously, an image recognition model is used to segment the boundaries of structural objects in the image and identify occlusion contours, thereby verifying the accuracy of obstacle locations detected by radar. After cross-validation of these three types of data, the system integrates the obstacle information into a structured grid map and maps it into the flight area cost layer.

[0120] Based on this, the dynamic cost weighting assessment unit initiates the path cost estimation process. First, the drone's current remaining battery life is regularly reported by the battery management system (BMS) as a percentage. The system then adjusts the weighting of the path length factor based on the remaining battery life. For example, when the battery life is below 30%, the system significantly deprioritizes long-range path segments and directs them to focus on nearby areas. When the battery life is sufficient (above 70%), the system allows for the planning of long-range clearance segments with wide coverage. Second, wind speed levels in the flight area are measured in real time by an onboard micro-anemometer or external data obtained through an online meteorological interface, in meters per second. The system sets a wind speed threshold range of 3 to 7 meters per second. When the detected wind speed exceeds 5 meters per second, the risk penalty factor is increased for crosswind-exposed areas of the path segment. Third, the unit path task load is calculated by multiplying the path segment's coverage area by the corresponding snow depth. The system uses the average snow volume per square meter as an indicator to assess task intensity, assigning clearance priority to path segments with high task density.

[0121] The three input parameters are normalized and mapped to a path length factor, a wind speed risk factor, and a mission intensity factor, respectively. A weighted model then calculates a composite cost for each candidate path segment and uses it as an input to the critical path score in heuristic searches (such as A* or Dijkstra). The system dynamically adjusts the three weighting coefficients to adapt to changes in mission phases and environmental disturbances, achieving a balance between safety, efficiency, and endurance.

[0122] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A UAV system for clearing snow from rooftop solar photovoltaic arrays, characterized in that: The system comprises: The multi-source sensor network module deploys a multi-source sensor network to conduct a comprehensive scan of the photovoltaic array coverage area to obtain raw data information on snow cover, including the thickness and preliminary distribution characteristics of snow in different areas, and obtain a preliminary snow cover dataset; The path planning module determines a preliminary clearing operation path plan based on the preliminary snow cover dataset. Based on the preliminary clearing operation path plan, it dynamically adjusts the path based on terrain obstacles and snow distribution density. If snow areas with density higher than a preset threshold are detected in the path, the flight altitude and operation sequence of the UAV are preferentially adjusted to obtain the optimized operation path. The operation execution and feedback collection module uses the optimized operation path to obtain real-time environmental feedback data from the drone during the clearing operation. It analyzes the residual snow in the feedback data and determines whether a second clearing operation is required in the area. The secondary clearing path planning module analyzes the results of real-time environmental feedback data. If it determines that the residual snow in a local area exceeds a preset threshold, it will re-plan the local path in the area and determine the order of secondary clearing operations; The control system module transmits the adjusted path information to the UAV control system through the secondary clearing operation sequence, performs precise operations on the residual snow in the local area, and obtains the final clearing operation completion data.

2. The UAV system for clearing snow from a rooftop solar photovoltaic array according to claim 1, characterized in that: Determining a preliminary clearing operation path plan based on the preliminary snow cover dataset includes: Based on the preliminary snow cover dataset, image processing technology is used to denoise and extract features from the data, and classification processing is performed on the snow distribution characteristics of different regions to determine a processed snow distribution feature dataset.

3. The UAV system for clearing snow from a rooftop solar photovoltaic array according to claim 2, characterized in that: Determining a preliminary clearing operation path plan based on the preliminary snow cover dataset also includes: By inputting the processed snow distribution feature dataset into a pre-established 3D modeling module, spatial mapping is performed on the span and density information of the snow distribution to generate the corresponding 3D visualization graphics of the snow distribution.

4. The UAV system for clearing snow from a rooftop solar photovoltaic array according to claim 3, characterized in that: Determining a preliminary clearing operation path plan based on the preliminary snow cover dataset also includes: Based on the generated three-dimensional visualization of snow distribution, the features of areas with snow density higher than the preset threshold and uneven distribution are extracted, and priority marking is performed on the characteristic areas to obtain a snow distribution graphic with priority labels.

5. The UAV system for clearing snow from a rooftop solar photovoltaic array according to claim 4, characterized in that: Determining a preliminary clearing operation path plan based on the preliminary snow cover dataset also includes: By using snow distribution graphics with priority labels and combining them with environmental data of complex terrain, a path optimization algorithm is used to plan the drone's operation path and determine the preliminary clearing operation path plan.

6. The UAV system for clearing snow from a rooftop solar photovoltaic array according to claim 1, characterized in that: The specific formula for dynamically adjusting the terrain obstacles and snow distribution density in the path based on the preliminary clearing operation path plan is as follows: ； in, represents the total path cost, Indicates the The flight length of the segment path, Indicates the The obstacle density on the segment path, Indicates the The thickness of snow on the path, Represents the path segment index, Represents the path segment index, represents the weighting coefficient of the path length, represents the weighting coefficient of obstacle density, Indicates the weighting factor of snow thickness.

7. The UAV system for clearing snow from a rooftop solar photovoltaic array according to claim 6, characterized in that: The weighting factor of the path length The specific formula is; ＝ ； in, represents the weighting coefficient of the path length, Indicates the current remaining battery power of the drone. Indicates the upper limit of the drone battery capacity. Indicates the average wind speed in the current area. Indicates the maximum safe wind speed of the drone. Represents the base of natural logarithms.

8. The UAV system for clearing snow from a rooftop solar photovoltaic array according to claim 6, characterized in that: The weighting coefficient of the obstacle density The specific formula is: ＝ ； in, represents the weighting coefficient of obstacle density, Represents the total projected area of ​​obstacles on the current path segment, represents the total traversable area of ​​the path segment, Indicates the shortest distance between the centerline of the path segment and the nearest obstacle, Indicates the safe flight control radius of the drone; described =1- - ; in, represents the weighting coefficient of the path length, represents the weighting coefficient of obstacle density, Indicates the weighting factor of snow thickness.

9. The UAV system for clearing snow from a rooftop solar photovoltaic array according to claim 1, characterized in that: The multi-source sensor network module includes a lidar unit, an infrared thermal imaging sensor unit and a high-definition camera unit. The lidar unit is used to scan the three-dimensional contours of the photovoltaic panel and the thickness of snow. The infrared thermal imaging sensor is used to identify areas with abnormal temperature differences. The high-definition camera is used to obtain visible light images for image recognition and auxiliary verification.

10. The UAV system for clearing snow from a rooftop solar photovoltaic array according to claim 1, characterized in that: The path planning module includes an obstacle recognition processing unit and a dynamic cost weight evaluation unit. The obstacle recognition processing unit is used to analyze the position, size and spatial distribution of obstacles above and around the photovoltaic array. The dynamic cost weight evaluation unit dynamically calculates the comprehensive cost of the path segment based on the current remaining power of the drone, the wind speed level in the flight area and the density of snow removal tasks covered by the unit path.

Images