Photovoltaic panel cleaning scheme determination method considering environment

Abstract

The invention provides an environment-considered photovoltaic panel cleaning scheme determination method, and belongs to the technical field of photovoltaic panel cleaning.The method comprises the steps that firstly, real-time weather data and prediction data are obtained, and an environment stain correlation model is established; a multispectral camera is adopted to identify the types of stains on the surface of the photovoltaic panel and calculate a distribution map; calculating a cleaning index by combining the generating capacity data and the solar irradiation intensity, and determining a cleaning priority; a weather adaptability scoring system is constructed, and a robot operation environment threshold value is set; dividing power generation time periods according to the sunlight intensity and calculating an optimal cleaning time window; cleaning schemes are matched according to different stain types, and parameters are adjusted according to environmental conditions; calling a stain cleaning efficiency optimization function to carry out multi-objective optimization; and finally, generating an optimal cleaning path and parameter configuration by utilizing the photovoltaic cleaning planning network model. The technical problem that a photovoltaic panel cleaning scheme cannot be adaptively optimized according to complex and changeable environmental factors is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of photovoltaic panel cleaning, and in particular relates to a method for determining a photovoltaic panel cleaning solution taking the environment into consideration. Background Art

[0002] Photovoltaic power generation, an important form of clean energy, significantly impacts its efficiency due to the surface cleanliness of photovoltaic panels. Traditional photovoltaic panel cleaning technologies primarily rely on periodic manual cleaning, mechanical cleaning with fixed parameters, or simple automated cleaning systems. These methods typically trigger cleaning operations at preset intervals or based on a single indicator, such as a drop in power generation. These methods often employ standardized cleaning parameters and processes, such as fixed water volumes, detergent ratios, and cleaning paths, making them difficult to dynamically adjust to changing environmental conditions.

[0003] However, in real-world environments, the climate conditions in which photovoltaic power plants operate are complex and highly variable. Affected by seasonal variations, geographic location, and weather systems, environmental parameters such as precipitation, wind speed, temperature, humidity, UV intensity, and dust content vary significantly. Traditional cleaning technologies are unable to respond to these environmental changes in real time, resulting in cleaning operations being performed under unsuitable conditions (e.g., wasting water resources during strong winds, or impacting power generation efficiency during high-temperature periods) or failing to identify optimal cleaning times (e.g., timely cleaning after rain effectively utilizes natural resources).

[0004] The photovoltaic cleaning industry urgently needs an intelligent cleaning solution that comprehensively considers complex and changing environmental factors. This approach aims to achieve the most efficient cleaning path, using the most appropriate cleaning technology, and completing the photovoltaic panel cleaning process under optimal environmental conditions. The core issue that existing technologies struggle to address is their inability to adaptively optimize photovoltaic panel cleaning solutions based on these complex and changing environmental factors. Summary of the Invention

[0005] In view of this, the present invention provides a method for determining a photovoltaic panel cleaning solution taking the environment into consideration, which can solve the technical problem in the prior art that the photovoltaic panel cleaning solution cannot be adaptively optimized according to complex and changeable environmental factors.

[0006] The present invention is implemented as follows: The present invention provides a method for determining a photovoltaic panel cleaning plan that takes the environment into consideration, including: obtaining real-time weather data and forecast data of the photovoltaic panel monitoring area, forming an environmental parameter matrix and establishing an environmental stain association model; using a multispectral camera to scan the surface of the photovoltaic panel, identifying the surface stain type and calculating the stain coverage area ratio and thickness distribution map; calculating the photovoltaic panel cleaning index based on the real-time power generation data of the photovoltaic power station, and establishing a mapping relationship between the stain coverage rate and the power generation efficiency loss rate; constructing a weather adaptability scoring system and setting the robot operation threshold; calculating the optimal cleaning time window based on the sunshine intensity time period distribution curve; matching the corresponding cleaning plan for the identified stain type; calling the stain cleaning efficiency optimization function to optimize the cleaning plan; inputting the environmental parameter matrix, stain distribution data, historical cleaning records and photovoltaic panel layout information into the photovoltaic cleaning planning network model to generate the optimal cleaning path and parameter configuration.

[0007] Among them, the acquisition of real-time weather data and forecast data in the photovoltaic panel monitoring area includes: obtaining precipitation, wind speed, temperature, humidity, ultraviolet intensity and dust content, forming an environmental parameter matrix through environmental monitoring sensor networking, and establishing an environmental stain association model in combination with historical weather and stain data.

[0008] Among them, the use of a multispectral camera to scan the surface of the photovoltaic panel includes: identifying the type of surface stains and classifying them into four types: water-soluble stains, grease stains, solid particle stains and crystalline stains, and calculating the coverage area ratio and thickness distribution map of each stain.

[0009] The photovoltaic panel cleanliness index is a dimensionless parameter that characterizes the cleanliness of the photovoltaic panel and is calculated by comparing the theoretical power generation with the actual power generation. The value range is 0 to 1, and the closer the value is to 1, the cleaner the photovoltaic panel.

[0010] Among them, the weather adaptability scoring system includes a robot passage difficulty index, which is a parameter that quantifies the impact of different weather conditions on the movement and operation of the cleaning robot, taking into account the factors affecting mechanical performance such as ground wetness, wind resistance and temperature.

[0011] Among them, the stain cleaning efficiency optimization function is used to calculate the optimal cleaning strategy parameter combination while considering multiple environmental factors and stain characteristics. The input includes the stain type distribution matrix, the environmental parameter matrix, the cleaning resource constraint vector, the power generation period weight coefficient and the robot performance parameters. The output is the cleaning scheme matrix and the cleaning effect index including the cleaning scheme selection, detergent ratio, operating pressure, operating speed and expected cleaning effect.

[0012] Among them, the structure of the photovoltaic cleaning planning network model is a hybrid architecture based on graph neural network and multi-head adaptive attention mechanism, which includes four main components: environment encoding module, stain representation module, path planning module and parameter generation module. It adopts an encoder-decoder architecture. The encoder is responsible for processing the input environment and stain information, and the decoder generates the cleaning path sequence and the corresponding operation parameter configuration.

[0013] Among them, the steps of establishing the training data set of the photovoltaic cleaning planning network model include: collecting environmental data, stain accumulation pattern data, cleaning operation records and power generation efficiency change data of multiple photovoltaic power stations under different seasons and climatic conditions; using multispectral cameras and infrared imaging equipment to perform high-precision scanning of the surface of photovoltaic panels to construct a real annotation set of stain type distribution; and recording the actual effects and resource consumption of different cleaning strategies under various environmental conditions.

[0014] This method, by constructing an environmental parameter matrix and a weather adaptability scoring system, combined with a multi-objective optimization algorithm and a photovoltaic cleaning planning network model, enables intelligent response and adaptive optimization of cleaning plans to environmental changes. This method adaptively selects the optimal cleaning time window, cleaning method, detergent ratio, operating parameters, and cleaning path based on real-time environmental data and weather forecasts.

[0015] Compared with traditional fixed-mode photovoltaic panel cleaning technology, the present invention can accurately perceive and analyze complex and changeable environmental factors, avoiding one-size-fits-all environmentally inadaptable cleaning; through the weather adaptability scoring system, it effectively avoids the risk of cleaning under unsuitable meteorological conditions; based on a comprehensive analysis of environmental parameters and sunshine intensity, it grasps the best cleaning time and minimizes the impact of cleaning operations on power generation efficiency; using multi-objective optimization algorithms and photovoltaic cleaning planning network models, it achieves the best match between cleaning plans and environmental conditions, solving the technical problem that photovoltaic panel cleaning plans cannot be adaptively optimized according to complex and changeable environmental factors.

[0016] The photovoltaic cleaning planning network model training process includes: pre-training the model infrastructure in a simulated environment; introducing real photovoltaic power station data for fine-tuning training; using reinforcement learning methods to train the path planning module; training the stain characterization module through comparative learning; using transfer learning technology to adapt the pre-trained model to photovoltaic power stations in different geographical locations and climatic conditions; and performing end-to-end joint optimization training.

[0017] Among them, the high power generation period specifically refers to the daytime period with the highest solar radiation intensity and longest duration, usually from 10 am to 2 pm. During this period, the photovoltaic panels have the highest power generation efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0020] like Figure 1 FIG. 1 is a flow chart of a method for determining a photovoltaic panel cleaning solution taking the environment into consideration provided by the present invention. The method includes the following steps:

[0021] S01. Obtain real-time weather data and forecast data for the photovoltaic panel monitoring area, including precipitation, wind speed, temperature, humidity, UV intensity, and dust content. Form an environmental parameter matrix through a network of environmental monitoring sensors, and establish an environmental contamination correlation model based on historical weather and contamination data.

[0022] S02. Scan the surface of the photovoltaic panel using a multispectral camera to identify and classify surface stains into four types: water-soluble stains, grease stains, solid particle stains, and crystalline stains. Calculate the coverage area ratio and thickness distribution of each stain.

[0023] S03. Calculate the photovoltaic panel cleanliness index based on the real-time power generation data of the photovoltaic power station and the solar radiation intensity, establish a mapping relationship between the stain coverage rate and the power generation efficiency loss rate, and determine the cleaning priority matrix;

[0024] S04. Build a weather adaptability scoring system based on the robot's navigation difficulty index and task completion efficiency under different weather conditions. Set robot operation thresholds based on three key indicators: wind speed, precipitation, and temperature, and generate an operation risk assessment report.

