A photovoltaic module cleaning operation task allocation method based on multi-unmanned aerial vehicle cooperation

Abstract

The present application relates to a photovoltaic module cleaning operation task allocation method based on multi-unmanned aerial vehicle cooperation, and belongs to the technical field of intelligent operation and maintenance of photovoltaic stations. The present application solves the problem of large energy consumption and long time of unmanned aerial vehicle cleaning photovoltaic stations in the prior art. Each photovoltaic module is clustered, and the weighted centroid and nest candidate point of each type of photovoltaic module are obtained. A nest site selection cost function is calculated according to the nest candidate point, the three-dimensional coordinates of the photovoltaic module, and the weighted centroid, and the final nest site is determined. An unmanned aerial vehicle energy consumption cost function is obtained according to the three-dimensional coordinates of each photovoltaic module and the final nest site. An unmanned aerial vehicle cleaning time cost function is obtained according to the dirtiness category of each photovoltaic module and the series connection relationship of the photovoltaic module. A reward function is constructed based on the unmanned aerial vehicle energy consumption cost function and the unmanned aerial vehicle cleaning time cost function, the final nest site, and the three-dimensional coordinates of each photovoltaic module, and then the unmanned aerial vehicle cleaning task is allocated. The photovoltaic module cleaning operation task allocation method based on multi-unmanned aerial vehicle cooperation is realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology for photovoltaic power plants, and in particular to a method for allocating photovoltaic module cleaning tasks based on multi-UAV collaboration. Background Technology

[0002] With the global energy transition accelerating, the demand for clean energy continues to grow, and photovoltaic (PV) power generation, with its sustainability advantages, is experiencing continuous expansion in installed capacity. Large-scale PV power plants typically occupy vast areas and are distributed across complex terrains such as plains, hills, and mountains. Outdoor PV modules, operating for extended periods, easily accumulate large amounts of dust and various contaminants. These pollutants interfere with the light transmittance and heat radiation of the PV modules, directly affecting the solar energy conversion efficiency. Therefore, PV cleaning has become an important maintenance task to ensure the efficiency of PV power generation.

[0003] Traditional methods for cleaning photovoltaic (PV) modules include manual cleaning and robotic cleaning. PV power plants are typically large-scale, and manual cleaning suffers from low efficiency, high cost, and high labor intensity, making it unsuitable for the large-scale operation and maintenance needs of such plants. While cleaning robots can operate flexibly within single strings, cross-string operations still require manual handling and scheduling, failing to achieve automation and being unsuitable for PV power plants with diverse terrains. Both methods fall short of meeting the requirements for efficient and intelligent operation and maintenance of PV power plants.

[0004] In recent years, the development of drone technology has brought new opportunities for photovoltaic (PV) cleaning. Drones can fly precisely point-to-point, quickly locate target components, and achieve unmanned autonomous operation, significantly reducing operation time. However, for large PV power plants, single-drone operations suffer from insufficient coverage, low cleaning volume per flight, and high energy consumption over long distances. Drone swarms, through distributed collaboration, can operate in different areas and modules based on terrain, dirt distribution, and other information, significantly reducing the flight distance of a single drone and improving operational efficiency. This has become a key direction for solving the cleaning challenges of large PV power plants.

[0005] However, existing drone swarm cleaning and scheduling technologies still have many shortcomings and are difficult to adapt to the actual needs of complex photovoltaic power plants. First, mountainous photovoltaic power plants have significant terrain undulations, and traditional scheduling algorithms mostly plan paths based on planar maps, failing to consider altitude changes, leading to discrepancies between planned and actual energy consumption. Second, the types of dirt on photovoltaic module surfaces are complex and diverse, including dust, bird droppings, algae, and oil stains. Different types of dirt have significantly different physicochemical properties, but existing technologies typically use uniform cleaning parameters, making targeted cleaning difficult. Third, existing technologies lack consideration for the topological characteristics of photovoltaic strings, potentially leading to frequent string replacements during cleaning, increasing additional energy consumption. Fourth, existing multi-drone scheduling algorithms often take off from and return from fixed nests, but returning to the starting nest after completing the task is often not the optimal decision; therefore, planning the drones' origin and destination nests is equally important.

[0006] In summary, existing drone cleaning and scheduling technologies for photovoltaic power plants have significant shortcomings in terms of terrain adaptability, targeted dirt and grime treatment, string coordination, and energy consumption optimization. Summary of the Invention

[0007] Based on the above analysis, the present invention aims to provide a method for allocating photovoltaic module cleaning tasks based on multi-UAV collaboration, in order to solve the problems of long time and high energy consumption in existing photovoltaic power plant UAV cleaning scheduling technology.

