Desert photovoltaic power station cleaning unmanned aerial vehicle control method, device and equipment

Abstract

The invention provides a desert photovoltaic power station cleaning unmanned aerial vehicle control method, device and equipment, belongs to the technical field of photovoltaic cleaning, and solves the problems that a desert photovoltaic power station is insufficient in automation degree and low in operation and maintenance efficiency. The method comprises the following steps: acquiring environmental data and multi-dimensional state data of a photovoltaic module; inputting the multi-dimensional state data and the environment data into a cleaning decision model for processing to obtain a cleaning trigger judgment result; according to the cleaning triggering judgment result, a cleaning operation instruction is generated, the cleaning unmanned aerial vehicle is controlled to execute, and the cleaning operation instruction comprises a flight route instruction and a cleaning parameter instruction; acquiring real-time monitoring data acquired by a plurality of monitoring devices on the cleaning unmanned aerial vehicle, wherein the real-time monitoring data comprises cleaning effect data and component thermal state data; and according to the real-time monitoring data, the cleaning parameter instruction is dynamically optimized through the cleaning decision model, an adjusted cleaning parameter instruction is generated, and the cleaning unmanned aerial vehicle is controlled to execute the cleaning parameter instruction. According to the scheme, waterless precise cleaning and multi-dimensional monitoring of the photovoltaic module are realized, the operation and maintenance efficiency is improved, and the cost is reduced.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic clean energy technology, and in particular to a method, device and equipment for controlling a clean drone for a desert photovoltaic power station. Background Technology

[0002] With the large-scale application of photovoltaic power generation in desert areas, the decline in power generation efficiency caused by dust accumulation on the surface of photovoltaic panels has become a core operation and maintenance problem restricting the operational efficiency of power plants. To address this issue, existing cleaning technologies have evolved from manual to mechanized, and then to preliminary automation. Early methods commonly employed manual cleaning, relying on personnel carrying tools to wipe or rinse, which was inefficient, posed high safety risks, and severely conflicted with the water-scarce desert environment. To improve efficiency, ground cleaning vehicles and tracked cleaning robots emerged. These robots operate by mechanically moving and carrying rotating brushes and spraying systems. While this improved coverage to some extent, their large size resulted in poor terrain adaptability, making them prone to getting stuck in sand, crushing fragile vegetation, and creating blind spots and secondary dust problems. In recent years, drawing on the development trend of low-altitude economic growth, drone cleaning technology has begun to be introduced into the photovoltaic operation and maintenance field. By carrying cleaning mechanisms on a flight platform, it has demonstrated excellent terrain-crossing flexibility.

[0003] However, existing drone-based cleaning solutions still suffer from a series of core shortcomings that urgently need to be addressed when dealing with the extreme desert environment and the demands of intelligent operation and maintenance. First, at the decision-making level, existing solutions lack autonomous decision-making capabilities based on big data and artificial intelligence. The initiation of cleaning operations and strategy selection rely on fixed procedures or manual instructions, failing to make accurate judgments and dynamic adjustments based on real-time data such as dust accumulation concentration and hot spot risk on photovoltaic panels. This results in cleaning operations being somewhat blind, lacking timeliness and preventative measures. Second, at the functional level, the monitoring dimensions are extremely limited. Existing equipment typically focuses only on the cleaning execution itself, without integrating hot spot detection and monitoring of the surrounding ecosystem (such as vegetation restoration), failing to meet the higher requirements of integrated operation and maintenance of desert power plants, encompassing both power generation equipment maintenance and ecological environmental protection. Furthermore, in terms of cleaning methods, most solutions still use or combine high-pressure spraying, consuming large amounts of precious water resources, fundamentally contradicting the extreme water scarcity in desert regions. Furthermore, regarding environmental adaptability, existing drone communication largely relies on public networks, resulting in poor stability in desert areas without signal. The operational capabilities and reliability of their cleaning systems at night and during sandstorms have not been fully validated. Energy utilization is also not deeply integrated with photovoltaic power generation itself. Finally, in terms of cleaning efficiency, fixed-speed fans or brushes cannot intelligently adjust their intensity based on the degree of soiling, easily leading to energy waste or incomplete cleaning. In summary, existing technologies have failed to systematically solve the complex technical problems faced by photovoltaic power plants in desert environments, including "lack of intelligent decision-making, low functional integration, resource consumption contradictions, and poor environmental adaptability," thus hindering the long-term economic, safe, and sustainable operation of photovoltaic power plants in harsh environments. Summary of the Invention

[0004] This invention provides a method, device, and equipment for controlling clean drones in desert photovoltaic power stations, which solves the problems of insufficient automation and low operation and maintenance efficiency in desert photovoltaic power stations.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: This invention provides a method for controlling a clean-up drone for a desert photovoltaic power station, comprising: Acquire environmental data and multi-dimensional status data of photovoltaic modules; The multi-dimensional state data and environmental data are input into the cleaning decision model for processing to obtain the cleaning trigger judgment result. Based on the cleaning trigger judgment result, a cleaning operation instruction is generated and the cleaning drone is controlled to execute it. The cleaning operation instruction includes a flight path instruction and a cleaning parameter instruction. Acquire real-time monitoring data collected by multiple monitoring devices on the cleaning drone, including cleaning effect data and component thermal state data; Based on the real-time monitoring data, the cleaning parameter commands are dynamically optimized through the cleaning decision model to generate adjusted cleaning parameter commands and control the cleaning drone to execute them.

[0006] Optionally, acquiring the environmental data and multi-dimensional status data of the photovoltaic module includes: Environmental data is acquired through meteorological sensors deployed in the photovoltaic power station, including real-time wind speed, ambient dust concentration, and ambient temperature. Multi-source image data of photovoltaic modules are acquired using image acquisition devices mounted on drones; The multi-source image data is extracted and processed to obtain the location data of the photovoltaic module, the first state data characterizing the degree of pollution on the surface of the photovoltaic panel, the second state data characterizing the risk of abnormal heating of the module, and the third state data characterizing the surrounding ecological conditions. The first state data is the dust coverage. C d The calculation formula is: C d = A d / A t ×100%, in, A d The pixel area of ​​the dust-covered region identified from the visible light image using an image segmentation algorithm; A t This refers to the standard reference total area of ​​photovoltaic panels.

[0007] Optionally, the cleaning decision model is obtained through the following process: Obtain historical photovoltaic power plant operation and maintenance datasets, which include environmental data collected in time series, photovoltaic module status data, manual cleaning intervention records, and performance improvement data; The environmental data and the state data of the photovoltaic module are integrated and processed to obtain the input feature vector; The data recorded from manual cleaning interventions and the performance improvement data are mapped to obtain the optimal cleaning decision label; Based on the input feature vector and the optimal cleaning decision label, a training sample set is obtained, which includes a training set and a validation set. Based on the training set, the initialized cleaning decision model is iteratively trained to obtain the target weight coefficients; The initial cleaning decision model is adjusted according to the target weight coefficients to obtain the trained cleaning decision model. Based on the validation set, the performance of the trained cleaning decision model is evaluated to obtain evaluation indicators. When the evaluation indicators reach the preset standard, the trained cleaning decision model is obtained.