[0025] S05. Divide the day into high power generation period, medium power generation period, and low power generation period according to the sunshine intensity distribution curve. Calculate the optimal cleaning time window based on the weight of the impact of stains on power generation in different periods.

[0026] S06. Based on the four identified stain types, the system matches the corresponding cleaning solutions, including pure water rinsing, additive rinsing, dry scrubbing, and combined cleaning methods. The cleaning solution ratio and spray pressure are adjusted according to the ambient temperature and humidity.

[0027] S07. Calling a stain cleaning efficiency optimization function to optimize the cleaning plan, taking the stain type distribution matrix, environmental parameter matrix, cleaning resource constraint vector, power generation period weight coefficient, and robot performance parameters as input parameters, and outputting an optimized cleaning plan matrix and an expected cleaning effect index;

[0028] S08. A photovoltaic cleaning planning network model is inputted with the environmental parameter matrix, stain distribution data, historical cleaning records, and photovoltaic panel layout information to generate an optimal cleaning path and parameter configuration to maximize resource utilization and minimize the impact on photovoltaic power generation.

[0029] Among them, the environmental parameter matrix specifically refers to a two-dimensional or three-dimensional data structure formed by arranging weather data collected by multiple environmental monitoring sensors in time and space dimensions, which is used to comprehensively describe the environmental status of the photovoltaic panels.

[0030] Among them, the cleanliness index specifically refers to a dimensionless parameter that characterizes the cleanliness of photovoltaic panels, which is calculated by comparing theoretical power generation with actual power generation. The value range is 0 to 1. The closer the value is to 1, the cleaner the photovoltaic panel.

[0031] Among them, the mapping relationship between the stain coverage rate and the power generation efficiency loss rate specifically refers to the established mathematical model, which describes the quantitative relationship between the impact of different types and degrees of stains on the power generation efficiency of photovoltaic panels, and is usually expressed as a nonlinear increasing function.

[0032] Among them, the robot passage difficulty index specifically refers to the parameters that quantify the impact of different weather conditions on the movement and operation of the cleaning robot, taking into account factors such as ground wetness, wind resistance and the impact of temperature on mechanical performance.

[0033] Among them, the high power generation period specifically refers to the daytime period with the highest solar radiation intensity and longest duration, usually from 10 am to 2 pm. During this period, the photovoltaic panels have the highest power generation efficiency.

[0034] Among them, the additive flushing method specifically refers to a cleaning method in which a surfactant or cleaning chemicals specified by the staff in a ratio set by the staff is added to the cleaning water to enhance the ability to remove greasy stains or crystalline stains.

[0035] Among them, the multi-objective optimization algorithm specifically refers to a mathematical solution method that simultaneously considers multiple mutually constrained decision-making objectives and finds the optimal balance point. This solution mainly balances the four dimensions of cleaning effect, resource consumption, environmental impact and power generation income.

[0036] Among them, the closed-loop feedback system specifically refers to an adaptive control system that dynamically adjusts cleaning parameters to achieve the best cleaning effect by real-time monitoring of changes in power generation efficiency and residual stains on the surface before and after cleaning during the cleaning process of the cleaning robot.

[0037] Among them, the stain cleaning efficiency optimization function is used to calculate the optimal cleaning strategy parameter combination while considering multiple environmental factors and stain characteristics. The input includes the stain type distribution matrix, environmental parameter matrix, cleaning resource constraint vector, power generation period weight coefficient and robot performance parameters. The output is the cleaning plan matrix and cleaning effect index including cleaning plan selection, detergent ratio, operating pressure, operating speed and expected cleaning effect.

[0038] Among them, the specific structure of the photovoltaic cleaning planning network model is a hybrid architecture based on graph neural network and multi-head adaptive attention mechanism, which includes four main components: environment encoding module, stain representation module, path planning module and parameter generation module. The number of attention heads of the multi-head adaptive attention mechanism is dynamically adjusted according to the scale and complexity of the photovoltaic array, and the attention weight distribution is determined according to the weather change gradient and stain distribution unevenness coefficient in the environmental parameter matrix, thereby realizing accurate modeling and decision-making of different environmental conditions and stain distribution conditions; the overall model adopts an encoder-decoder architecture, the encoder is responsible for processing the input environment and stain information, and the decoder generates the cleaning path sequence and the corresponding operation parameter configuration.

[0039] Among them, the steps for establishing the training data set in the photovoltaic cleaning planning network model training process specifically include collecting environmental data, stain accumulation pattern data, cleaning operation records and power generation efficiency change data of multiple photovoltaic power stations under different seasons and climatic conditions; using multispectral cameras and infrared imaging equipment to perform high-precision scanning of the photovoltaic panel surface to construct a real annotation set of stain type distribution; recording the actual effects and resource consumption of different cleaning strategies under various environmental conditions; establishing the correlation data between environmental parameters and stain accumulation rate through a combination of physical simulation and actual measurement; using data enhancement technology to simulate cleaning scenarios under extreme weather conditions to expand the coverage of the data set; and finally forming a comprehensive training set that includes environmental parameters, stain distribution, cleaning strategies and effect evaluation.

[0040] The training steps for the photovoltaic cleaning planning network model include: first, pre-training the model infrastructure in a simulated environment to enable it to identify the impact patterns of different weather conditions on the cleaning effect; then, introducing real photovoltaic power station data for fine-tuning training to optimize the model's adaptability to the actual environment; using reinforcement learning methods to train the path planning module to maximize cleaning efficiency and minimize resource consumption and interference with power generation; training the stain characterization module through comparative learning to enhance the recognition accuracy of different stain types and the accuracy of cleaning difficulty assessment; using transfer learning technology to adapt the pre-trained model to photovoltaic power stations in different geographical locations and climatic conditions; and finally, conducting end-to-end joint optimization training to enhance the synergy between the various modules of the model.

[0041] The specific implementation of the above steps is described in detail below.

[0042] The specific implementation of step S01 is to build a sensor network by deploying multiple environmental monitoring sensors. The sensor network consists of precipitation sensors, wind speed sensors, temperature sensors, humidity sensors, ultraviolet intensity sensors, and particulate matter concentration sensors. The sensors are distributed in a grid pattern with the photovoltaic array as the center, and the collection frequency is once every 10 minutes. The collected environmental data undergoes data preprocessing, including outlier detection and missing value interpolation, and a moving average filter algorithm is used to remove noise. The environmental parameter matrix is arranged according to the time and space dimensions to form M env A matrix is constructed, where rows represent time series, columns represent sensor nodes at different locations, and matrix elements are environmental parameter vectors corresponding to the time and space points. A Bayesian network is used to establish an environmental stain association model using historical weather data and stain data, and the probability distribution of various stain formations under different environmental conditions is calculated. The wind speed threshold is set at 10 m / s, the precipitation threshold is 5 mm / h, the temperature range is -10 to 50°C, the humidity range is 20 to 95%, the UV intensity range is 0 to 10, and the dust content range is 0 to 500 μg / m 3 This step aims to obtain comprehensive data on the environment in which the photovoltaic panels are located, providing a data basis for subsequent stain formation prediction and cleaning strategy formulation.

[0043] The specific implementation method of step S02 is to use a drone or robot equipped with a multispectral camera to perform high-resolution scanning on the surface of the photovoltaic panel. The multispectral camera includes visible light band, near-infrared band and short-wave infrared band, and the resolution is not less than 0.5mm / pixel. The scanned data undergoes image preprocessing, including geometric correction, radiation correction and enhancement processing. Stain recognition adopts a deep learning model, specifically an improved U-Net convolutional neural network structure, which divides stains into four types: water-soluble stains, greasy stains, solid particle stains and crystalline stains through semantic segmentation methods. Each type of stain is identified by different spectral features. For example, greasy stains have a characteristic absorption peak in the near-infrared band. The stain coverage area calculation is based on the pixel statistics method. The number of identified stain pixels is divided by the total number of pixels of the photovoltaic panel to obtain the coverage rate R cover The stain thickness is estimated by comparing the spectral reflectance with the thickness standard curve to generate a thickness distribution heat map H thick This step aims to accurately identify the stains on the surface of the photovoltaic panel and provide detailed stain distribution information for cleaning solution design.

[0044] The specific implementation of step S03 is to collect real-time power generation data of the photovoltaic power station, with a sampling frequency of once every 5 minutes, and simultaneously obtain solar radiation intensity data at the location of the photovoltaic panels. By comparing the theoretical power generation with the actual power generation, the photovoltaic panel cleanliness index CI is calculated. The calculation formula is the ratio of actual power generation to theoretical power generation. The theoretical power generation is calculated based on solar radiation intensity, photovoltaic panel conversion efficiency, and area. A mapping relationship model f(R) is established between the stain coverage rate and the power generation efficiency loss rate. cover ), using piecewise polynomial fitting method, establish the respective mapping functions f for different types of stains i (R cover，i Model parameters were derived through regression analysis of historical data. For example, for water-soluble stains, the power generation efficiency loss is approximately linear when the coverage is below 30%, but increases exponentially after coverage exceeds 30%. The cleaning priority matrix P is calculated based on the power generation efficiency loss rate, the cleaning difficulty coefficient of the stain type, and the regional importance weight. This step aims to quantify the impact of stains on power generation efficiency, determine cleaning priorities, and achieve efficient allocation of cleaning resources.