[0008] This invention provides a method for allocating photovoltaic module cleaning tasks based on multi-UAV collaboration, including: S1. Obtain the three-dimensional coordinates and dirt category of each photovoltaic module; S2. Cluster each photovoltaic module and obtain the weighted centroid three-dimensional coordinates and candidate points of the cluster for each type of photovoltaic module; S3. Based on the location of candidate cell sites, the three-dimensional coordinates of photovoltaic modules, and the three-dimensional coordinates of the weighted centroid, calculate the cell site selection cost function to determine the final three-dimensional coordinates of the cell site. S4. Obtain the energy consumption cost function of the inspection drone based on the three-dimensional coordinates of each photovoltaic module and the three-dimensional coordinates of the final drone nest location; obtain the cleaning time cost function of the inspection drone based on the dirt type of each photovoltaic module and the series relationship of the photovoltaic modules. S5. Based on the energy consumption cost function and cleaning time cost function of the inspection drone, the three-dimensional coordinates of the final drone nest location, and the three-dimensional coordinates of each photovoltaic module, a reward function is constructed, and then the cleaning task of the inspection drone is allocated.

[0009] Furthermore, the photovoltaic modules are clustered to obtain the weighted centroid 3D coordinates and candidate node locations for each type of photovoltaic module, including: S21. Cluster the photovoltaic modules according to their three-dimensional coordinates to obtain the initial centroids of each type of photovoltaic module; S22. Select the initial position of the nest within the neighborhood of each initial centroid; S23. Calculate the weighted centroid three-dimensional coordinates of various photovoltaic modules based on the initial position of the cell and the distance between them, and select candidate cell positions within the neighborhood of the weighted centroid.

[0010] Furthermore, based on the three-dimensional coordinates of each photovoltaic module, clustering is performed on each photovoltaic module to obtain the initial centroids of each type of photovoltaic module, including the following steps: S211. Randomly select a photovoltaic module as the initial centroid; S212. Calculate the distance between each photovoltaic module and the initial centroid; S213. Obtain the probability that each photovoltaic module is selected as the first new centroid based on the distance between each photovoltaic module and the initial centroid; select the photovoltaic module with the highest probability as the first new centroid. S214. Calculate the distance between each photovoltaic module and each centroid, and classify each photovoltaic module into different categories according to the distance; wherein, each centroid includes an initial centroid and a newly added centroid, and each centroid corresponds to a category; take the minimum value of the distance between each photovoltaic module and each centroid as the nearest centroid distance of each photovoltaic module. S215. Obtain the probability of each photovoltaic module being selected as the new centroid based on the nearest centroid distance of each photovoltaic module; select the photovoltaic module with the highest probability as the new centroid; repeat steps S4-S5 until all photovoltaic modules are clustered together. Class, the initial centroid and -1 new centroids constitute the initial centroids of various photovoltaic modules.

[0011] Furthermore, the formula for calculating the distance between each photovoltaic module and each centroid is as follows:

[0012] in, The first clustered cluster The location of each photovoltaic module in a photovoltaic module-like structure. for The X-axis coordinate, for Y-coordinate, for Z-direction coordinates; For the first The center of mass of photovoltaic modules for The X-axis coordinate, for Y-coordinate, for Z-direction coordinates; This is the altitude weighting coefficient.

[0013] Furthermore, the step of obtaining the probability of each photovoltaic module being selected as the first new centroid based on the distance between each photovoltaic module and the initial centroid includes: calculating the square of each distance and the sum of the squares of each distance based on the distance between each photovoltaic module and the initial centroid; and obtaining the probability of each photovoltaic module being selected as the first new centroid based on the ratio of the square of each distance and the sum of the squares of each distance.

[0014] Furthermore, the step of obtaining the probability of each photovoltaic module being selected as a new centroid based on the nearest centroid distance of each photovoltaic module includes: calculating the square of each nearest centroid distance and the sum of the squares of the nearest centroid distances; and obtaining the probability of each photovoltaic module being selected as a new centroid based on the ratio of the square of each nearest centroid distance to the sum of the squares of the nearest centroid distances.

[0015] Furthermore, the formula for calculating the nesting location cost function is as follows: , in, This represents the candidate location for the nest. The first clustered cluster The location of each photovoltaic module in a photovoltaic module-like structure. For the first The weighted centroid of photovoltaic modules For the first Weighted centroid and candidate location of solar cell for photovoltaic modules Distance between The first clustered cluster The distance between each photovoltaic module and the candidate location of the solar cell in a photovoltaic module system.

[0016] Furthermore, the energy consumption cost function of the inspection drone is calculated using the following formula: , in, The inspection drone consumes energy to ascend. To inspect the energy consumption of the drone during horizontal flight, The energy consumption of inspection drones is increasing. The horizontal flight energy consumption weight for inspection drones is defined as M, where M is the number of horizontal waypoints for all inspection drones, and N is the number of waypoints divided according to altitude. A value of 1 indicates that the photovoltaic modules to be cleaned are in the current flight segment. A value of 0 indicates that the photovoltaic modules to be cleaned are not in the current flight segment.