[0008] Optionally, the multi-dimensional state data and environmental data are input into the cleaning decision model for processing to obtain the cleaning trigger judgment result, including: The cleaning decision model processes the multi-dimensional state and environmental data using an objective function to obtain a comprehensive evaluation value for cleaning triggering. Q The objective function is calculated as follows: Q = α · f ( C d )+ β · g ( R h )- c · h ( v , C S , T ); in, α , β , c The weight coefficients are determined through model training. f ( C d Let be a monotonically increasing function of dust cover. g ( R h Let be a mapping function for hot spot risk. h (v,Cs,T) is the penalty function characterizing the suitability of the work environment. R h For hot spot risk identification in the second state data, v For ambient wind speed, C s The concentration of dust in the environment. T The ambient temperature.

[0009] Optionally, based on the cleaning trigger determination result, a cleaning operation instruction is generated, including: When the comprehensive evaluation value of the cleaning trigger is greater than the preset trigger threshold, it is determined that the cleaning trigger condition is met, and the preset initial cleaning strategy parameters are output. The initial cleaning strategy parameters include the first fan speed and the first operating flight speed of the blower. According to the preset encoding rules, the initial cleaning strategy parameters are compiled to obtain cleaning parameter instructions; Based on the pre-stored digital map of the photovoltaic power station, path planning is performed on the location data of the photovoltaic modules to generate flight route instructions.

[0010] Optionally, real-time monitoring data collected by multiple monitoring devices on the cleaning drone can be acquired, including: The visible light image sequence on the surface of the photovoltaic panel is collected using a visible light camera mounted on a drone; The drone was equipped with an infrared thermal imaging camera to collect a sequence of temperature distribution images of the corresponding area of ​​the photovoltaic panel. Based on the visible light image sequence, determine cleaning effect data reflecting the current cleaning effect; Based on the temperature distribution image sequence, the thermal state data of the component is determined.

[0011] Optionally, based on the real-time monitoring data, the cleaning parameter instructions are dynamically optimized using the cleaning decision model to generate adjusted cleaning parameter instructions, including: The cleaning effect data is input into the cleaning decision model. Based on the cleaning effect data and a preset cleaning effect target value, the cleaning decision model determines the adjustment amount of the cleaning parameters. The calculation formula for the cleaning parameter adjustment amount is as follows: , in, To adjust the gain coefficient, The deviation normalization coefficient is... For cleaning parameter adjustment amount, The deviation between the cleaning effect data and the preset cleaning effect target value; The component thermal state data is input into the cleaning decision model. The cleaning decision model generates enhanced cleaning parameters based on the temperature exceedance level data and area data of the hot spot area in the component thermal state data. The enhanced cleaning parameters include a second flight speed, a second fan speed, and the number of reciprocating cleaning cycles. The cleaning parameter adjustment amount and the enhanced cleaning parameter are compiled and processed according to a preset code to obtain the adjusted cleaning parameter instruction.

[0012] This invention also provides a control device for a clean-up drone used in a desert photovoltaic power station, comprising: The acquisition module is used to acquire environmental data and multi-dimensional status data of photovoltaic modules; The processing module is used to input the multi-dimensional state data and environmental data into the cleaning decision model for processing to obtain the cleaning trigger judgment result; based on the cleaning trigger judgment result, it generates cleaning operation instructions and controls the cleaning drone to execute them, the cleaning operation instructions including flight route instructions and cleaning parameter instructions; The acquisition module is also used to acquire real-time monitoring data collected by multiple monitoring devices on the cleaning drone, including cleaning effect data and component thermal state data; The processing module is also used to dynamically optimize the cleaning parameter instructions based on the real-time monitoring data and the cleaning decision model, generate adjusted cleaning parameter instructions, and control the cleaning drone to execute them.

[0013] This invention also provides a computing device, including: a processor and a memory storing a computer program, wherein the computer program, when run by the processor, executes the above-described method.

[0014] This invention also provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the above-described method.

[0015] The technical solution of the present invention has at least the following effects: The above-described solution of the present invention acquires environmental data and multi-dimensional status data of photovoltaic modules; inputs the multi-dimensional status data and environmental data into a cleaning decision model for processing to obtain a cleaning trigger judgment result; generates a cleaning operation command based on the cleaning trigger judgment result and controls a cleaning drone to execute it, the cleaning operation command including flight path command and cleaning parameter command; acquires real-time monitoring data collected by multiple monitoring devices on the cleaning drone, the real-time monitoring data including cleaning effect data and module thermal status data; and dynamically optimizes the cleaning parameter command through the cleaning decision model based on the real-time monitoring data, generating adjusted cleaning parameter commands and controlling the cleaning drone to execute them. This solution achieves waterless and precise cleaning and multi-dimensional monitoring of photovoltaic modules, improving operation and maintenance efficiency and reducing costs. Attached Figure Description

[0016] Figure 1 This is a flowchart of the desert photovoltaic power station cleaning drone control method provided in an embodiment of the present invention; Figure 2 This is a structural diagram of the desert photovoltaic power station clean drone control device provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of the computing device provided in an embodiment of the present invention. Detailed Implementation

[0017] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this invention will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.

[0018] like Figure 1 As shown, an embodiment of the present invention proposes a control method for a clean-up drone for a desert photovoltaic power station, comprising: Step 11: Obtain environmental data and multi-dimensional status data of the photovoltaic modules; Step 12: Input the multi-dimensional state data and environmental data into the cleaning decision model for processing to obtain the cleaning trigger judgment result; Step 13: Based on the cleaning trigger judgment result, generate a cleaning operation instruction and control the cleaning drone to execute it. The cleaning operation instruction includes a flight path instruction and a cleaning parameter instruction. Step 14: Obtain real-time monitoring data collected by multiple monitoring devices on the cleaning drone, including cleaning effect data and component thermal state data; Step 15: Based on the real-time monitoring data, dynamically optimize the cleaning parameter instructions through the cleaning decision model, generate adjusted cleaning parameter instructions, and control the cleaning drone to execute them.

[0019] In step 11 of this embodiment, environmental data and multi-dimensional status data of the photovoltaic modules are acquired. Environmental data includes temperature, humidity, wind speed, and dust concentration in desert areas, collected in real time by meteorological sensors deployed in the photovoltaic power station and airborne sensors on drones. Multi-dimensional status data includes the surface dirt coverage of the photovoltaic modules, the rate of power generation efficiency degradation, and the module temperature distribution. Dirt coverage is quantified by capturing surface images with a high-definition camera and using image processing algorithms. The rate of power generation efficiency degradation is calculated by comparing the inverter's output power with a standard power. The module temperature distribution is obtained through infrared thermal imaging. These data collectively constitute the basic input for cleanliness decisions.

[0020] In step 12, the cleaning decision model is a mathematical model based on multivariate analysis, which assesses the necessity of cleaning by calculating a cleaning demand index. When the cleaning demand index exceeds a preset threshold, the model outputs a cleaning trigger judgment result indicating that cleaning is required; otherwise, cleaning is not required.

[0021] In step 13, when the system determines that cleaning is required, it generates a cleaning operation instruction, which includes a flight path instruction and cleaning parameter instructions. The flight path instruction, based on the latitude and longitude coordinates of the photovoltaic array and the component arrangement topology, uses a path planning algorithm to generate the optimal flight path, ensuring coverage of all components to be cleaned. The cleaning parameter instructions include the rotational speed of the cleaning brush, the water volume per unit area of ​​the water spray system, and the spray pressure; these parameters are initially set according to the type and degree of dirt. The instruction is transmitted to the cleaning drone via a wireless communication network, and the drone autonomously executes the cleaning operation according to the instructions.