[0045] The specific implementation of step S04 involves constructing a weather adaptability scoring system to assess the feasibility and efficiency of the robot's cleaning tasks under different weather conditions. This system, based on fuzzy logic control theory, uses wind speed, precipitation, and temperature as input variables, and the robot's difficulty index and task completion efficiency as output variables. The wind speed input membership function is divided into low (0-3 m / s), medium (3-7 m / s), and high (above 7 m / s); the precipitation input membership function is divided into none (0 mm / h), low (0-2 mm / h), and high (above 2 mm / h); and the temperature input membership function is divided into low (below 5°C), moderate (5-35°C), and high (above 35°C). The robot's operating thresholds are set as follows: wind speed no more than 12 m / s, precipitation no more than 3 mm / h, and a temperature range of -5 to 45°C. Based on real-time weather data and forecast data, the robot's operating risk index (RI) is calculated for each time period within the next 24 hours, generating an operating risk assessment report. This step aims to assess the impact of weather on the robot's cleaning operations, avoid performing cleaning tasks under adverse weather conditions, and ensure the robot's safety and cleaning effectiveness.

[0046] The specific implementation of step S05 is to divide the 24 hours of a day into high power generation period (usually 10:00-14:00), medium power generation period (usually 7:00-10:00 and 14:00-17:00) and low power generation period (usually before sunrise and after 17:00) based on the solar radiation intensity time distribution curve. The time series analysis method is used to analyze the historical radiation intensity data and establish a typical sunshine pattern library. For different stain types, the influence weight W of each stain in different power generation period is calculated. impactFor example, the impact of solid particle stains is higher during high power generation periods. Combining the cleaning priority matrix P in step S03 and the operation risk index RI in step S04, a multi-objective optimization algorithm is used to calculate the optimal cleaning time window T opt Optimization objectives include minimizing the impact on power generation, maximizing cleaning efficiency, and minimizing operational risks. Optimization constraints include daylight conditions, robot operating hours, and weather conditions. This step aims to determine the optimal cleaning time, balancing the impact on power generation and cleaning efficiency, to maximize the overall benefits of the PV system.

[0047] The specific implementation method of step S06 is to match the corresponding cleaning scheme based on the four types of stains identified in step S02. Water-soluble stains are rinsed with pure water, using deionized water, the water temperature is set to 20-30°C, and the spraying pressure is 0.3-0.6 MPa. Grease stains are rinsed with an additive, adding anionic surfactants to deionized water at a concentration of 0.1-0.3%, and the spraying pressure is 0.5-0.8 MPa. Solid particle stains are rinsed with a combination of dry scrubbing or low-pressure water rinsing, with a scrubbing speed of 60-120 rpm and a contact pressure of 0.1-0.2 kPa. Crystalline stains are rinsed with an additive, adding a weak acid detergent to deionized water at a concentration of 0.2-0.5%, a spraying pressure of 0.6-1.0 MPa, and an action time of 60-120 seconds. The cleaning solution ratio and spray pressure are adjusted in real time based on ambient temperature and humidity. In low-temperature environments (below 5°C), the cleaning solution temperature is increased, while in high-humidity environments (above 80%), the water volume is reduced and wind-assisted drying is increased. This step aims to select the most appropriate cleaning method for different stain types, improving cleaning efficiency and effectiveness while reducing water consumption and wear on photovoltaic panels.

[0048] The specific implementation of step S07 is to call the stain cleaning efficiency optimization function to optimize the cleaning scheme in multiple dimensions. This function is implemented based on the genetic algorithm and the stain type distribution matrix M is used. dirt , environmental parameter matrix M env , Clean resource constraint vector V resource , power generation period weight coefficient W generation and robot performance parameters P robot As input. The stain type distribution matrix comes from the identification results of step S02, the environmental parameter matrix comes from the monitoring data of step S01, the cleaning resource constraint vector includes the available clean water, cleaning dosage and energy limit, the power generation period weight coefficient comes from the period division of step S05, and the robot performance parameters include maximum movement speed, battery life, water tank capacity and other parameters. The optimization process includes population initialization, fitness evaluation, selection, crossover and mutation operations. After multiple generations of evolution, the output is the optimized cleaning solution matrix M solutionand expected cleaning effect index E clean The cleaning plan matrix includes parameters such as the cleaning method selection, cleaning agent ratio, operating pressure, and operating speed for each area. The expected cleaning effect index is used to evaluate the expected results after the plan is implemented. This step aims to comprehensively consider multiple factors to generate the optimal cleaning strategy, balancing cleaning effect, resource consumption, environmental impact, and power generation benefits.

[0049] The specific implementation of step S08 is to transform the environmental parameter matrix M env , stain distribution data D dirt 、Historical cleaning record H clean and photovoltaic panel layout information L panel This data is input into the PV cleaning planning network model to generate the optimal cleaning path and parameter configuration. The PV cleaning planning network model is based on a hybrid architecture combining a graph neural network and a multi-head adaptive attention mechanism. It consists of four main components: an environment encoding module, a stain representation module, a path planning module, and a parameter generation module. The environment encoding module uses a bidirectional long short-term memory (BiLSTM) network to process time-series environmental data and capture environmental trends. The stain representation module uses a convolutional neural network to extract stain distribution characteristics and generate a stain density heat map. The path planning module generates cleaning paths based on an improved ant colony algorithm, optimizing the path length and maximizing cleaning coverage. The parameter generation module uses a fully connected neural network to generate cleaning parameters, including nozzle angle, water pressure, and scrubbing force, based on stain type and environmental conditions. The number of attention heads in the multi-head adaptive attention mechanism is dynamically adjusted based on the size and complexity of the PV array, and the attention weight distribution is determined based on the weather gradient and stain distribution non-uniformity coefficient in the environmental parameter matrix. This step aims to generate an efficient cleaning execution plan that maximizes resource utilization and minimizes the impact on PV power generation.

[0050] The detailed structure of the photovoltaic cleaning planning network model adopts an encoder-decoder architecture. The encoder part includes an environment encoding module and a stain representation module. The environment encoding module is composed of a bidirectional long short-term memory network (BiLSTM), and the input is the environment parameter matrix M env , the hidden layer dimension is 256, the number of layers is 3, and the output is the environment feature vector F env The stain representation module is composed of a residual convolutional neural network, which contains 5 residual blocks. Each residual block contains 2 3×3 convolutional layers and a jump connection. The input is the stain distribution data D dirt , the output is the stain feature map F dirtThe decoder part includes a path planning module and a parameter generation module. The path planning module is implemented based on the graph attention network (GAT), which represents the photovoltaic panel layout as a graph structure G = (V, E), where the node V represents the area to be cleaned and the edge E represents the feasible movement path. The node embedding is calculated through multiple rounds of message passing, and the improved ant colony algorithm is combined to generate the cleaning path sequence P. clean The parameter generation module adopts a multi-layer perceptron (MLP) structure, and the input is the environment feature vector F env , stain feature map F dirt and the current path location embedded in E pos The output is the clean parameter vector V param The multi-head adaptive attention mechanism establishes a connection between the encoder and the decoder, with the number of attention heads N head According to the formula N head =max(4,log2(N panel )) Determine, where N panel is the number of photovoltaic panels. Each attention head calculates the association weights between different environmental and stain characteristics, focusing on the importance of different regions and features. The entire model uses an end-to-end training approach, enabling each module to work together to optimize cleaning performance and resource utilization.

[0051] The specific implementation method for establishing a training dataset for the photovoltaic cleaning planning network model includes the following steps: first, collecting environmental data, stain accumulation pattern data, cleaning operation records, and power generation efficiency change data from multiple photovoltaic power stations under different seasonal and climatic conditions. The data collection period is no less than 12 months, covering the changes of the four seasons. Multispectral cameras and infrared imaging equipment are used to perform high-precision scans of the photovoltaic panel surfaces with a resolution of 0.3mm / pixel to construct a real-world annotated set of stain distributions, including four types of stains: water-soluble stains, grease stains, solid particle stains, and crystalline stains. The actual effects and resource consumption of different cleaning strategies under various environmental conditions are recorded, including changes in power generation efficiency before and after cleaning, cleaning duration, and water and detergent usage. Through a combination of physical simulation and field measurement, correlation data between environmental parameters and stain accumulation rates is established, with a simulation duration of no less than 100 typical weather cycles. Data augmentation technology is used to simulate cleaning scenarios under extreme weather conditions, including strong winds, heavy rain, high temperatures, and low temperatures, to expand the coverage of the dataset. Finally, a comprehensive training set is formed, which includes environmental parameters, stain distribution, cleaning strategies, and effect evaluation, with no less than 10,000 training samples. Photovoltaic Clean Planning Network

[0052] The specific implementation method for training the photovoltaic cleaning planning network model includes the following steps: first, pre-training the model infrastructure in a simulated environment, using a synthetic dataset, with 200 training rounds and a learning rate of 0.001; then introducing real photovoltaic power station data for fine-tuning training, with 100 training rounds and a learning rate of 0.0001; using the policy gradient reinforcement learning method to train the path planning module, and the reward function design includes cleaning coverage reward, path length penalty, resource consumption penalty and power generation interference penalty; training the stain representation module through comparative learning, using the InfoNCE loss function to enhance the recognition accuracy of different stain types and the accuracy of cleaning difficulty assessment; using transfer learning technology to adapt the pre-trained model to photovoltaic power stations with different geographical locations and climatic conditions, freezing the underlying feature extraction layer, and fine-tuning the high-level decision layer; finally, performing end-to-end joint optimization training, using a weighted multi-task loss function, and simultaneously optimizing the cleaning path planning loss, parameter generation loss and power generation impact loss, with 50 training rounds and a learning rate of 0.00005.