[0017] Furthermore, based on the type of dirt on each photovoltaic module and the series connection of the photovoltaic modules, the cleaning time cost function of the inspection drone is obtained, including: , , in, The cleaning time for the photovoltaic strings on each flight segment is determined by the type of dirt. C is the time optimization coefficient for photovoltaic strings, and C represents the photovoltaic strings that cluster into the same class in a certain flight segment.

[0018] Furthermore, the formula for constructing the reward function is as follows: , ; in, For the reward function, This is a weighting factor for cleaning time. Z represents the energy consumption weighting coefficient for the inspection drone; Z represents the overall task allocation cost function. Assign weights to the overall task based on the inverse of the cost function. The sum of the distances each drone travels from its starting nest to each photovoltaic module to be cleaned, and then back to its destination nest; The sum of the distances between the starting and ending nests and The absolute value of the difference.

[0019] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects: 1. Cluster the photovoltaic modules and obtain the weighted centroid 3D coordinates and candidate nest locations for each type of photovoltaic module; calculate the nest location cost function based on the candidate nest location, the 3D coordinates of the photovoltaic modules, and the weighted centroid 3D coordinates, thereby determining the final nest location 3D coordinates; the selected final nest location minimizes the sum of the distances from each UAV to the clustered photovoltaic modules, saving both flight energy consumption and flight time.

[0020] 2. Based on the three-dimensional coordinates of each photovoltaic module and the final three-dimensional coordinates of the drone nest location, an energy consumption cost function for the inspection drone is obtained. Based on the type of contamination of each photovoltaic module and the series connection of the photovoltaic modules, a cleaning time cost function for the inspection drone is obtained. Based on the energy consumption cost function, the cleaning time cost function, the final three-dimensional coordinates of the drone nest location, and the three-dimensional coordinates of each photovoltaic module, a reward function is constructed to allocate cleaning tasks for the inspection drone. This innovatively proposes a multi-dimensional target-constrained task allocation model, fully considering actual operation and maintenance interference factors of photovoltaic power plants, such as terrain, contamination distribution, and drone energy consumption. This method can be widely applied and adapted to various types of photovoltaic power plants.

[0021] 3. The reward function constructed in this invention simultaneously considers the task allocation quality of the photovoltaic modules to be cleaned at the starting and ending drone nests, as well as the impact of energy consumption and time on task allocation, and designs a reward mechanism to achieve the high-efficiency, energy-saving, and automated operation and maintenance requirements of photovoltaic power plants. Based on the multi-drone operation mode, the topological association characteristics of photovoltaic strings are utilized to reduce the drone transfer time between photovoltaic modules, enabling modular and rapid processing of photovoltaic module dirt cleaning tasks; the energy consumption optimization effect is outstanding. By integrating terrain information, dirt distribution, and other constraint information, a full-process energy consumption model is constructed to collaboratively optimize the allocation of drone cleaning tasks, shorten the flight area span, and achieve a breakthrough in energy saving.

[0022] 4. This invention is based on an improved YOLOv8 model to accurately detect and identify contamination information of photovoltaic modules. A branched Backbone structure is proposed to process visible light images and thermal infrared images separately. Based on the original structure of the Head part, three fully connected layers are added to output the probability distribution of contamination types. At the same time, temperature outliers are used to help optimize the classification results, thereby improving the accuracy of contamination identification.

[0023] The inspection drone accurately locates the dirt and identifies the type of dirt, thereby estimating the required cleaning time and providing precise input for cleaning task allocation. It also avoids missed cleaning and repeated cleaning through a series-level continuous cleaning mode.

[0024] 5. The photovoltaic module clustering process obtains the probability of each photovoltaic module being selected as a new centroid based on the nearest centroid distance. This includes: calculating the square of each nearest centroid distance and the sum of the squares of the nearest centroid distances; and obtaining the probability of each photovoltaic module being selected as a new centroid based on the ratio of the square of each nearest centroid distance to the sum of the squares of the nearest centroid distances. The formula for the distance between each photovoltaic module and each centroid includes an altitude variation term plus a weight adjustment coefficient, taking into account the significant impact of terrain undulations on the energy consumption of drones in mountainous photovoltaic power stations.

[0025] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0026] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.

[0027] Figure 1 This is a schematic diagram of the photovoltaic module cleaning task allocation method of the present invention; Figure 2 This is a block diagram of the technical solution of an embodiment of the present invention. Detailed Implementation

[0028] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.

[0029] A specific embodiment of the present invention discloses a method for allocating photovoltaic module cleaning tasks based on multi-UAV collaboration, the flowchart of which is shown below. Figure 1 As shown. Specifically, it includes steps S1-S5.

[0030] The technical solution block diagram of the embodiment of the present invention is as follows: Figure 2 As shown.

[0031] S1. Obtain the three-dimensional coordinates and dirt category of each photovoltaic module.

[0032] During the data acquisition phase, visual-laser mapping technology is used to accurately collect 3D terrain and component distribution information of the photovoltaic power station. Based on the collected images of the photovoltaic components, target detection algorithms are used to collect and analyze the types of contamination on the photovoltaic components, providing high-precision basic data support for subsequent task allocation. Specifically, S1 includes S11-S12.