[0022] In step 14, the monitoring device includes a high-definition camera and an infrared thermal imager (camera). The high-definition camera continuously captures images of the component surface during the cleaning process and extracts cleaning effect data, i.e., dirt residue rate, through image processing algorithms. The infrared thermal imager simultaneously scans the component surface to obtain component thermal state data, i.e., a temperature distribution map, which is used to identify hot spots caused by dirt accumulation. This real-time monitoring data is transmitted back to the control system via the UAV's onboard communication module.

[0023] In step 15, based on real-time monitoring data, the cleaning parameter commands are dynamically optimized using a cleaning decision model, generating adjusted cleaning parameter commands and controlling the cleaning drone to execute them. The dynamic optimization process is based on feedback control principles; the cleaning decision model adjusts the cleaning parameters according to real-time cleaning effect data and component thermal state data. For example, when the dirt residue rate is higher than the target value, the model proportionally increases the water spray volume or cleaning brush pressure; when the component hotspot temperature exceeds a safety threshold, the model reduces the cleaning speed or pauses cleaning to avoid thermal damage. The optimized cleaning parameter commands are updated in real time and sent to the drone, achieving adaptive cleaning control.

[0024] The technical solution described in this embodiment achieves integrated intelligent operation and maintenance of desert photovoltaic power stations by constructing an analysis and decision-making system centered on big data, integrating multiple cameras for collaborative monitoring, and incorporating a dry-type variable frequency purging system and a fully automated control module. This solution first automatically determines cleaning trigger conditions and matches purging parameters through multi-source data collection and model analysis. During execution, it dynamically adjusts the purging strategy based on real-time feedback from multiple cameras regarding cleaning effectiveness, hot spot status, and vegetation conditions. Employing waterless variable frequency purging technology completely avoids water consumption and optimizes energy consumption through frequency conversion. This technical solution achieves precise cleaning decisions, multi-dimensional operation and maintenance monitoring, unmanned operation, and minimal resource consumption, significantly improving the operational reliability and ecological sustainability of photovoltaic power stations in extreme environments.

[0025] In an optional embodiment of the present invention, step 11, obtaining environmental data and multi-dimensional status data of the photovoltaic module, may include: Step 111: Obtain environmental data through meteorological sensors installed in the photovoltaic power station. The environmental data includes real-time wind speed, ambient dust concentration, and ambient temperature. Step 112: Acquire multi-source image data of the photovoltaic modules using the image acquisition device mounted on the drone; Step 113: Extract and process the multi-source image data to obtain the location data of the photovoltaic module, first state data characterizing the degree of surface pollution of the photovoltaic panel, second state data characterizing the risk of abnormal heating of the module, and third state data characterizing the surrounding ecological conditions; wherein, the first state data is the dust coverage. C dThe calculation formula is: C d = A d / A t ×100%, in, A d The pixel area of ​​the dust-covered region identified from the visible light image using an image segmentation algorithm; A t This refers to the standard reference total area of ​​photovoltaic panels.

[0026] In step 111 of this embodiment, environmental data is collected and transmitted in real time by a meteorological sensor network deployed within the photovoltaic power station. Specifically, real-time wind speed is obtained using ultrasonic anemometers installed on weather towers or key monitoring points; this data is used to determine the safety of wind conditions during drone take-off, landing, and operations. Environmental dust concentration is obtained using laser dust meters or beta-ray particulate matter monitors; this concentration value is used to assess the intensity of external environmental pollution input. Environmental temperature is obtained using weather-resistant temperature sensors; this data is used to confirm the operating temperature range of the equipment, ensuring that electronic components and batteries operate under permissible conditions. All sensor data is aggregated to the power station monitoring data platform via IoT nodes and used as a real-time data stream for subsequent analysis.

[0027] In step 112, multi-source image data is simultaneously acquired by a heterogeneous image acquisition device carried by the UAV during its inspection flight. The UAV flies over the photovoltaic array along a pre-programmed inspection route, and its gimbal integrates a high-resolution visible light camera, an uncooled infrared thermal imaging camera, and a multispectral camera. The visible light camera captures color images of the photovoltaic panel surface at fixed intervals for visual analysis; the infrared thermal imaging camera simultaneously captures temperature distribution heat maps within the same field of view for thermal feature analysis; and the multispectral camera acquires spectral images containing specific bands such as red-edge and near-infrared for ground feature classification and vegetation analysis. The three types of images are aligned and correlated using a unified spatiotemporal stamp during acquisition.

[0028] In step 113, the synchronously acquired multi-source image data undergoes automated extraction processing. First, based on visible light images, a pre-trained semantic segmentation deep learning model identifies the precise outline of each photovoltaic panel. Combined with UAV GPS and visual odometry data, the geographical coordinates of each photovoltaic panel within the power station are calculated, forming location data. Second, within the same photovoltaic panel outline area, the semantic segmentation model further classifies pixels into clean glass areas and dust-covered areas. By accumulating the total number of pixels in the dust-covered area and multiplying it by the actual area corresponding to a single pixel, the area of ​​the dust-covered area is calculated. A d Standard reference total area for photovoltaic panelsA t These are pre-set fixed parameters based on the photovoltaic panel model and specifications. Dust cover. C d The result is calculated using the formula: C d = A d / A t ×100%, This percentage value represents the first state data characterizing the degree of surface contamination. Next, by analyzing the infrared thermogram of the corresponding area, the presence of localized high-temperature points significantly higher than the surrounding average temperature is detected. Combining these temperature values ​​with spatial distribution characteristics, second state data characterizing the risk level of abnormal heating of the components is generated. Finally, by calculating vegetation indices, such as the Normalized Difference Vegetation Index (NDVI), from multispectral images, the quantified vegetation activity value of the ground surface or protective belt surrounding the photovoltaic array is obtained, serving as the third state data characterizing the surrounding ecological condition. All extracted state data and environmental data are packaged into a structured data package for use by subsequent decision-making models.

[0029] In an optional embodiment of the present invention, the cleaning decision model is obtained through the following process: Step 121: Obtain historical photovoltaic power plant operation and maintenance dataset. The historical photovoltaic power plant operation and maintenance dataset includes environmental data collected in time series, photovoltaic module status data, manual cleaning intervention record data, and performance improvement data. Step 122: Integrate the environmental data and the state data of the photovoltaic module to obtain the input feature vector; Step 123: Map the recorded data of manual cleaning intervention and the performance improvement data to obtain the optimal cleaning decision label; Step 124: Based on the input feature vector and the optimal cleaning decision label, a training sample set is obtained, which includes a training set and a validation set. Step 125: Based on the training set, iteratively train the initialized cleaning decision model to obtain the target weight coefficients; Step 126: Adjust the initialized cleaning decision model according to the target weight coefficients to obtain the trained cleaning decision model; Step 127: Based on the validation set, perform performance evaluation on the trained cleaning decision model to obtain evaluation indicators. When the evaluation indicators reach the preset standard, the trained cleaning decision model is obtained.