[0053] The mathematical model or calculation process involved in the present invention is described in detail below.

[0054] In step S01, the environmental parameter matrix M env The construction process is specifically shown as follows:

[0055] M env ={m i，j} n×k ={(p i，j , v i，j , t i，j , h i，j ,u i，j , d i，j )} n×k ;

[0056] Where M env is the environmental parameter matrix; m i，j is the environmental parameter vector collected by the jth sensor node at the i-th time point; n is the length of the time series; k is the number of sensor nodes; P i，j is the precipitation data, the unit is mm / h; v i，j is the wind speed data, in m / s; t i，j is the temperature data, the unit is ℃; h i，j Humidity data, unit is %; u i，j is the ultraviolet intensity data, dimensionless; d i，j Dust content data, unit is μg / m 3 .

[0057] The environmental stain association model is constructed using the Bayesian network, and its conditional probability distribution is expressed as follows:

[0058]

[0059] Where, P(D t |E t , D t-1 ) is the given current environmental condition E t and the stain state D at the previous moment t-1 The current stain state D t The conditional probability of i (D t |E t , D t-1 ) is the conditional probability distribution of the i-th type of stain; w i is the weight coefficient of each type of stain, and satisfies ε is a random error term, which obeys the normal distribution N(0, σ 2 ), the value range of σ is 0.01~0.05.

[0060] The parameter acquisition method is: environmental parameter data p i，j 、v i，j , t i，j 、h i，j 、u i，j d i，j The data is collected in real time by sensors, with a frequency of once every 10 minutes. The sensors are deployed using a grid method, around the photovoltaic array and in the center, with a spacing of no more than 50 meters. The sensor data is detected for outliers, using the triple standard deviation method to identify outliers, and linear interpolation to supplement missing values. i (D t |E t , D t-1 ) is obtained through statistical analysis of historical data, and at least one year of historical data is required as a training sample. Weight coefficient w i The maximum likelihood estimation method was used to determine the model accuracy, and cross-validation was used to evaluate the model accuracy.

[0061] In step S02, the stain coverage P cover The calculation formula is:

[0062]

[0063] Where R cover，i is the coverage of the i-th type of stain, ranging from 0 to 100%; A dirt，i is the pixel area covered by the i-th type of stain; A total is the total pixel area of the photovoltaic panel.

[0064] The stain thickness estimation model is specifically expressed as:

[0065]

[0066] Where H thick，i (x, y) is the estimated thickness of the i-th type of stain at the coordinate (x, y), in mm; R spec，i (x, y) is the spectral reflectance of the i-th type of stain at coordinate (x, y); f i is the thickness estimation function of the i-th type of stain; a i 、b i 、c i is the fitting coefficient; ε i is the error term, which obeys the normal distribution σ i The value range is 0.01~0.1.

[0067] The parameter acquisition method is: stain coverage area A dirt，i Obtained through multispectral image processing, the improved U-Net convolutional neural network is used for semantic segmentation, with a segmentation accuracy of no less than 95%. Spectral reflectance R spec，i (x, y) is obtained by measuring the reflection intensity in different bands using a multispectral camera. The measurement bands include visible light (400-700nm), near infrared (700-1100nm) and short-wave infrared (1100-2500nm). Fitting coefficient a i 、b i 、c i It is obtained through laboratory sample preparation and measurement. First, standard stain samples of different thicknesses are prepared, and the actual thickness is measured using a thickness gauge with micron-level precision. Then, the corresponding spectral reflectance is measured using a multispectral camera. Finally, the parameter value is obtained by least squares fitting.

[0068] In step S03, the calculation formula of the photovoltaic panel cleanliness index CI is:

[0069]

[0070] Where, CI is the cleanliness index of photovoltaic panels, dimensionless, ranging from 0 to 1; P actual is the actual power generation, in kW; P theoretical is the theoretical power generation, in kW; I is the solar radiation intensity, in kW / m 2 ; A is the area of the photovoltaic panel, in m 2 ; η is the photovoltaic panel conversion efficiency, dimensionless; L system is the system inherent loss, dimensionless.

[0071] The mapping relationship model between the stain coverage rate and the power generation efficiency loss rate f(R cover ) is specifically expressed as:

[0072]

[0073] Where, L eff is the power generation efficiency loss rate, ranging from 0 to 1; f(R cover ) is the overall mapping function; f i (R cover，i ) is the mapping function of the i-th type of stain; α i is the influence weight of the i-th type of stain, and satisfies R cover，i is the coverage of the i-th type of stain.

[0074] Mapping function f for various types of stains i (R cover，i ) using a piecewise polynomial function:

[0075]

[0076] Where k 1i 、k 2i 、k 3i is the fitting coefficient; a i is the power coefficient, usually greater than 1; R threshold，i is the threshold coverage, water-soluble stain R threshold，1 30%, greasy stains R threshold，2 20%, solid particle stains R threshold，3 25%, crystalline stains R threshold，4 is 15%.

[0077] The calculation formula of the cleaning priority matrix P is:

[0078] P x，y =w1·L eff (x, u) + w2·D(x, y) + w3·I area (x, y) + ε;

[0079] Where, P x，y is the cleaning priority at coordinate (x, y); L eff (x, y) is the power generation efficiency loss rate at the coordinate (x, y); D(x, y) is the difficulty coefficient of stain cleaning at the coordinate (x, y); I area (x, y) is the importance weight of the region where the coordinate (x, y) is located; w1, w2, w3 are weight coefficients, and they satisfy w1+w2+w3=1; ε is the random error term, which obeys the normal distribution N(0, σ 2 ), the value range of σ is 0.01~0.05.

[0080] The parameter acquisition method is: actual power generation P actualThe data is obtained through the photovoltaic power station monitoring system with a sampling frequency of every 5 minutes. The solar radiation intensity I is measured in real time by a irradiance meter. The photovoltaic panel area A, conversion efficiency η and system inherent loss L system is the design parameter of the photovoltaic system. The pollution influence weight α i The test was obtained through experimental testing. The test process was to apply different types of stains with different coverage rates to the surface of the photovoltaic panel under controlled conditions, measure the changes in power generation efficiency, and obtain the weight coefficient through regression analysis. Fitting coefficient k 1k 、i 2i 、k 3i and power coefficient a i The stain cleaning difficulty coefficient D(x, y) is also obtained through experimental testing and regression analysis. The stain cleaning difficulty coefficient D(x, y) is determined based on the stain type and thickness. The cleaning difficulty benchmark values of different stain types are obtained through experimental testing and then calculated based on the thickness distribution. area (x, y) is determined based on the location of the photovoltaic panels and their connection to the series connection. Panels at the edge have lower weights, while panels at the center and key connection locations have higher weights. Weight coefficients w1, w2, and w3 are determined using a multi-objective optimization algorithm, with the goal of maximizing the ratio of post-cleaning power generation efficiency gain to cleaning costs.

[0081] In step S04, the calculation formula of the robot operation risk index RI is based on fuzzy logic control theory:

[0082]

[0083] Where RI is the robot operation risk index, ranging from 0 to 1; μ j (v, p, t) is the membership degree of the jth fuzzy rule; r j is the risk value corresponding to the jth rule; m is the total number of fuzzy rules; v is the wind speed; p is the precipitation; t is the temperature; ε is the error term, which obeys the normal distribution N(0, σ 2 ), the value range of σ is 0.01~0.05.

[0084] The membership functions of wind speed, precipitation, and temperature are:

[0085]

[0086]

[0087] The fuzzy rules are designed based on the impact of wind speed, precipitation, and temperature on robot operations. For example, "If the wind speed is high and the precipitation is heavy, the operation risk is high." The total number of fuzzy rules, m, is determined by the number of membership functions for wind speed, precipitation, and temperature: m = 3 × 3 × 3 = 27. The risk value r corresponding to each rule is jDetermined based on expert experience and historical data, ranging from 0 to 1.

[0088] The parameter acquisition method is: wind speed v, precipitation p and temperature t are acquired in real time through environmental monitoring sensors. The risk value r of the fuzzy rule j Based on historical data statistics and expert experience, data such as the success rate, completion time, and failure rate of robot operations under different weather conditions must be collected to establish a mapping between weather conditions and operational risks. The parameters of the membership function are determined through historical data analysis to ensure that the function shape aligns with actual risk trends. When building a fuzzy rule base, the membership function for each input variable is first defined. Then, rules are formulated based on expert experience and historical data. Finally, the rule base is optimized through repeated testing and adjustment.