[0033] S11. Obtain three-dimensional information of photovoltaic power station.

[0034] The inspection drone is equipped with a visible light camera and a lidar sensor. The visible light camera acquires the appearance features and topography of the photovoltaic power station, while the lidar sensor acquires the three-dimensional terrain data of the photovoltaic power station and the height information of the photovoltaic modules themselves.

[0035] First, based on the geographical location and area information of the photovoltaic power station, an inspection route is set for the inspection robot. The inspection drone flies autonomously along the preset route, simultaneously collecting visible light images and point cloud data. After the flight, the visible light images and point cloud data collected by the drone are transmitted back to the ground control station. After receiving the visible light images and lidar information from the inspection drone, the ground control station uses the R3live algorithm to fuse the 3D true-color information and point cloud information to generate a 3D true-color map. The center location of the photovoltaic power station is selected as the origin of the coordinate system to construct a world coordinate system, and the position coordinates of each photovoltaic module are marked on the map.

[0036] S12. Obtain the dirt category of the photovoltaic module.

[0037] The inspection drone is equipped with a high-resolution visible light camera and a thermal infrared camera. The high-resolution visible light camera captures the appearance features and surface contamination of the photovoltaic panels, while the thermal infrared camera detects the temperature distribution of the photovoltaic panels and identifies temperature anomalies caused by contamination. Simultaneously, the drone also carries a target detection edge computing module deployed with a YOLOv8 network model, enabling real-time identification of the type of contamination on the photovoltaic panel surface.

[0038] The inspection drones conduct a full inspection of the photovoltaic power station according to a preset route, collecting visible light and thermal infrared images of the surface of all photovoltaic modules. The YOLOv8 network model obtains the type of dirt on the surface of the photovoltaic modules. After the inspection, the geographical location information of the dirty photovoltaic modules and the dirt type information are uploaded to the ground control station, which serves as a constraint for drone task allocation and is fed back to the drone swarm.

[0039] Depending on the location of the photovoltaic power station, the types of dirt introduced include: wind and sand dust, fallen vegetation, algae and moss, bird droppings, biological secretions, watermarks, and oil stains.

[0040] After acquiring visible light and infrared images, the thermal infrared image is first spatially aligned with the visible light image using the SIFT feature point matching algorithm to ensure that the same dirty area has no coordinate deviation in the two modalities. Then, both images are input into the YOLOv8 network. A branch structure is designed in the backbone: texture and color features of the visible light image are extracted using CSPDarknet, while temperature distribution features of the thermal infrared image are extracted using the lightweight convolutional network MobileNetV3. The features from the two branches are weighted and fused in the Neck layer using an attention mechanism to enhance the salient features of the dirty area. In the head part, three fully connected layers are added to the original structure to output the probability distribution of dirt types. Temperature outliers are also used to assist in optimizing the classification results, improving the accuracy of dirt identification.

[0041] S2. Cluster each photovoltaic module and obtain the weighted centroid three-dimensional coordinates and candidate points of the cluster for each type of photovoltaic module, specifically including steps S21-S22.

[0042] S21. Cluster the photovoltaic modules based on their three-dimensional coordinates to obtain the initial centroids of each type of photovoltaic module. This includes the following steps: S211. Randomly select a photovoltaic module as the initial centroid.

[0043] S212. Calculate the distance between each photovoltaic module and the initial centroid.

[0044] S213. Obtain the probability that each photovoltaic module is selected as the first new centroid based on the distance between each photovoltaic module and the initial centroid; select the photovoltaic module with the highest probability as the first new centroid.

[0045] The step of obtaining the probability of each photovoltaic module being selected as the first new centroid based on the distance between each photovoltaic module and the initial centroid includes: calculating the square of each distance and the sum of the squares of each distance based on the distance between each photovoltaic module and the initial centroid; and obtaining the probability of each photovoltaic module being selected as the first new centroid based on the ratio of the square of each distance and the sum of the squares of each distance.

[0046] S214. Calculate the distance between each photovoltaic module and each centroid, and classify each photovoltaic module into different categories according to the distance; wherein, each centroid includes an initial centroid and a newly added centroid, and each centroid corresponds to a category; take the minimum value of the distance between each photovoltaic module and each centroid as the nearest centroid distance of each photovoltaic module.

[0047] The formula for calculating the distance between each photovoltaic module and each centroid is: , in, The first clustered cluster The location of each photovoltaic module in a photovoltaic module-like structure. for The X-axis coordinate, for Y-coordinate, for Z-direction coordinates; For the first The center of mass of photovoltaic modules for The X-axis coordinate, for Y-coordinate, for Z-direction coordinates; This is the altitude weighting coefficient.

[0048] A value of 0 indicates that altitude is not considered, and clustering is performed only on a two-dimensional plane; A value of 1 indicates that altitude and coordinate information have equal influence. >>1 indicates that altitude has a significant impact; two photovoltaic modules with a large altitude difference will not be classified into the same category.