[0030] In step 121 of this embodiment, the specific process of obtaining the historical photovoltaic power plant operation and maintenance dataset is as follows: historical data records from the past few months to several years are extracted from the central monitoring database of the photovoltaic power plant. These data are strictly aligned by timestamps to form a continuous time series. Environmental data comes from the weather station deployed within the power plant, including real-time wind speed recorded every five minutes. v Environmental dust concentration C s and ambient temperature T The status data for photovoltaic modules comes from historical manual inspection reports and early rudimentary monitoring equipment, including dust cover estimated through visible light photographs. C d Hot spot risk markings identified using a handheld infrared thermal imager R h The system also includes a general overview of vegetation cover recorded through regular photography. Manual cleaning intervention data is extracted from the maintenance work order system, detailing the initiation time, triggering reason, specific location, and method used for each manual cleaning operation. Performance improvement data is calculated by correlating power generation monitoring data of the same photovoltaic string before and after cleaning, quantifying the percentage of power generation efficiency recovery resulting from the cleaning action.

[0031] In step 122, the method for integrating the environmental and state data to obtain the input feature vector is as follows: Taking each candidate time point in the historical data that might trigger a cleaning decision as the center, extract all environmental and state data sequences within a fixed time period preceding it. For each data sequence, such as dust cover... C d Calculate the statistical characteristics within this time window, such as the mean, maximum, standard deviation, and rate of change over the most recent period. Concatenate the calculated statistical characteristics of all environmental and state variables into a high-dimensional numerical vector in a predefined order. This vector comprehensively represents the panoramic information of the environmental conditions faced by the power plant and the health status of the photovoltaic modules at a specific historical moment; this is the input feature vector.

[0032] In step 123, the logic for mapping manual cleaning intervention records and performance improvement data to obtain the optimal cleaning decision label is as follows: For each candidate time point and its input feature vector, analyze the maintenance records for a period of time afterward. If effective manual cleaning occurs shortly after the time point and performance improvement data confirms a significant improvement in power generation efficiency, then that time point is determined to be the correct decision point of "cleaning should be done," and a label indicating "triggered cleaning" is generated for it, with the actual intensity of manual cleaning used as a reference. If cleaning is not done for a long time after the time point and performance data continues to decline, and subsequent analysis confirms it as a "missed cleaning" error, then the label "triggered cleaning" is also assigned to it. If the state is stable after the time point and cleaning is not required, then the label "no cleaning triggered" is assigned. This process transforms complex historical maintenance experience into clear decision labels that can be used for supervised learning.

[0033] In step 124, the method for constructing the training sample set based on the input feature vector and the optimal clean decision label is as follows: each pair consisting of the input feature vector and its corresponding decision label is called a training sample. After arranging all historical samples in chronological order, the first 70% to 80% of the samples are typically divided into the training set for learning the model parameters. The remaining samples serve as the validation set, which does not participate in parameter updates during the training process but is only used to periodically evaluate the model's generalization ability and prevent the model from over-memorizing specific patterns from the training data.

[0034] In step 125, the iterative training process of the initialized cleaning decision model based on the training set is as follows: First, a multi-layer neural network model is initialized, whose input layer dimension matches the feature vector, and whose output layer can provide cleaning trigger judgments and policy parameters. The core part of the model contains a learnable objective function used to calculate the comprehensive evaluation value Q of cleaning triggers, and its expression is: Q = α · f ( C d )+ β · g ( R h )- c · h ( v , C S , T ); in, α , β , c These are the weight coefficients determined through model training, used to balance the importance of different influencing factors. f ( C d(This is about dust cover) C d A monotonically increasing function, such as a linear function or a saturated growth function, indicates that the more severe the pollution, the higher the tendency to clean. g ( R h This is about hot spot risk markings. R h The mapping function outputs a higher positive value when there is a risk of hot spots. h (v, C s (T) is the penalty function, whose value varies with wind speed. v Environmental dust concentration C s Increase or temperature T An increase exceeding the suitable range indicates the constraints of harsh environments on operations. During training, training set samples are input into the model, and forward propagation is used to calculate the predicted values. Q The values ​​are compared with the true labels, and the weight coefficients within the model are continuously adjusted through backpropagation and gradient descent optimizers. α , β , c Other connection weights are added to minimize the prediction error. The coefficients determined after iteration until the loss function converges are the target weight coefficients.

[0035] In step 126, the process of adjusting the initialized model based on the target weight coefficients is a direct parameter substitution. This involves replacing all target weight coefficients obtained after training convergence, including those in the objective function. α , β , c The connection weights of each layer of the neural network are fully written into the initialized model structure, overwriting the original random initial values. This results in a neural network model with optimized and finalized parameters, possessing decision-making capabilities—the trained clean decision model.

[0036] In step 127, the method for evaluating the performance of the trained cleaning decision model based on the validation set is as follows: all input feature vectors in the validation set are input into the tuned model, and the model outputs a comprehensive evaluation value for cleaning triggering for each sample. Q . This Q The value is compared with an adjustable threshold to obtain a binary judgment of "trigger" or "do not trigger". This judgment is then compared with the true optimal cleaning decision label of the validation sample to calculate evaluation metrics such as accuracy, recall, and F1 score. If key metrics, such as the F1 score, reach a pre-set high standard, such as above 0.95, it indicates that the model has good generalization ability and decision accuracy, and the model training can be considered complete and usable in a real system. If the standard is not met, it is necessary to go back and adjust the model structure or training parameters and retrain.

[0037] In an optional embodiment of the present invention, step 13, generating a cleaning operation instruction based on the cleaning trigger determination result, may include: Step 131: When the comprehensive evaluation value of the cleaning trigger is greater than the preset trigger threshold, it is determined that the cleaning trigger condition is met, and the preset initial cleaning strategy parameters are output. The initial cleaning strategy parameters include the first fan speed and the first operating flight speed of the blower. Step 132: Compile the initial cleaning strategy parameters according to the preset encoding rules to obtain cleaning parameter instructions; Step 133: Based on the pre-stored digital map of the photovoltaic power station, perform path planning on the location data of the photovoltaic modules to obtain the flight route command.

[0038] In step 131 of this embodiment, the cleaning trigger determination is the logical starting point for generating the work instruction. The cleaning trigger comprehensive evaluation value... Q The cleaning decision model described above quantifies the necessity and urgency of immediately executing cleaning operations given the current multi-dimensional state and environmental data. A preset trigger threshold is also included. Q th It is a threshold value set through historical data statistical analysis or expert experience, used to define a clear decision boundary within a continuous set of evaluation values. When the calculated... Q Value greater than Q th At this point, the system logic determines that the current operating conditions meet the trigger conditions for automatic cleaning. Instead of simply outputting a "yes / no" signal, the system simultaneously outputs a set of preset initial cleaning strategy parameters. These parameters serve as a baseline operating guide, directly instructing the initial actions of the cleaning actuators and the aircraft. The rotational speed of the first fan is determined based on the dust coverage. C d The initial assessment level is matched to the airflow speed, with higher pollution levels corresponding to higher initial rotation speeds to ensure sufficient purging capacity at startup. The initial flight speed is set by taking into account both the photovoltaic array's density and the current wind speed, aiming to maintain flight stability while ensuring efficient cleaning coverage.