[0089] In step S05, the optimal cleaning time window T opt The calculation formula is:

[0090] T opt =argmin T [w1·∫ T L power (t)dt+w2·∫ T E clean (t)dt+w3·∫ T RI(t)dt]+ε;

[0091] Where, T opt is the optimal cleaning time window; T is the candidate time window; L power (t) is the power generation loss function at time t; E clean (t) is the cleaning efficiency function at time t; RI(t) is the operation risk index at time t; w1, w2, w3 are weight coefficients, and they satisfy w1+w2+w3=1; ε is the random error term, which obeys the normal distribution N(0, σ 2 ), the value range of σ is 0.01~0.05.

[0092] Power generation loss function L power The calculation formula for (t) is:

[0093] L power (t) = W impact (t)·P theoretical (t)·L eff ;

[0094] Where, L power (t) is the power generation loss at time t; W impact (t) is the influence weight of time t; P theoretical (t) is the theoretical power generation at time t; L eff is the power generation efficiency loss rate.

[0095] Influence weight W impact (t) Determined by time period:

[0096]

[0097] Where, T high It is the high power generation period (10:00~14:00); T mid The middle power generation period (7:00~10:00 and 14:00~17:00); T low Low power generation period (before sunrise and after 17:00); high 、w mid 、w low is the weight coefficient of each period, usually satisfying w high >w mid >w low , the typical value is w high =1.0,w mid =0.6, w low =0.2.

[0098] Cleaning efficiency function E clean The calculation formula for (t) is:

[0099] E clean (t) = E base ·f temp (t)·f light (t)·f humid (t)+ε E ;

[0100] Where, E clean (t) is the cleaning efficiency at time t; E base is the benchmark cleaning efficiency; f temp (t) is the temperature influence factor; f light (t) is the light impact factor; f humid (t) is the humidity influence factor; ε E is the error term, which obeys the normal distribution σ E The value range is 0.05~0.1.

[0101] The parameter acquisition method is: Theoretical power generation P theoretical (t) Calculated based on the solar radiation intensity forecast data. Temperature influence factor f temp (t), light impact factor f light (t) and humidity influence factor f humid(t) Obtained through experimental testing, measuring the changes in cleaning efficiency under different temperature, light and humidity conditions, and establishing a relationship model between environmental factors and cleaning efficiency. Baseline cleaning efficiency E base Determined based on robot design parameters and historical operation data. Weight coefficient w for each time period high 、w mid 、w low The ratio of power generation in each period to the total power generation for the day is determined by analyzing historical power generation data and used as a weight reference value.

[0102] In step S07, the stain cleaning efficiency optimization function is expressed as:

[0103] {M solution , E clean}=argmax X F(X,M dirt , M env , V resource , W generation , P robot )+ε;

[0104] Where M solution is the optimized cleaning solution matrix; E clean is the expected cleaning effect index; X is the decision variable matrix, including parameters such as cleaning method selection, detergent ratio, operating pressure, and operating speed; F is the objective function; M dirt is the stain type distribution matrix; M env is the environmental parameter matrix; V resource is the clean resource constraint vector; W generation is the power generation period weight coefficient; P robot is the robot performance parameter; ε is the random error term, which obeys the normal distribution N(0, σ 2 ), the value range of σ is 0.01~0.05.

[0105] The specific expression of the objective function F is:

[0106]

[0107] Where, E eff C is the cleaning efficiency index; resource is the resource consumption indicator; I env G is the environmental impact indicator; power is the power generation income index; w1, w2, w3, w4 are weight coefficients, and they satisfy w1+w2+w3+w4=1; ε F is the error term, which obeys the normal distribution σ F The value range is 0.05~0.1.

[0108] Cleaning efficiency index E eff The calculation formula is:

[0109]

[0110] Where, E eff It is the cleaning efficiency index, ranging from 0 to 1; A i，j is the area of the jth type of stain in the ith region; R i，j is the removal rate of the jth type of stain in the ith area; e i，j is the cleaning efficiency coefficient of the jth type of stain in the ith area; n is the total number of areas.

[0111] Resource consumption indicator C resource The calculation formula is:

[0112]

[0113] Where C resource is the resource consumption indicator; A i，j is the area of the jth type of stain in the ith region; V water，i，j is the water consumption per unit area of the jth type of stain in the ith area; V agent，i，j E is the unit area cleaning agent consumption for the jth type of stain in the i-th area; energy，i，j is the energy consumption per unit area of the jth type of stain in the ith area; w water 、w agent 、w energy is the weight coefficient of each resource, and satisfies w water +w agent +w energy =1.

[0114] The parameter acquisition method is: stain type distribution matrix M dirt From the recognition result of step S02, the matrix elements represent the distribution of the type of stains at the corresponding position. env The matrix elements of the monitoring data from step S01 represent the environmental parameter vectors at the corresponding time and space points. resource Determined based on the design parameters of the robot cleaning system, including parameters such as maximum water tank capacity, detergent inventory, and battery capacity. Power generation period weight coefficient W generation The time period division from step S05, the weight coefficient of different time periods reflects the proportion of the power generation of the time period to the total power generation of the whole day. robot Determined according to the robot design specifications, including parameters such as maximum moving speed, battery life, water tank capacity, etc. Cleaning plan matrix M solution and expected cleaning effect index E cleanThe solution is obtained through genetic algorithm, which includes initial population generation, fitness evaluation, selection, crossover, mutation and elite retention strategy. After multiple generations of evolution, the optimal solution is output. i，j and cleaning efficiency coefficient e i，j The results were obtained through experimental testing, and the stain removal performance was measured under different cleaning parameters, and a relationship model between cleaning parameters and cleaning effect was established. water，i，j 、V agent，i，j and E energy，i，j The same is obtained through experimental testing.

[0115] The construction principles and meanings of the above equations are explained as follows:

[0116] Environmental parameter matrix M env The system uses a matrix representation, using time and space as its two dimensions. Each matrix element represents an environmental parameter vector, encompassing six parameters: precipitation, wind speed, temperature, humidity, UV intensity, and dust content. This representation method comprehensively captures the spatiotemporal variations in the environment, providing a data foundation for subsequent stain prediction and cleaning strategy development.

[0117] The environmental stain association model uses a Bayesian network. Its conditional probability distribution expression is based on probability theory and statistical principles, and takes into account the impact of the current environmental conditions and the stain state at the previous moment on the current stain state. The weight coefficient w is introduced into the model. i , reflecting the importance of different types of stains; an error term, ε, is introduced to account for the randomness and uncertainty of the model's predictions. This model can accurately predict the probability of various types of stains forming under different environmental conditions, providing a theoretical basis for the formulation of cleaning strategies.

[0118] Stain coverage R cover The calculation formula is based on the area ratio principle. The percentage of stain coverage is calculated by dividing the pixel area covered by the stain by the total pixel area of the photovoltaic panel. This calculation method is intuitive and easy to implement, and can accurately quantify the degree of stain coverage.

[0119] The stain thickness estimation model uses a quadratic polynomial function to establish the relationship between spectral reflectance and stain thickness. The quadratic term is introduced in the model because the reflectance and thickness usually have a nonlinear relationship. The rate of change of reflectance slows down when the thickness increases. The error term ε is introduced i , taking into account the random errors in the measurement and fitting process. The model can estimate the stain thickness based on the spectral reflectance data and generate a thickness distribution heat map, which provides an important basis for the optimization of cleaning parameters.

[0120] The calculation formula for the PV panel cleanliness index (CI) is based on the principle of energy conversion, using the ratio of actual power generation to theoretical power generation as the cleanliness index. Theoretical power generation takes into account factors such as solar radiation intensity, panel area, conversion efficiency, and inherent system losses. This index provides a direct reflection of the cleanliness of PV panels, providing a quantitative basis for cleaning decisions.

[0121] The mapping model between the stain coverage and the power generation efficiency loss rate is represented by a piecewise polynomial function, with different parameters for different types of stains. When the stain coverage is below the threshold, a linear relationship is adopted; when the coverage is above the threshold, a power term is introduced, resulting in a nonlinear growth relationship. This piecewise representation method is consistent with actual observations, namely that at low coverage, the efficiency loss and coverage are approximately linearly related; at high coverage, the efficiency loss growth rate accelerates, showing an exponential growth trend. The stain type weight α is introduced into the model. i , reflecting the different degrees of influence of different types of stains on power generation efficiency.

[0122] The calculation formula for the cleaning priority matrix P comprehensively considers three factors: power generation efficiency loss rate, stain cleaning difficulty, and area importance. The cleaning priority of each location is calculated through a weighted summation. Weight coefficients w1, w2, and w3 are introduced to adjust the importance of each factor based on actual needs. An error term ε is introduced to account for random errors in the calculation process. This matrix provides priority guidance for cleaning route planning, ensuring that limited cleaning resources are allocated to the areas most in need.

[0123] The calculation formula for the Robot Operation Risk Index (RI) is based on fuzzy logic control theory. It uses wind speed, precipitation, and temperature as input variables and applies fuzzy rules to calculate the operation risk index. Fuzzy logic effectively handles the uncertainty and ambiguity of these input variables, making it more consistent with practical decision-making processes. The membership functions for wind speed, precipitation, and temperature use piecewise linear functions, which are simple, intuitive, and computationally efficient. This index quantifies the operational risk under different weather conditions, providing a basis for decision-making regarding cleaning task scheduling.