[0049] S215. Obtain the probability of each photovoltaic module being selected as the new centroid based on the nearest centroid distance of each photovoltaic module; select the photovoltaic module with the highest probability as the new centroid; repeat steps S214-S215 until all photovoltaic modules are clustered together. Class, the initial centroid and -1 new centroids constitute the initial centroids of various photovoltaic modules.

[0050] The method of obtaining the probability of each photovoltaic module being selected as a new centroid based on the nearest centroid distance of each photovoltaic module includes: calculating the square of each nearest centroid distance and the sum of the squares of the nearest centroid distances; and obtaining the probability of each photovoltaic module being selected as a new centroid based on the ratio of the square of each nearest centroid distance to the sum of the squares of the nearest centroid distances.

[0051] S22. Select the initial position of the nest within the neighborhood of each initial centroid.

[0052] Specifically, the initial centroid is the geometric centroid, not the centroid usable in actual engineering. Since the cell cannot be directly built in the photovoltaic module group, the feasibility range of cell site selection should be generated in its neighborhood based on the initial centroid location and engineering constraints. Within the feasibility range, a location should be randomly selected as the initial location of the cell.

[0053] Optionally, its neighborhood is within 100 meters of the initial centroid.

[0054] S23. Calculate the weighted centroid three-dimensional coordinates of various photovoltaic modules based on the initial position of the cell and the distance between them, and select candidate cell positions within the neighborhood of the weighted centroid.

[0055] The weighted centroid three-dimensional coordinates are calculated using the following formula: , in, The first clustered cluster The initial position of the solar cell corresponding to a photovoltaic module type and the distance weighting coefficient between the solar cells and other photovoltaic modules in that type. For the first cluster The location of each photovoltaic module in a photovoltaic module-like structure. for The X-axis coordinate, for Y-coordinate, for The Z-axis coordinate.

[0056] Specifically, the closer the photovoltaic module is to the initial position of the corresponding cell, the better. The larger the value, the greater the impact on adjusting the initial position of the nest.

[0057] All points within a feasible neighborhood with a weighted centroid radius of r are considered as candidate nest locations. For example, r is 100 meters.

[0058] S3. Based on the candidate locations of the solar cell, the three-dimensional coordinates of the photovoltaic modules, and the weighted centroid three-dimensional coordinates, calculate the solar cell location cost function to determine the final three-dimensional coordinates of the solar cell location.

[0059] To evaluate the impact of nest location schemes on operational efficiency, the optimization function value for each candidate nest location is calculated, and the candidate nest location with the lowest cost is selected as the first nest location in that cluster. The final location of the solar cell corresponding to photovoltaic modules.

[0060] The formula for calculating the nest location cost function is as follows: , in, This represents the candidate location for the nest. The first clustered cluster The location of each photovoltaic module in a photovoltaic module-like structure. For the first The weighted centroid of photovoltaic modules For the first Weighted centroid and candidate location of solar cell for photovoltaic modules Distance between The first clustered cluster The distance between each photovoltaic module and the candidate location of the solar cell in a photovoltaic module system.

[0061] Based on the nest location cost function, the candidate nest location that minimizes the cost function value is selected as the final nest location.

[0062] The final nest location represents the location where the drone takes off and returns to its nest. This method significantly improves the planning efficiency and robustness of photovoltaic module cleaning operations in complex environments through improved distance measurement and adaptive adjustment mechanisms.

[0063] S4. Obtain the energy consumption cost function of the inspection drone based on the three-dimensional coordinates of each photovoltaic module and the three-dimensional coordinates of the final drone nest location; obtain the cleaning time cost function of the inspection drone based on the dirt type of each photovoltaic module and the series relationship of the photovoltaic modules.

[0064] The formula for calculating the energy consumption cost function of the inspection drone is as follows: , in, The inspection drone consumes energy to ascend. To inspect the energy consumption of the drone during horizontal flight, The energy consumption of inspection drones is increasing. The horizontal flight energy consumption weight for inspection drones is defined as M, where M is the number of horizontal waypoints for all inspection drones, and N is the number of waypoints divided according to altitude. A value of 1 indicates that the photovoltaic modules to be cleaned are in the current flight segment. A value of 0 indicates that the photovoltaic modules to be cleaned are not in the current flight segment.

[0065] Specifically, the waypoints represent the location of the drone's nest or the photovoltaic module where the drone lands. Some photovoltaic power plants are built in mountainous areas, requiring drones to climb to a certain altitude during operation. Compared to horizontal flight, climbing altitude consumes more energy, and the drone can complete fewer cleaning tasks. Therefore, altitude gain is considered a cost indicator. For drones whose operational routes require climbing, we aim for a more concentrated operational range, avoiding sparse, wide-area flight, and completing more cleaning tasks with limited energy consumption.