[0039] Step 132, which involves compiling the initial cleaning strategy parameters, is a crucial step in transforming high-level decision-making instructions into recognizable and executable commands for the lower-level control system. Predefined encoding rules define the data structure, field order, verification method, and transmission protocol of the instructions. For example, the values ​​of the first fan speed and the first operational flight speed, along with the instruction type identifier and target device address, are combined in a fixed byte order, and a cyclic redundancy check (CRC) code is calculated and appended to the end of the data packet. This compilation process ensures that the generated cleaning parameter instructions possess clear semantics, complete control information, and high transmission reliability, enabling them to be flawlessly transmitted to the UAV flight controller and variable frequency fan driver via the communication link, driving them to execute actions according to the instructions.

[0040] In step 133, the generation of flight path instructions relies on precise spatial information and intelligent planning algorithms. The pre-stored digital map of the photovoltaic power station contains the precise geographic coordinates, component dimensions, array arrangement, and obstacle information of all photovoltaic modules within the power station. The location data of the photovoltaic modules indicates the specific set of target panels to be covered in this cleaning task. The path planning algorithm uses this information as input, and its core objective is to generate an optimal or suboptimal flight trajectory that starts from the takeoff point, efficiently covers all cleaning positions directly above the target photovoltaic panels, and finally returns to the landing point. The planning process must adhere to several constraints, including but not limited to: safe flight distances, UAV turning performance limitations, and ensuring that the sweeping airflow effectively covers the panel surface. The final generated flight path instructions are a series of ordered spatial coordinate points, or directly a specific waypoint file format that the UAV flight control system can parse, used to guide the UAV to complete autonomous and precise fixed-point flight operations.

[0041] In an optional embodiment of the present invention, step 14, acquiring real-time monitoring data collected by multiple monitoring devices on the cleaning drone, may include: Step 141: Collect a sequence of visible light images of the photovoltaic panel surface using a visible light camera mounted on a drone; Step 142: Use the infrared thermal imaging camera mounted on the drone to collect a sequence of temperature distribution images of the corresponding area of ​​the photovoltaic panel; Step 143: Determine cleaning effect data reflecting the current cleaning effect based on the visible light image sequence; Step 144: Determine the component thermal state data based on the temperature distribution image sequence.

[0042] In step 141 of this embodiment, a visible light image sequence of the photovoltaic panel surface is acquired using a visible light camera mounted on a drone. Specifically, after the drone flies to a preset cleaning altitude above the target photovoltaic panel according to flight path instructions, the visible light camera continuously captures images at fixed time intervals or triggered by the drone's flight position. The camera's optical axis is adjusted to be approximately perpendicular to the photovoltaic panel plane to ensure that the acquired images completely and directly reflect the panel's surface condition. The acquired visible light image sequence strictly corresponds to the drone's flight trajectory in time and space, forming a set of continuous frame data recording the changes in the visual state of the photovoltaic panel surface over time or position, providing a raw visual information source for subsequent quantitative analysis of the cleaning effect.

[0043] In step 142 of this embodiment, an infrared thermal imaging camera mounted on a drone simultaneously acquires a sequence of temperature distribution images of the corresponding area of ​​the photovoltaic panel. The infrared camera and the visible light camera are physically calibrated so that their fields of view are spatially largely overlapped. At the same moment the visible light camera captures an image, the infrared thermal imaging camera simultaneously captures the infrared radiation information of the target area and generates a two-dimensional thermal image using different colors or grayscale values ​​to represent temperature levels. The resulting thermal image sequence is strictly synchronized in time and spatially registered with the visible light image sequence, thus enabling a precise correlation between the visual characteristics of the photovoltaic panel surface and its thermodynamic state, providing a direct temperature measurement data basis for hot spot detection.

[0044] In step 143 of this embodiment, cleaning effect data reflecting the current cleaning effect is determined based on the visible light image sequence. This process is achieved by analyzing the real-time acquired visible light images. First, an image processing algorithm is used to segment the main area of ​​the photovoltaic panel from a single frame image. Then, within this area, the proportion of area still attached to the panel surface is identified and quantified by calculating the texture features and grayscale distribution of pixels in the image, or by comparing it with a pre-stored clean panel surface reference image. This proportion is calculated as the current real-time cleanliness, or by comparing it with the initial cleanliness before the cleaning operation begins, the immediate efficiency improvement percentage of this cleaning operation is calculated. This quantification result is the cleaning effect data, which objectively reflects the real-time effectiveness of the purging operation.

[0045] In step 144 of this embodiment, the thermal status data of the module is determined based on the temperature distribution image sequence. This process is based on the analysis of synchronously acquired infrared thermal image sequences. For each thermal image, the temperature distribution within the identified photovoltaic panel area is statistically analyzed, and the average temperature is calculated as a baseline. Next, all localized high-temperature points or areas with temperatures significantly higher than the average temperature of the area and exceeding a preset safety threshold are detected and located. For these abnormal areas, their specific location coordinates on the photovoltaic panel, their highest temperature value, and the difference between their abnormal temperature and the average temperature of the area are recorded. This information collectively constitutes the module's thermal status data, which directly indicates the potential hot spot effect caused by dirt shading and its severity. It is a key basis for assessing the operational safety of the photovoltaic module and determining whether intensive cleaning or subsequent maintenance is required.

[0046] In an optional embodiment of the present invention, step 15, dynamically optimizing the cleaning parameter instructions based on the real-time monitoring data using the cleaning decision model to generate adjusted cleaning parameter instructions, may include: Step 151: Input the cleaning effect data into the cleaning decision model. The cleaning decision model determines the adjustment amount of the cleaning parameters based on the cleaning effect data and the preset cleaning effect target value. The calculation formula for the adjustment amount of the cleaning parameters is: Δ F = or ·tanh(Δ E / s ), in, or To adjust the gain coefficient, s Δ is the deviation normalization coefficient. F Δ is the adjustment amount for cleaning parameters. E The deviation between the cleaning effect data and the preset cleaning effect target value; Step 152: Input the component thermal state data into the cleaning decision model. The cleaning decision model generates enhanced cleaning parameters based on the temperature exceedance level data and area data of the hot spot area in the component thermal state data. The enhanced cleaning parameters include the second flight speed, the second fan speed, and the number of reciprocating cleaning cycles. Step 153: Compile the cleaning parameter adjustment amount and the enhanced cleaning parameter according to the preset code to obtain the adjusted cleaning parameter instruction.

[0047] In step 151 of this embodiment, the real-time acquired cleaning effect data is input into the cleaning decision model to initiate closed-loop dynamic optimization of the cleaning parameters. The cleaning effect data is typically represented as the percentage of cleanliness of the currently cleaned area or the improvement rate compared to the state before cleaning. The model internally compares its real-time values ​​with preset cleaning effect target values ​​to calculate the instantaneous deviation. This deviation reflects the gap between the current operational results and the expected standards. Cleaning parameter adjustment amount. This deviation is calculated using a specific feedback control formula, which is as follows: , in, The gain coefficient determines the strength of the system's response to deviations. This is the deviation normalization coefficient, used to normalize the deviation. Scale to the appropriate size; The value represents the adjustment amount that needs to be applied to the current fan speed. A positive value indicates that purging needs to be strengthened, while a negative value indicates that it can be weakened to save energy. The formula uses the hyperbolic tangent function tanh as the core mapping relationship because this function has smooth and bounded characteristics, which can ensure that the adjustment amount increases smoothly with the increase of deviation and eventually tends to saturate. This avoids violent oscillations in control commands caused by error fluctuations, and achieves a stable balance between pursuing cleaning effect and preventing energy waste.