[0124] Optimal cleaning time window T opt The calculation formula adopts the optimization method, and the objective function comprehensively considers the three factors of power generation loss, cleaning efficiency and operation risk. power (t) takes into account the impact weight of the time period, theoretical power generation and efficiency loss rate, reflecting the different degrees of impact of cleaning activities on power generation in different time periods. Cleaning efficiency function E clean (t) considers the impact of environmental factors such as temperature, light, and humidity on cleaning efficiency, combining the effects of each factor multiplicatively. This optimization method can find the optimal cleaning time window that balances multiple factors and maximizes cleaning efficiency.

[0125] The stain cleaning efficiency optimization function adopts a multi-objective optimization method. The objective function F comprehensively considers four dimensions: cleaning efficiency, resource consumption, environmental impact, and power generation income. Cleaning efficiency index E eff The weighted cleaning efficiency of each area and each type of stain was calculated; the resource consumption index C resource The weighted consumption of water, cleaning agents, and energy was calculated. The objective function uses an inverse formula for resource consumption and environmental impact, indicating that lower values are preferable. Meanwhile, a positive formula is used for cleaning efficiency and power generation revenue, indicating that higher values are preferable. This function can find the optimal balance between multiple mutually constrained objectives, generating a cleaning solution with the best overall benefits.

[0126] Optionally, in step S06, the cleaning parameter adjustment formula is:

[0127] P spray (T, H) = P base ·f T (T)·f H (H)+ε P ;

[0128] Where, P spray (T, H) is the adjusted spraying pressure, in MPa; P base is the reference spraying pressure, in MPa; f T (T) is the temperature adjustment factor; f H (H) is the humidity adjustment factor; ε P is the error term, which obeys the normal distribution σ P The value range is 0.01~0.05.

[0129] Temperature adjustment factor f T (T) and humidity adjustment factor f H The calculation formula for (H) is:

[0130]

[0131] Where, T ref is the reference temperature, usually 20℃; H ref is the reference humidity, usually 60%; k T is the temperature adjustment coefficient, the typical value is 0.01~0.03 / ℃; k H is the humidity adjustment coefficient, and its typical value is 0.005~0.01 / %.

[0132] Optionally, the detergent ratio adjustment formula is:

[0133] C agent (T)=Cbase ·g T (T)+ε C ;

[0134] Where C agent (T) is the adjusted detergent concentration, in %; C base is the base detergent concentration, in %; g T (T) is the temperature ratio adjustment factor; ε C is the error term, which obeys the normal distribution σ C The value range is 0.01~0.05.

[0135] Temperature ratio adjustment factor g T The calculation formula for (T) is:

[0136] g T (T) = 1 + k C ·(T ref -T);

[0137] Where, T ref is the reference temperature, usually 20°C; k C is the ratio adjustment coefficient, and its typical value is 0.005~0.02 / ℃.

[0138] The parameter acquisition method is: reference spraying pressure P base and the baseline detergent concentration C base Determined according to the type of stain, the reference values for different types of stains are given in step S06. Temperature T and humidity H are obtained in real time through environmental monitoring sensors. Temperature adjustment coefficient k T , humidity adjustment coefficient k H and ratio adjustment coefficient k C The cleaning effect was tested under different temperature and humidity conditions, and the optimal parameter values were determined through regression analysis.

[0139] The cleaning parameter adjustment formula is constructed based on the principle of considering the impact of ambient temperature and humidity on cleaning performance. In low-temperature environments, the spray pressure and detergent concentration are increased to improve cleaning performance; in high-humidity environments, the spray pressure is reduced to avoid excessive moisture and prolonged drying time. Using a piecewise function form ensures that the adjustment process is more tailored to actual needs, avoiding over-adjustment that wastes resources or reduces cleaning performance.

[0140] Specifically, the core technology of this invention is to build a complete closed-loop system for environmental perception, environmental adaptability assessment, and solution optimization. By combining data-driven and model-based prediction, the photovoltaic cleaning solution can intelligently respond to environmental changes. Its working principle can be divided into the following key aspects:

[0141] First, the present invention establishes a correlation mechanism between environment, contamination, and power generation efficiency. A network of environmental monitoring sensors collects real-time and forecasted weather data, including precipitation, wind speed, temperature, humidity, UV intensity, and dust content, to form an environmental parameter matrix. This data is combined with historical weather and contamination data to establish an environmental contamination correlation model, quantifying the impact of different environmental conditions on contamination formation. Furthermore, the present invention combines real-time power generation data from photovoltaic power plants and solar radiation intensity to calculate the photovoltaic panel cleanliness index, establishing a mapping between contamination coverage and power generation efficiency loss. This correlation mechanism provides a data foundation for environmentally adaptive cleaning decisions.

[0142] Secondly, the present invention introduces a decision-making mechanism for cleaning timing and methods based on environmental conditions. Through a weather adaptability scoring system, robot operation thresholds are set based on three key indicators: wind speed, precipitation, and temperature. Operation risk assessment reports are generated to avoid high-risk cleaning operations in harsh environmental conditions. Based on the sunlight intensity distribution curve, high, medium, and low power generation periods are divided and the optimal cleaning time window is calculated to minimize the impact of the cleaning process on power generation. According to the ambient temperature and humidity conditions, the cleaning liquid ratio and spray pressure are dynamically adjusted to ensure cleaning results while reducing resource consumption.

[0143] Third, the present invention constructs an environmental adaptability optimization system based on a photovoltaic cleaning planning network model. This model uses a hybrid architecture of a graph neural network and a multi-head adaptive attention mechanism, comprising an environmental encoding module, a stain characterization module, a path planning module, and a parameter generation module. Through deep learning and analysis of the environmental parameter matrix, historical cleaning records, and photovoltaic panel layout information, it generates a cleaning path and parameter configuration that best suits the current environmental conditions. In particular, the multi-head adaptive attention mechanism in the model can determine attention weights based on the environmental change gradient, enabling accurate modeling and decision-making for different environmental conditions.

[0144] Finally, the present invention achieves multi-objective environmental adaptation optimization of cleaning solutions. Using a stain cleaning efficiency optimization function, the system takes the environmental parameter matrix, the cleaning resource constraint vector, the power generation period weight coefficient, and the robot performance parameters as inputs. It comprehensively considers multiple mutually constrained decision-making objectives, including environmental adaptability, cleaning effectiveness, resource consumption, and power generation revenue, and outputs the optimized cleaning solution matrix and expected cleaning effectiveness index.

[0145] It is through these organically combined technical principles that the present invention can optimize the environmental adaptability of photovoltaic panel cleaning solutions, solving the technical problem in traditional technologies that are unable to adaptively optimize photovoltaic panel cleaning solutions based on complex and changeable environmental factors.

[0146] A specific embodiment 1 of the present invention is provided below. The specific implementation of each step in this embodiment 1 is described in detail as follows.

[0147] The specific implementation of step S01 is to build a sensor network by deploying multiple environmental monitoring sensors. The sensor network consists of precipitation sensors, wind speed sensors, temperature sensors, humidity sensors, ultraviolet intensity sensors, and particulate matter concentration sensors. The sensors are distributed in a grid pattern with the photovoltaic array as the center, and the collection frequency is once every 10 minutes. The collected environmental data undergoes data preprocessing, including outlier detection and missing value interpolation, and a moving average filter algorithm is used to remove noise. The environmental parameter matrix is arranged according to the time and space dimensions to form M env Matrix, where rows represent time series, columns represent sensor nodes at different locations, and matrix elements are environmental parameter vectors corresponding to time and space points. Environmental parameter matrix M env The construction process of M is specifically expressed as follows: env ={m i，j} n×k ={(p i，j , v i，j , t i，j , h i，j ,u i，j , d i，j )} n×k Where, M env is the environmental parameter matrix; m i，j is the environmental parameter vector collected by the jth sensor node at the i-th time point; n is the length of the time series; k is the number of sensor nodes; p i，j is the precipitation data, the unit is mm / h; v i，j is the wind speed data, in m / s; t i，j is the temperature data, the unit is ℃; h i，j Humidity data, unit is %; u i，j is the ultraviolet intensity data, dimensionless; d i，j Dust content data, unit is μg / m 3 The historical weather data and the stain data are used to establish an environmental stain association model through the Bayesian network, and the probability distribution of various types of stains under different environmental conditions is calculated. The environmental stain association model is constructed using the Bayesian network, and its conditional probability distribution is expressed as: Where, P(D t |E t , D t-1 ) is the given current environmental condition E t and the stain state D at the previous moment t-1 The current stain state D t The conditional probability of i (D t |E t , D t-1 ) is the conditional probability distribution of the i-th type of stain; w iis the weight coefficient of each type of stain, and satisfies ε is a random error term, which obeys the normal distribution N(0, σ 2 ), the value range of σ is 0.01 to 0.05. The wind speed threshold is set to 10 m / s, the precipitation threshold is 5 mm / h, the temperature range is -10 to 50 °C, the humidity range is 20 to 95%, the UV intensity range is 0 to 10, and the dust content range is 0 to 500 μg / m 3 This step aims to obtain comprehensive data on the environment in which the photovoltaic panels are located, providing a data basis for subsequent stain formation prediction and cleaning strategy formulation.