[0066] The cleaning time cost function of the inspection drone is obtained based on the type of dirt on each photovoltaic module and the series relationship of the photovoltaic modules, including: , , in, The cleaning time for the photovoltaic strings on each flight segment is determined by the type of dirt. C is the time optimization coefficient for photovoltaic strings, and C represents the photovoltaic strings that cluster into the same class in a certain flight segment.

[0067] Specifically, the cleaning time required for different types of dirt on photovoltaic modules varies, and the time that drones need to hover and wait in the air also varies. When assigning tasks to drones, this invention calculates the cleaning time required for different types of dirt, avoiding the situation where a large amount of difficult-to-clean dirt is assigned to a certain drone, which would lead to the depletion of power during the cleaning operation.

[0068] In many photovoltaic (PV) power plants, PV modules are connected in series to form PV strings. Cleaning robots can glide and move along these strings, eliminating the need for drones to lift them up during cleaning. Therefore, when different PV panels on the same string need cleaning, a drone can place the cleaning robot on any PV panel, allowing the robot to complete the cleaning task for the entire string independently. Thus, when allocating tasks to drones, the level of dirt on the same PV panel is considered. Complete string cleaning tasks are packaged and assigned to individual drones, saving time for drones to move between different PV strings and allowing for dynamic adjustment of drone numbers based on dirt distribution.

[0069] If some dirt is on a set of photovoltaic strings, the cleaning task for that set of strings can be packaged and assigned to a drone during task allocation. This can save the time of frequent drone take-off and landing, and reduce cleaning time by more than 30%.

[0070] Under various constraints, drone swarms can allocate and execute multiple tasks. By statistically analyzing altitude gain and types of dirt, they can accurately estimate the cleaning time of photovoltaic modules and the energy consumption of drones.

[0071] S5. Based on the energy consumption cost function and cleaning time cost function of the inspection drone, the three-dimensional coordinates of the final drone nest location, and the three-dimensional coordinates of each photovoltaic module, a reward function is constructed, and then the cleaning task of the inspection drone is allocated.

[0072] Based on the population algorithm, the drone swarm is regarded as a population, and each drone is an individual in the population. The cleaning needs of photovoltaic modules are the tasks to be assigned. The task allocation scheme is obtained with the goal of maximizing the reward function value.

[0073] The formula for constructing the reward function is as follows: (1) in, ; in, For the reward function, This is a weighting factor for cleaning time. Z represents the energy consumption weighting coefficient for the inspection drone; Z represents the overall task allocation cost function. Assign weights to the overall task based on the inverse of the cost function. The sum of the distances each drone travels from its starting nest to each photovoltaic module to be cleaned, and then back to its destination nest; The sum of the distances between the starting and ending nests and The absolute value of the difference.

[0074] Specifically, the goal of this embodiment is to optimize the relationship between the task requirements of each UAV and the cleaning requirements of photovoltaic modules, constructing it as a multi-constraint, multi-objective optimization problem. The UAV swarm is considered as a population, and each UAV is an individual within that population. Each individual will be assigned a series of tasks, with the cleaning requirements of the photovoltaic modules being the tasks to be assigned. When constructing tasks, constraints based on photovoltaic strings are introduced, and photovoltaic modules located in the same string to be cleaned are packaged into a task package. During assignment, tasks within the task package are assigned in a packaged manner.

[0075] Furthermore, after completing its mission, the drone needs to return to its destination nest, but it may not necessarily return to its origin nest. Therefore, when assigning tasks, it is necessary to decide not only which mission packages the drone will execute, but also which nest it will ultimately fly back to. Thus, the executed tasks are affected by both the origin nest and the destination nest.

[0076] Let the set of drones be The set of all tasks in the photovoltaic power station is , , in, Represents an independent task entity. This indicates the altitude at which the photovoltaic module is located. Indicates the types of dirt on photovoltaic modules. This indicates the series connection of photovoltaic modules. This indicates that the drone consumes energy to climb.

[0077] The overall task allocation cost function is: , in These are the weighting coefficients for cleaning time and climbing energy consumption, respectively.

[0078] The same task cannot be assigned to two drones; therefore, a binary decision variable is designed in the cost function to ensure the uniqueness of task allocation.

[0079] Where j represents the drone, with a total of M drones, and i represents the photovoltaic cleaning task, with a total of N tasks.

[0080] Both the starting and ending flight nests will evaluate the tasks in the photovoltaic power station. If the cost of a task to both the starting and ending flight nests is small, the cost function will be converted into a reward function to increase the probability of the UAV selecting the task. The reward function is shown in (1).

[0081] The smaller the cost function Z, the lower the distance cost. / The smaller the value, the larger the reward function value, and the better the task matches the start and end nests; conversely, the larger the value, the less suitable the task is for the start and end nests.