[0048] In step 152, component thermal state data is input into the cleaning decision model to address hotspot risks in a targeted manner. The model analyzes the temperature exceedance level and spatial area of ​​the hotspot regions identified in the input data. Higher temperature exceedance levels and larger areas indicate a stronger thermal effect likely caused by severe dirt obstruction, requiring more thorough cleaning intervention. Based on this assessment, the model generates a set of enhanced cleaning parameters. The second flight speed is set below the baseline operating speed to ensure the drone has a longer dwell time above the hotspot area. The second fan speed is set above the current operating speed to provide stronger instantaneous blowing force to remove potentially more stubborn dust. Simultaneously, the model generates a reciprocating cleaning command greater than 1, controlling the drone to perform multiple reciprocating cleaning cycles in the area to ensure cleaning effectiveness. These parameters together constitute an enhanced cleaning sub-strategy focused on local anomalies.

[0049] In step 153, the calculated general cleaning parameter adjustment amount is... Integrate with enhanced cleaning parameters for specific areas. The integration logic is: first, combine... A new global speed setpoint is obtained by applying the current baseline fan speed. Then, for specific locations with hot spots, enhanced parameters such as the second fan speed and second flight speed at that location are temporarily overlaid or superimposed on the global setpoint. Finally, following the same preset coding rules as the initial command generation, these adjusted and optimized parameter values, along with their target location or time period information, are recompiled into a new, structured data command package. This data package is the adjusted cleaning parameter command, which contains comprehensive control information of global optimization and local enhancement, and is executed through the communication link to achieve precise and adaptive control of the cleaning operation process.

[0050] A specific embodiment of the desert photovoltaic power station cleaning drone control method provided by the present invention is as follows: Step 1: Multidimensional data collection and quantitative extraction.

[0051] The system acquires real-time environmental data, including wind speed, through meteorological sensors at the photovoltaic power station. v Environmental dust concentration C s and ambient temperature T Simultaneously, the drone, equipped with visible light, infrared, and multispectral cameras, photographed the photovoltaic modules. Image processing algorithms were then used to segment and calculate the area of ​​the dust-covered region from the visible light images. A d Combined with the standard total area of ​​photovoltaic panels A t According to the formula: C d = A d / A t ×100% to calculate the dust cover. C d This serves as the first state data. Hotspot risk areas are identified and marked from the infrared image, generating hotspot risk labels. R h This serves as the second state data.

[0052] Step 2: Intelligent decision triggering and initial strategy generation.

[0053] Environmental and state data are input into a pre-trained cleaning decision model. The model's core decision function calculates a comprehensive evaluation value for cleaning triggers. Q = α · f ( C d )+ β · g ( R h )- c · h ( v , C S , T );in, α , β , c The weight coefficients are determined through model training. f ( C d Let be a monotonically increasing function of dust cover. g ( R h Let be a mapping function for hot spot risk. h (v, C S Let T be a penalty function characterizing the suitability of the work environment. Q Greater than the preset threshold Q th (dust cover) There may be a risk of hot spots. And wind speed Level, temperature Dust concentration If the current condition is positive, then cleaning is triggered. The model determines this based on the current condition. C d and C S Through the relation: ; Calculate the reference speed of the purging blower. F base This generates an initial cleaning operation instruction that includes this parameter and a reference flight speed; among which, and The scaling factor is a constant determined through model training and calibration, used to adjust the dust coverage on the plate surface. With environmental dust concentration Weighting of the impact on the wind turbine's reference speed; This is a characteristic coefficient of dust cover, used to measure dust cover. Perform normalization processing; This is a characteristic coefficient of environmental dust concentration, used to analyze environmental dust concentration. Perform normalization processing; The minimum speed threshold for maintaining effective clean airflow is a constant determined by the fan's own performance and the minimum effective working air pressure; The model then generates initial cleaning operation instructions. The instruction consists of two parts: first, the optimal flight path planned based on the digital map of the photovoltaic power station and the location of the target components; and second, the baseline cleaning parameters matched according to the initial dust coverage, including the initial rotation speed of the blower and the drone's cruising speed. The instruction is encoded and transmitted wirelessly to the standby cleaning drone.

[0054] Step 3: Automated job execution and collaborative monitoring.

[0055] After receiving instructions, the drone takes off autonomously and operates along the planned route. The cleaning execution module operates at a reference speed. F base Dry purging is initiated. During the operation, three cameras work together to monitor the process: a visible light camera captures images after cleaning, an infrared camera captures temperature distribution images, and a multispectral camera captures ambient spectral images. Real-time image data is transmitted back to the edge computing unit for preliminary analysis.

[0056] Step 4: Real-time performance evaluation and feature extraction.

[0057] Online processing is performed on the real-time transmitted visible light images to calculate the cleanliness of the current cleaning area and compare it with the target value of the cleaning effect to obtain the real-time deviation. Infrared images are analyzed to identify pixel areas with temperatures exceeding a threshold, and their area and highest temperature are recorded to generate hotspot state vectors. The Normalized Difference Vegetation Index (NDVI) is calculated from multispectral images as ecological monitoring data.

[0058] Step 5: Dynamically optimize decision-making and parameter regeneration.

[0059] The real-time deviation and hotspot state vector are input into the dynamic adjustment submodule of the cleaning decision model. This module first adjusts the model based on the cleaning deviation and the formula... Calculate the global adjustment amount of the fan speed. ,in This is the gain coefficient. This is the normalization coefficient. Next, for the identified hotspot areas, based on their area and the degree of temperature exceeding the limit, local enhancement commands are generated, including reducing flight speed, setting specific fan speeds, and specifying the number of reciprocating cleaning cycles. N hot .

[0060] Step 6: Closed-loop adaptive cleaning control.

[0061] The UAV flight control and execution system receives and executes optimized commands. The wind turbine speed is globally adjusted to... A localized enhanced cleaning strategy was implemented for the hot spot area: the drone flew over the area at a reduced speed, while the fan switched to a higher speed and cleaned along the area. N hotThis system performs repeated purging flights, achieving an online, real-time closed loop of "monitoring-decision-execution".

[0062] Step 7: Operation termination verification and autonomous return.

[0063] Once all planned areas have been cleaned, and the real-time monitored cleanliness level consistently exceeds the target value with no new hot spots, the model determines the operation is complete. The drone automatically returns to base. During the return journey, the system generates an operation summary report containing the final cleanliness level of each area, hot spot elimination status, and energy consumption.

[0064] Step 8: Data archiving and model iterative updates.

[0065] The complete operation data package, including initial data, all intermediate instructions, real-time monitoring sequences, and the final report, is transmitted to the cloud data center for storage. Using this new data, the weight coefficients α, β, γ, and function parameters in the cleaning decision model are incrementally trained, updating the model parameters and achieving continuous self-optimization of model performance, thus completing a full-cycle intelligent operation and maintenance closed loop.

[0066] This invention proposes the above-mentioned technical solution, which constructs a comprehensive evaluation model for clean-triggered operation and maintenance that integrates multi-source sensing data and embeds specific decision-making formulas, and designs a dynamic parameter adjustment algorithm based on real-time effect feedback. This achieves full-process automation, precise decision-making, and energy efficiency optimization for clean operation and maintenance of photovoltaic power plants in a water-scarce desert environment, significantly improving operation and maintenance efficiency and power generation benefits.