[0148] The specific implementation method of step S02 is to use a drone or robot equipped with a multispectral camera to perform high-resolution scanning on the surface of the photovoltaic panel. The multispectral camera includes visible light band, near-infrared band and short-wave infrared band, and the resolution is not less than 0.5mm / pixel. The scanned data undergoes image preprocessing, including geometric correction, radiation correction and enhancement processing. Stain recognition adopts a deep learning model, specifically an improved U-Net convolutional neural network structure, which divides stains into four types: water-soluble stains, greasy stains, solid particle stains and crystalline stains through semantic segmentation methods. Each type of stain is identified by different spectral features. For example, greasy stains have a characteristic absorption peak in the near-infrared band. The stain coverage area calculation is based on the pixel statistics method. The number of identified stain pixels is divided by the total number of pixels of the photovoltaic panel to obtain the coverage rate R cover . Stain coverage R cover The calculation formula is: Where R cover，i is the coverage of the i-th type of stain, ranging from 0 to 100%; A dirt，i is the pixel area covered by the i-th type of stain; A total is the total pixel area of the photovoltaic panel. The stain thickness is estimated by comparing the spectral reflectance with the thickness standard curve to generate the thickness distribution heat map H thick The stain thickness estimation model is specifically expressed as: Where H thick，i (x, y) is the estimated thickness of the i-th type of stain at the coordinate (x, y), in mm; R spec，i (x, y) is the spectral reflectance of the i-th type of stain at coordinate (x, y); f i is the thickness estimation function of the i-th type of stain; a i 、b i 、c i is the fitting coefficient; ε i is the error term, which obeys the normal distribution σ iThe value range of is 0.01 to 0.1. This step aims to accurately identify the stains on the surface of the photovoltaic panel and provide detailed stain distribution information for cleaning solution design.

[0149] The specific implementation of step S03 is to collect real-time power generation data from the photovoltaic power station at a sampling frequency of once every 5 minutes, and simultaneously obtain solar radiation intensity data at the location of the photovoltaic panels. By comparing the theoretical power generation with the actual power generation, the photovoltaic panel cleanliness index CI is calculated. The calculation formula is the ratio of actual power generation to theoretical power generation. The calculation formula for the photovoltaic panel cleanliness index CI is: Where, CI is the cleanliness index of photovoltaic panels, dimensionless, ranging from 0 to 1; P actual is the actual power generation, in kW; P theoretical is the theoretical power generation, in kW; I is the solar radiation intensity, in kW / m 2 ; A is the area of the photovoltaic panel, in m 2 ; η is the photovoltaic panel conversion efficiency, dimensionless; L system is the inherent loss of the system, dimensionless. The theoretical power generation is calculated based on the solar radiation intensity, photovoltaic panel conversion efficiency and area. A mapping relationship model between the stain coverage rate and the power generation efficiency loss rate is established. cover ), using piecewise polynomial fitting method, establish the respective mapping functions f for different types of stains i (R cover，i ). The mapping relationship model between the stain coverage rate and the power generation efficiency loss rate f(R cover ) is specifically expressed as: f i (R cover，i );where L eff is the power generation efficiency loss rate, ranging from 0 to 1; f(R cover ) is the overall mapping function; f i (R cover，i ) is the mapping function of the i-th type of stain; α i is the influence weight of the i-th type of stain, and satisfies R cover，i is the coverage of the i-th type of stain. The mapping function f of each type of stain i (R cover，i ) using a piecewise polynomial function: Where k 1i 、k 2i 、k 3i is the fitting coefficient; a i is the power coefficient, usually greater than 1; R threshold，i is the threshold coverage, water-soluble stain R threshold，1 30%, greasy stains R threshold，2 20%, solid particle stains R threshold，325%, crystalline stains R threshold，4 The model parameters are obtained through historical data regression analysis. For example, when the coverage of water-soluble stains is less than 30%, the power generation efficiency loss is approximately linear, and when the coverage exceeds 30%, it increases exponentially. The cleaning priority matrix P is calculated based on the power generation efficiency loss rate, the cleaning difficulty coefficient of the stain type, and the regional importance weight. The calculation formula of the cleaning priority matrix P is: P x，y =w1·L eff (x, y) + w2·D(x, y) + w3·L area (x, y) + ε; where P x，y is the cleaning priority at coordinate (x, y); L eff (x, y) is the power generation efficiency loss rate at the coordinate (x, y); D(x, y) is the difficulty coefficient of stain cleaning at the coordinate (x, y); I area (x, y) is the importance weight of the region where the coordinate (x, y) is located; w1, w2, w3 are weight coefficients, and they satisfy w1+w2+w3=1; ε is the random error term, which obeys the normal distribution N(0, σ 2 ), with σ ranging from 0.01 to 0.05. This step aims to quantify the impact of contamination on power generation efficiency, determine cleaning priorities, and achieve efficient allocation of cleaning resources.

[0150] The specific implementation method of step S04 is to construct a weather adaptability scoring system to evaluate the feasibility and efficiency of the robot in performing cleaning tasks under different weather conditions. The system is based on fuzzy logic control theory, with wind speed, precipitation and temperature as input variables, and the robot's passage difficulty index and task completion efficiency as output variables. The wind speed input membership function is divided into low (0-3m / s), medium (3-7m / s) and high (above 7m / s); the precipitation input membership function is divided into none (0mm / h), small (0-2mm / h) and large (above 2mm / h); the temperature input membership function is divided into low (below 5°C), moderate (5-35°C) and high (above 35°C). The calculation formula of the robot operation risk index RI is based on fuzzy logic control theory:

[0151] Where RI is the robot operation risk index, ranging from 0 to 1; μ j (v, p, t) is the membership degree of the jth fuzzy rule; r j is the risk value corresponding to the jth rule; m is the total number of fuzzy rules; v is the wind speed; p is the precipitation; t is the temperature; ε is the error term, which obeys the normal distribution N(0, σ 2), with σ ranging from 0.01 to 0.05. The specific membership function expressions for wind speed, precipitation, and temperature are omitted. The robot operation thresholds are set as follows: wind speed no more than 12 m / s, precipitation no more than 3 mm / h, and temperature range from -5°C to 45°C. Based on real-time weather data and forecast data, the robot operation risk index (RI) is calculated for each time period within the next 24 hours to generate an operation risk assessment report. This step aims to assess the impact of weather on robot cleaning operations, avoid performing cleaning tasks under adverse weather conditions, and ensure robot safety and cleaning effectiveness.

[0152] The specific implementation of step S05 is to divide the 24 hours of a day into high power generation period (usually 10:00-14:00), medium power generation period (usually 7:00-10:00 and 14:00-17:00) and low power generation period (usually before sunrise and after 17:00) based on the solar radiation intensity time distribution curve. The time series analysis method is used to analyze the historical radiation intensity data and establish a typical sunshine pattern library. For different stain types, the influence weight W of each stain in different power generation period is calculated. impact For example, the impact of solid particle stains is higher during high power generation periods. The optimal cleaning time window T opt The calculation formula is: opt =argmin T [w1·∫ T L power (t)dt+w2·∫ T E clean (t)dt+w3·∫ T RI(t)dt]+ε; where T opt is the optimal cleaning time window; T is the candidate time window; L power (t) is the power generation loss function at time t; E clean (t) is the cleaning efficiency function at time t; RI(t) is the operation risk index at time t; w1, w2, w3 are weight coefficients, and they satisfy w1+w2+w3=1; ε is the random error term, which obeys the normal distribution N(0, σ 2 ), the value range of σ is 0.01~0.05. Power generation loss function L power The calculation formula of (t) is: power (t) = W impact (t)·P theoretical (t)·L eff Where, L power (t) is the power generation loss at time t; W impact (t) is the influence weight of time t; P theoretical (t) is the theoretical power generation at time t; L eff Is the power generation efficiency loss rate. Impact weight Wimpact (t) Determined by time period: Where, T high It is the high power generation period (10:00~14:00); T mid The middle power generation period (7:00~10:00 and 14:00~17:00); T low Low power generation period (before sunrise and after 17:00); high 、w mid 、w low is the weight coefficient of each period, usually satisfying w high >w mid >w low , the typical value is w high =1.0,w high =0.6, w low = 0.2. This step aims to determine the optimal cleaning time, balance the impact of cleaning on power generation and cleaning efficiency, and maximize the overall benefits of the photovoltaic system.