[0082] Compared with existing technologies, the method provided in this embodiment clusters each photovoltaic module and obtains the weighted centroid 3D coordinates and candidate nest locations for each type of photovoltaic module. Based on the candidate nest locations, the 3D coordinates of the photovoltaic modules, and the weighted centroid 3D coordinates, a nest location cost function is calculated to determine the final nest location 3D coordinates. This final nest location minimizes the sum of distances from each drone to the clustered photovoltaic modules, saving both flight energy and flight time. An inspection drone energy consumption cost function is obtained based on the 3D coordinates of each photovoltaic module and the final nest location 3D coordinates. A cleaning time cost function is obtained based on the dirt type of each photovoltaic module and the series relationship of the photovoltaic modules. A reward function is constructed based on the inspection drone energy consumption cost function, the inspection drone cleaning time cost function, the final nest location 3D coordinates, and the 3D coordinates of each photovoltaic module, and then the cleaning task of the inspection drone is allocated. This innovatively proposes a multi-dimensional target-constrained task allocation mode, fully considering the actual operation and maintenance interference factors of photovoltaic power plants, such as terrain, dirt distribution, and drone energy consumption. This method can be extended and adapted to various types of photovoltaic power plants. The reward function constructed in this embodiment considers the quality of task allocation for cleaning photovoltaic modules at both the starting and ending drone nests, as well as the impact of energy consumption and time on task allocation. A reward mechanism is designed to meet the needs of efficient, energy-saving, and automated operation and maintenance of photovoltaic power plants. Based on a multi-drone operation mode, the topological association characteristics of photovoltaic strings reduce the time for drones to transfer between photovoltaic modules, enabling modular and rapid processing of photovoltaic module cleaning tasks; energy consumption optimization is significant. By integrating terrain information, dirt distribution, and other constraint information, a full-process energy consumption model is constructed to collaboratively optimize drone cleaning task allocation, shorten the flight area span, and achieve energy-saving breakthroughs. This embodiment is based on an improved YOLOv8 model for accurate detection and identification of photovoltaic module dirt information. A branched Backbone structure is proposed to process visible light and thermal infrared images separately; based on the original structure of the Head part, three fully connected layers are added to output the probability distribution of dirt types, and temperature anomalies are used to assist in optimizing the classification results, improving the accuracy of dirt identification. The inspection drone precisely locates the dirt positions and identifies the types of dirt, thereby estimating the required cleaning time and providing accurate input for cleaning task allocation. A continuous string-level cleaning mode avoids missed or repeated cleaning. In this embodiment, the photovoltaic module clustering process obtains the probability of each photovoltaic module being selected as a new centroid based on the nearest centroid distance. This includes: calculating the square of each nearest centroid distance and the sum of the squares of the nearest centroid distances; and obtaining the probability of each photovoltaic module being selected as a new centroid based on the ratio of the square of each nearest centroid distance to the sum of the squares of the nearest centroid distances. The formula for the distance between each photovoltaic module and each centroid includes an altitude variation term plus a weighting adjustment coefficient, taking into account the significant impact of terrain undulations on the drone's energy consumption in mountainous photovoltaic power stations.

[0083] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0084] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for allocating photovoltaic module cleaning tasks based on multi-UAV collaboration, characterized in that, include: S1. Obtain the three-dimensional coordinates and dirt categories of each photovoltaic module, including: The inspection drone is equipped with a high-resolution visible light camera and a thermal infrared camera. The high-resolution visible light camera obtains the appearance features and surface dirt of the photovoltaic panels on the photovoltaic modules, and the thermal infrared camera detects the temperature distribution of the photovoltaic panels to identify temperature anomalies caused by dirt. The inspection drone is also equipped with a target detection edge computing module with a YOLOv8 network model, which can identify the dirt categories on the surface of the photovoltaic panels in real time. In the YOLOv8 network, a branch structure is designed in the backbone structure. The visible light image extracts texture and color features through CSPDarknet, and the thermal infrared image extracts temperature distribution features through the lightweight convolutional network MobileNetV3. The features of the two branches are weighted and fused in the Neck layer through an attention mechanism to enhance the salient features of the dirt area. The Head part adds three fully connected layers to output the probability distribution of dirt types on the basis of the original structure. At the same time, the classification results are optimized by temperature anomalies to improve the accuracy of dirt identification. S2. Cluster the photovoltaic modules and obtain the weighted centroid 3D coordinates and candidate node positions for each type of photovoltaic module, including: S21. Cluster the photovoltaic modules according to their three-dimensional coordinates to obtain the initial centroids of each type of photovoltaic module; S22. Select the initial position of the nest within the neighborhood of each initial centroid; S23. Calculate the weighted centroid three-dimensional coordinates of various photovoltaic modules based on the initial position of the cell and the distance between them, and select candidate cell positions within the neighborhood of the weighted centroid. S3. Based on the location of candidate cell sites, the three-dimensional coordinates of photovoltaic modules, and the three-dimensional coordinates of the weighted centroid, calculate the cell site selection cost function to determine the final three-dimensional coordinates of the cell site. S4. Obtain the energy consumption cost function of the inspection drone based on the three-dimensional coordinates of each photovoltaic module and the final three-dimensional coordinates of the drone nest location; the calculation formula for the energy consumption cost function of the inspection drone is as follows: ， in, The inspection drone consumes energy to ascend. To inspect the energy consumption of the drone during horizontal flight, To increase the energy consumption of inspection drones, The horizontal flight energy consumption weight for inspection drones is defined as M, where M is the number of horizontal waypoints for all inspection drones, and N is the number of waypoints divided according to altitude. A value of 1 indicates that the photovoltaic modules to be cleaned are in the current flight segment. A value of 0 indicates that the photovoltaic modules to be cleaned are not in the current flight segment; The cleaning time cost function of the inspection drone is obtained based on the type of dirt on each photovoltaic module and the series relationship of the photovoltaic modules, including: ， ， in, The cleaning time for the photovoltaic strings on each flight segment is determined by the type of dirt. S5 is the photovoltaic string time optimization coefficient, where C is the photovoltaic strings clustered into the same class in a certain flight segment; S6 is to construct a reward function based on the energy consumption cost function of the inspection drone, the cleaning time cost function of the inspection drone, the three-dimensional coordinates of the final nest location, and the three-dimensional coordinates of each photovoltaic module, and then allocate the cleaning task of the inspection drone.