[0067] like Figure 2 As shown, this embodiment of the invention also provides a desert photovoltaic power station clean drone control device 20, comprising: Module 21 is used to acquire environmental data and multi-dimensional status data of photovoltaic modules; Processing module 22 is used to input the multi-dimensional state data and environmental data into the cleaning decision model for processing to obtain the cleaning trigger judgment result; based on the cleaning trigger judgment result, it generates cleaning operation instructions and controls the cleaning drone to execute them, the cleaning operation instructions including flight route instructions and cleaning parameter instructions; The acquisition module 21 is also used to acquire real-time monitoring data collected by multiple monitoring devices on the cleaning drone, the real-time monitoring data including cleaning effect data and component thermal state data; The processing module 22 is also used to dynamically optimize the cleaning parameter instructions based on the real-time monitoring data and the cleaning decision model, generate adjusted cleaning parameter instructions, and control the cleaning drone to execute them.

[0068] Optionally, module 21 is specifically used for: Environmental data is acquired through meteorological sensors deployed in the photovoltaic power station, including real-time wind speed, ambient dust concentration, and ambient temperature. Multi-source image data of photovoltaic modules are acquired using image acquisition devices mounted on drones; The multi-source image data is extracted and processed to obtain the location data of the photovoltaic module, the first state data characterizing the degree of pollution on the surface of the photovoltaic panel, the second state data characterizing the risk of abnormal heating of the module, and the third state data characterizing the surrounding ecological conditions. The first state data is the dust coverage. C d The calculation formula is: C d = A d / A t ×100%, in, A d The pixel area of ​​the dust-covered region identified from the visible light image using an image segmentation algorithm; A t This refers to the standard reference total area of ​​photovoltaic panels.

[0069] Optionally, the cleaning decision model is obtained through the following process: Obtain historical photovoltaic power plant operation and maintenance datasets, which include environmental data collected in time series, photovoltaic module status data, manual cleaning intervention records, and performance improvement data; The environmental data and the state data of the photovoltaic module are integrated and processed to obtain the input feature vector; The data recorded from manual cleaning interventions and the performance improvement data are mapped to obtain the optimal cleaning decision label; Based on the input feature vector and the optimal cleaning decision label, a training sample set is obtained, which includes a training set and a validation set. Based on the training set, the initialized cleaning decision model is iteratively trained to obtain the target weight coefficients; The initial cleaning decision model is adjusted according to the target weight coefficients to obtain the trained cleaning decision model. Based on the validation set, the performance of the trained cleaning decision model is evaluated to obtain evaluation indicators. When the evaluation indicators reach the preset standard, the trained cleaning decision model is obtained.

[0070] Optionally, processing module 22 is specifically used for: The cleaning decision model processes the multi-dimensional state and environmental data using an objective function to obtain a comprehensive evaluation value for cleaning triggering. Q The objective function is calculated as follows: Q = α · f ( C d )+ β · g ( R h )- c · h ( v , C S , T ); in, α , β , c The weight coefficients are determined through model training. f ( C d Let be a monotonically increasing function of dust cover. g ( R h Let be a mapping function for hot spot risk. h (v,Cs,T) is the penalty function characterizing the suitability of the work environment. R h For hot spot risk identification in the second state data, v For ambient wind speed, C s The concentration of dust in the environment. T The ambient temperature.

[0071] Optionally, the processing module 22 is also specifically used for: When the comprehensive evaluation value of the cleaning trigger is greater than the preset trigger threshold, it is determined that the cleaning trigger condition is met, and the preset initial cleaning strategy parameters are output. The initial cleaning strategy parameters include the first fan speed and the first operating flight speed of the blower. According to the preset encoding rules, the initial cleaning strategy parameters are compiled to obtain cleaning parameter instructions; Based on the pre-stored digital map of the photovoltaic power station, path planning is performed on the location data of the photovoltaic modules to generate flight route instructions.

[0072] Optionally, module 21 is also specifically used for: The visible light image sequence on the surface of the photovoltaic panel is collected using a visible light camera mounted on a drone; The drone was equipped with an infrared thermal imaging camera to collect a sequence of temperature distribution images of the corresponding area of ​​the photovoltaic panel. Based on the visible light image sequence, determine cleaning effect data reflecting the current cleaning effect; Based on the temperature distribution image sequence, the thermal state data of the component is determined.

[0073] Optionally, the processing module 22 is also specifically used for: The cleaning effect data is input into the cleaning decision model. Based on the cleaning effect data and a preset cleaning effect target value, the cleaning decision model determines the adjustment amount of the cleaning parameters. The calculation formula for the cleaning parameter adjustment amount is as follows: , in, To adjust the gain coefficient, The deviation normalization coefficient is... For cleaning parameter adjustment amount, The deviation between the cleaning effect data and the preset cleaning effect target value; The component thermal state data is input into the cleaning decision model. The cleaning decision model generates enhanced cleaning parameters based on the temperature exceedance level data and area data of the hot spot area in the component thermal state data. The enhanced cleaning parameters include a second flight speed, a second fan speed, and the number of reciprocating cleaning cycles. The cleaning parameter adjustment amount and the enhanced cleaning parameter are compiled and processed according to a preset code to obtain the adjusted cleaning parameter instruction.

[0074] It should be noted that this device is a device corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0075] like Figure 3 As shown, this embodiment of the invention also provides a computing device 30, including a processor 31, a memory 32, and a program or instructions stored in the memory 32 and executable on the processor 31. When the program or instructions are executed by the processor 31, they implement the various processes of the above-described desert photovoltaic power station clean drone control method embodiment and achieve the same technical effects. To avoid repetition, further details are omitted here. It should be noted that the computing device in this embodiment of the invention includes the aforementioned mobile electronic devices and non-mobile electronic devices.

[0076] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0077] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0078] In the embodiments provided by this invention, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0079] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0080] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0081] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, ROM, RAM, magnetic disks, or optical disks.

[0082] Furthermore, it should be noted that in the apparatus and method of the present invention, it is obvious that the components or steps can be decomposed and / or recombined. These decompositions and / or recombinations should be considered equivalent solutions of the present invention. Moreover, the steps performing the above series of processes can naturally be executed in the order described, but are not necessarily required to be executed in chronological order; some steps can be executed in parallel or independently of each other. Those skilled in the art will understand that all or any step or component of the method and apparatus of the present invention can be implemented in any computing device (including processors, storage media, etc.) or network of computing devices, in hardware, firmware, software, or a combination thereof. This is something that those skilled in the art can achieve using their basic programming skills after reading the description of the present invention.

[0083] Therefore, the object of the present invention can also be achieved by running a program or a set of programs on any computing device. The computing device can be a known general-purpose device. Therefore, the object of the present invention can also be achieved simply by providing a program product containing program code for implementing the method or apparatus. That is, such a program product also constitutes the present invention, and the storage medium storing such a program product also constitutes the present invention. Obviously, the storage medium can be any known storage medium or any storage medium developed in the future. It should also be noted that in the apparatus and method of the present invention, it is obvious that the components or steps can be decomposed and / or recombined. These decompositions and / or recombinations should be considered equivalent to the present invention. Furthermore, the steps for performing the above series of processes can naturally be performed in the order described, but are not necessarily required to be performed in chronological order. Some steps can be performed in parallel or independently of each other.