2. The photovoltaic module cleaning task allocation method based on multi-UAV collaboration according to claim 1, characterized in that, Clustering of photovoltaic modules based on their three-dimensional coordinates yields the initial centroids for each module, including the following steps: S211. Randomly select a photovoltaic module as the initial centroid; S212. Calculate the distance between each photovoltaic module and the initial centroid; S213. Obtain the probability that each photovoltaic module is selected as the first new centroid based on the distance between each photovoltaic module and the initial centroid; select the photovoltaic module with the highest probability as the first new centroid. S214. Calculate the distance between each photovoltaic module and each centroid, and classify each photovoltaic module into different categories based on the distance between each photovoltaic module and each centroid; wherein, each centroid includes the initial centroid and the newly added centroid, and each centroid corresponds to a category; take the minimum value among the distances between each photovoltaic module and each centroid as the nearest centroid distance of each photovoltaic module. S215. Obtain the probability of each photovoltaic module being selected as the new centroid based on the nearest centroid distance of each photovoltaic module; select the photovoltaic module with the highest probability as the new centroid; repeat steps S4-S5 until all photovoltaic modules are clustered together. Class, the initial centroid and -1 new centroids constitute the initial centroids of various photovoltaic modules.

3. The photovoltaic module cleaning task allocation method based on multi-UAV collaboration according to claim 2, characterized in that, The formula for calculating the distance between each photovoltaic module and each centroid is: , in, The first clustered cluster The location of each photovoltaic module in a photovoltaic module-like structure. for The X-axis coordinate, for Y-axis coordinate, for Z-axis coordinate; For the first The center of mass of photovoltaic modules for The X-axis coordinate, for Y-axis coordinate, for Z-axis coordinate; This is the altitude weighting coefficient.

4. The photovoltaic module cleaning task allocation method based on multi-UAV collaboration according to claim 2, characterized in that, The step of obtaining the probability of each photovoltaic module being selected as the first new centroid based on the distance between each photovoltaic module and the initial centroid includes: calculating the square of each distance and the sum of the squares of each distance based on the distance between each photovoltaic module and the initial centroid; and obtaining the probability of each photovoltaic module being selected as the first new centroid based on the ratio of the square of each distance and the sum of the squares of each distance.

5. The photovoltaic module cleaning task allocation method based on multi-UAV collaboration according to claim 2, characterized in that, The method of obtaining the probability of each photovoltaic module being selected as a new centroid based on the nearest centroid distance of each photovoltaic module includes: calculating the square of each nearest centroid distance and the sum of the squares of the nearest centroid distances; and obtaining the probability of each photovoltaic module being selected as a new centroid based on the ratio of the square of each nearest centroid distance to the sum of the squares of the nearest centroid distances.

6. The photovoltaic module cleaning task allocation method based on multi-UAV collaboration according to claim 1, characterized in that, The formula for calculating the nest location cost function is as follows: ， in, This represents the candidate location for the nest. The first clustered cluster The location of each photovoltaic module in a photovoltaic module-like structure. For the first The weighted centroid of photovoltaic modules For the first Weighted centroid and candidate location of solar cell for photovoltaic modules Distance between The first clustered cluster The distance between each photovoltaic module and the candidate location of the solar cell in a photovoltaic module system.

7. The photovoltaic module cleaning task allocation method based on multi-UAV collaboration according to claim 1, characterized in that, The formula for constructing the reward function is as follows: ， ； in, For the reward function, This is a weighting factor for cleaning time. Z represents the energy consumption weighting coefficient for the inspection drone; Z represents the overall task allocation cost function. Assign weights to the overall task based on the inverse of the cost function. The sum of the distances each drone travels from its starting nest to each photovoltaic module to be cleaned, and then back to its destination nest; The sum of the distances between the starting and ending nests and The absolute value of the difference.

Images