[0084] The above are preferred embodiments of the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for controlling a cleaning drone for a desert photovoltaic power plant, characterized in that, The method comprises the following steps: acquiring environmental data and multi-dimensional state data of a photovoltaic module; inputting the multi-dimensional state data and the environmental data into a cleaning decision model for processing to obtain a cleaning trigger judgment result; generating a cleaning operation instruction according to the cleaning trigger judgment result and controlling a cleaning unmanned aerial vehicle to execute the cleaning operation instruction, wherein the cleaning operation instruction comprises a flight route instruction and a cleaning parameter instruction; acquiring real-time monitoring data collected by a plurality of monitoring devices on the cleaning unmanned aerial vehicle, wherein the real-time monitoring data comprises cleaning effect data and module thermal state data; dynamically optimizing the cleaning parameter instruction through the cleaning decision model according to the real-time monitoring data, generating an adjusted cleaning parameter instruction, and controlling the cleaning unmanned aerial vehicle to execute the adjusted cleaning parameter instruction.

2. The method of claim 1, wherein, The method of acquiring environmental data and multi-dimensional state data of a photovoltaic module comprises the following steps: acquiring environmental data through a meteorological sensor arranged in a photovoltaic power station, wherein the environmental data comprises real-time wind speed, environmental dust concentration and environmental temperature; acquiring multi-source image data of the photovoltaic module through an image acquisition device carried by the unmanned aerial vehicle; extracting and processing the multi-source image data to obtain position data of the photovoltaic module, first state data representing a pollution level of a surface of a photovoltaic panel, second state data representing an abnormal heating risk of the module, and third state data representing a surrounding ecological condition; The first state data is sand and dust coverage C d The calculation formula is: C d = A d / A t ×100%， wherein, A d is the area of the dust covered region identified from the visible light image by the image segmentation algorithm; A t is the standard reference total area of the photovoltaic panel.

3. The method of claim 2, wherein, The cleaning decision model is obtained through the following process: acquiring a historical photovoltaic power station operation and maintenance data set, wherein the historical photovoltaic power station operation and maintenance data set comprises environmental data, state data of a photovoltaic module, manual cleaning intervention record data, and performance improvement data collected in chronological order; integrating and processing the environmental data and the state data of the photovoltaic module to obtain an input feature vector; mapping and processing the manual cleaning intervention record data and the performance improvement data to obtain an optimal cleaning decision label; obtaining a training sample set according to the input feature vector and the optimal cleaning decision label, wherein the training sample set comprises a training set and a validation set; iteratively training an initialized cleaning decision model according to the training set to obtain a target weight coefficient; adjusting the initialized cleaning decision model according to the target weight coefficient to obtain a trained cleaning decision model; performing performance evaluation on the trained cleaning decision model according to the validation set to obtain an evaluation index, and when the evaluation index reaches a preset standard, a training completed cleaning decision model is obtained.

4. The method of claim 2, wherein the method further comprises: Inputting the multi-dimensional state data and the environmental data into a cleaning decision model for processing to obtain a cleaning trigger judgment result, comprising: The cleaning decision model processes the multi-dimensional state data and environment data through a target function to obtain a cleaning trigger comprehensive evaluation value Q , and a calculation formula of the target function is Q = α · f ( C d )+ β · g ( R h )- According to the cleaning trigger judgment result, a cleaning operation instruction is generated, comprising: · h ( v , C S , T )； wherein, α , β , when the comprehensive evaluation value of the cleaning trigger is greater than a preset trigger threshold, it is determined that the cleaning trigger condition is met, and a preset initial cleaning strategy parameter is output, wherein the initial cleaning strategy parameter comprises a first fan speed of a blowing fan and a first operation flight speed; is a weight coefficient determined by model training, f C d is a monotonically increasing function with respect to the sand cover degree, g R h is a mapping function with respect to the hot spot risk, h (v, Cs, T) is a penalty function representing the environmental operation suitability, R h is a hot spot risk identifier in the second state data, v is an environmental wind speed, C s is an environmental sand concentration, T is an environmental temperature.​​ 5. The method of claim 4, wherein, compiling and processing the initial cleaning strategy parameter according to a preset coding rule to obtain a cleaning parameter instruction; According to a pre-stored digital map of the photovoltaic power station, the position data of the photovoltaic module is path planned to obtain a generated flight route instruction. ​ ​ 6. The method of controlling a desert photovoltaic plant cleaning drone according to claim 1, wherein, Obtaining real-time monitoring data collected by a plurality of monitoring devices on the cleaning drone, including: Through the visible light camera carried by the drone, a visible light image sequence of the surface of the photovoltaic panel is collected; Through the infrared thermal imaging camera carried by the drone, a temperature distribution image sequence of the corresponding area of the photovoltaic panel is collected; According to the visible light image sequence, cleaning effect data reflecting the current cleaning effect is determined; According to the temperature distribution image sequence, component thermal state data is determined.

7. The method of controlling a desert photovoltaic plant cleaning drone according to claim 1, characterized in that, According to the real-time monitoring data, the cleaning parameter instruction is dynamically optimized through the cleaning decision model to generate an adjusted cleaning parameter instruction, including: The cleaning effect data is input into the cleaning decision model, and the cleaning decision model determines a cleaning parameter adjustment amount according to the cleaning effect data and a preset cleaning effect target value, and the calculation formula of the cleaning parameter adjustment amount is: ， wherein, is a gain coefficient, is a bias normalization coefficient, is a cleaning parameter adjustment amount, is a deviation of the cleaning effect data from a preset cleaning effect target value; The component thermal state data is input into the cleaning decision model, and the cleaning decision model generates an enhanced cleaning parameter according to the temperature exceeding standard level data and area data of the hot spot area in the component thermal state data, and the enhanced cleaning parameter includes a second flight speed, a second fan speed and a reciprocating cleaning frequency; The cleaning parameter adjustment amount and the enhanced cleaning parameter are compiled according to a preset coding to obtain the adjusted cleaning parameter instruction.

8. A control device for a desert photovoltaic power plant cleaning drone, characterized in that, Including: The acquisition module is used to acquire the environmental data and multi-dimensional state data of the photovoltaic module; The processing module is used to input the multi-dimensional state data and environmental data into the cleaning decision model for processing to obtain a cleaning trigger judgment result; according to the cleaning trigger judgment result, a cleaning operation instruction is generated and a cleaning drone is controlled to execute, and the cleaning operation instruction includes a flight route instruction and a cleaning parameter instruction; The acquisition module is also used to acquire real-time monitoring data collected by a plurality of monitoring devices on the cleaning drone, and the real-time monitoring data includes cleaning effect data and component thermal state data; The processing module is also used to dynamically optimize the cleaning parameter instruction through the cleaning decision model according to the real-time monitoring data to generate an adjusted cleaning parameter instruction and control the cleaning drone to execute.

9. A computing device, comprising: Including: The processor, the memory storing the computer program, when the computer program is run by the processor, executes the method of any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The instructions stored in the computer when the instructions are run on the computer, so that the computer executes the method of any one of claims 1 to 7.

Images