[0153] The specific implementation method of step S06 is to match the corresponding cleaning scheme based on the four types of stains identified in step S02. Water-soluble stains are rinsed with pure water, using deionized water, the water temperature is set to 20-30°C, and the spraying pressure is 0.3-0.6MPa. Grease stains are rinsed with an additive, adding anionic surfactants to deionized water at a concentration of 0.1-0.3%, and a spraying pressure of 0.5-0.8MPa. Solid particle stains are rinsed with a combination of dry scrubbing or low-pressure water rinsing, with a scrubbing speed of 60-120 rpm and a contact pressure of 0.1-0.2kPa. Crystalline stains are rinsed with an additive, adding a weak acid detergent to deionized water at a concentration of 0.2-0.5%, a spraying pressure of 0.6-1.0MPa, and an action time of 60-120 seconds. Adjust the cleaning liquid ratio and spraying pressure in real time according to the ambient temperature and humidity. The cleaning parameter adjustment formula is: P spray (T, H) = P base ·f T (T)·f H (H)+ε P Where, P spray (T, H) is the adjusted spraying pressure, in MPa; P base is the reference spraying pressure, in MPa; f T (T) is the temperature adjustment factor; f H (H) is the humidity adjustment factor; ε P is the error term, which obeys the normal distribution σ P The value range is 0.01~0.05. Temperature adjustment factor f T (T) and humidity adjustment factor fH The calculation formula for (H) is:

[0154] Where, T ref is the reference temperature, usually 20℃; H ref is the reference humidity, usually 60%; k T is the temperature adjustment coefficient, the typical value is 0.01~0.03 / ℃; k H is the humidity adjustment coefficient, with a typical value of 0.005 to 0.01%. The detergent ratio adjustment formula is: C agent (T)=C base ·g T (T)+ε C Where, C agent (T) is the adjusted detergent concentration, in %; C base is the base detergent concentration, in %; g T (T) is the temperature ratio adjustment factor; ε C is the error term, which obeys the normal distribution σ C The value range is 0.01 to 0.05. In low-temperature environments (below 5°C), the cleaning solution temperature is increased. In high-humidity environments (above 80%), the water volume is reduced and wind-assisted drying is added. This step aims to select the most appropriate cleaning method for different stain types, improving cleaning efficiency and effectiveness while reducing water consumption and wear on photovoltaic panels.

[0155] The specific implementation of step S07 is to call the stain cleaning efficiency optimization function to optimize the cleaning scheme in multiple dimensions. This function is implemented based on the genetic algorithm and the stain type distribution matrix M is used. dirt , environmental parameter matrix M env , Clean resource constraint vector V resource , power generation period weight coefficient W generation and robot performance parameters P robot As input. The expression of the stain cleaning efficiency optimization function is: {M solution , E clean}=argmax X F(X,M dirt , M env , V resource , W generation , P robot )+ε; where M solution is the optimized cleaning solution matrix; E clean is the expected cleaning effect index; X is the decision variable matrix, including parameters such as cleaning method selection, detergent ratio, operating pressure, and operating speed; F is the objective function; M dirtis the stain type distribution matrix; M env is the environmental parameter matrix; V resource is the clean resource constraint vector; W generation is the power generation period weight coefficient; P robot is the robot performance parameter; ε is the random error term, which obeys the normal distribution N(0, σ 2 ), the value range of σ is 0.01~0.05. The specific expression of the objective function F is: Where, E eff C is the cleaning efficiency index; resource is the resource consumption indicator; I env G is the environmental impact indicator; power is the power generation income index; w1, w2, w3, w4 are weight coefficients, and they satisfy w1+w2+w3+w4=1; ε F is the error term, which obeys the normal distribution σ F The value range is 0.05~0.1. Cleaning efficiency index E eff The calculation formula is: Where, E eff It is the cleaning efficiency index, ranging from 0 to 1; A i，j is the area of the jth type of stain in the ith region; R i，j is the removal rate of the jth type of stain in the ith area; e i，j is the cleaning efficiency coefficient of the jth type of stain in the ith region; n is the total number of regions. The optimization process includes population initialization, fitness evaluation, selection, crossover and mutation operations. After multiple generations of evolution, the optimized cleaning solution matrix M is output. solution and expected cleaning effect index E clean The cleaning plan matrix includes parameters such as the cleaning method selection, cleaning agent ratio, operating pressure, and operating speed for each area. The expected cleaning effect index is used to evaluate the expected results after the plan is implemented. This step aims to comprehensively consider multiple factors to generate the optimal cleaning strategy, balancing cleaning effect, resource consumption, environmental impact, and power generation benefits.

[0156] The specific implementation of step S08 is to transform the environmental parameter matrix M env , stain distribution data D dirt 、Historical cleaning record H clean and photovoltaic panel layout information L panelThis input is fed into the PV cleaning planning network model to generate the optimal cleaning path and parameter configuration. The PV cleaning planning network model is based on a hybrid architecture combining a graph neural network and a multi-head adaptive attention mechanism. It consists of four main components: an environment encoding module, a soiling representation module, a path planning module, and a parameter generation module. The environment encoding module uses a bidirectional long short-term memory (BiLSTM) network to process time-series environmental data and capture environmental trends. The soiling representation module uses a convolutional neural network to extract soiling distribution features and generate a soiling density heatmap. The path planning module generates cleaning paths based on an improved ant colony algorithm, optimizing to minimize path length and maximize cleaning coverage. The parameter generation module uses a fully connected neural network to generate cleaning parameters, including nozzle angle, water pressure, and scrubbing force, based on soiling type and environmental conditions. The number of attention heads in the multi-head adaptive attention mechanism is dynamically adjusted based on the size and complexity of the PV array, and the attention weight distribution is determined based on the weather gradient and soiling distribution non-uniformity coefficient in the environmental parameter matrix. The model employs an encoder-decoder architecture, with the encoder processing the input environmental and soiling information and the decoder generating a cleaning path sequence and corresponding operation parameter configuration. The PV cleaning planning network model training process first collects environmental data, stain accumulation patterns, cleaning operation records, and power generation efficiency change data from multiple PV plants under different seasons and climate conditions. Multispectral cameras and infrared imaging equipment are used to perform high-precision scans of PV panel surfaces to construct a realistic annotated set of stain type distributions. Physical simulations are combined with field measurements to establish correlations between environmental parameters and stain accumulation rates. Data augmentation techniques are used to simulate cleaning scenarios under extreme weather conditions to expand the dataset's coverage. The training process involves pre-training the model infrastructure in a simulated environment, followed by fine-tuning with real PV plant data. Reinforcement learning is used to train the path planning module, while comparative learning is used to train the stain characterization module. Transfer learning is used to adapt the pre-trained model to PV plants in different locations and climates. Finally, end-to-end joint optimization training is performed. This step aims to generate efficient cleaning execution plans that maximize resource utilization and minimize the impact on PV power generation.

[0157] To better understand and implement the present invention, Example 2, a specific application scenario, is provided below: A photovoltaic power station located on the edge of a desert has an installed capacity of 50 MW, comprises 162,000 solar panels, and covers an area of approximately 80 hectares. The area is prone to frequent wind and sandstorms year-round, which easily accumulates dust on the panels. Seasonal sandstorms are also frequent, severely impacting power generation efficiency.

[0158] First, the researchers deployed an environmental monitoring sensor network within the photovoltaic power plant area, comprising 48 sensor nodes. Each node integrated sensors for precipitation, wind speed, temperature, humidity, UV intensity, and particulate matter concentration, forming an environmental parameter matrix. Monitoring data was collected every 10 minutes for 12 consecutive months, accumulating over 2.5 million data points. Based on this data, an environmental contamination association model was constructed, with the corresponding weight coefficients for each environmental parameter shown in Table 1:

[0159] Table 1 Environmental parameter weight coefficient table

[0160]

[0161] The researchers then used a drone equipped with a multispectral camera to scan the surface of the photovoltaic panels. The camera covers visible light (400-700 nm), near-infrared (700-1100 nm), and short-wave infrared (1100-2500 nm) bands, with a resolution of 0.3 mm / pixel. The scanned data was processed using an improved U-Net convolutional neural network to identify the distribution of four types of stains. The distribution of stains on the photovoltaic panel surface in a certain area is shown in Table 2:

[0162] Table 2 Distribution of photovoltaic panel stains in a certain area

[0163] Stain Type Coverage rate (%) Average thickness (mm) Water-soluble stains 18.5 0.25 Grease stains 5.2 0.18 Solid particle stains 35.6 0.42 Crystalline stains 8.7 0.31

[0164] By analyzing the real-time power generation data of the photovoltaic power station, the photovoltaic panel cleanliness index was calculated to be 0.78. According to the mapping relationship model between stain coverage and power generation efficiency loss rate, the mapping function parameters for various types of stains are shown in Table 3:

[0165] Table 3 Parameters of stain mapping function

[0166]

[0167]

[0168] Based on the above data, the cleaning priority matrix was calculated and the cleaning priority was determined. The weather conditions and robot operation risk index for a certain day are shown in Table 4:

[0169] Table 4 Weather conditions and operation risk index for a certain day

[0170] Time period Wind speed (m / s) Precipitation (mm / h) Temperature (℃) Operational Risk Index 06:00-08:00 2.5 0 12 0.15 08:00-10:00 3.8 0 18 0.22 10:00-12:00 6.2 0 25 0.45 12:00-14:00 7.5 0 28 0.62 14:00-16:00 6.8 0 26 0.48 16:00-18:00 4.2 0 22 0.25 18:00-20:00 2.0 0 16 0.12

[0171] Based on the sunlight intensity distribution curve, the day is divided into high power generation periods (10:00-14:00), medium power generation periods (07:00-10:00 and 14:00-17:00), and low power generation periods (the rest of the day). Weight coefficients for each period are 1.0, 0.6, and 0.2, respectively. The optimal cleaning time windows were calculated using a multi-objective optimization algorithm and determined to be 06:00-08:00 and 18:00-20:00.

[0172] For the four types of stains identified, the corresponding cleaning solutions are matched. Water-soluble stains are rinsed with pure water, using deionized water, the water temperature is set to 25°C, and the spraying pressure is 0.4MPa. Grease stains are rinsed with an additive, adding anionic surfactants to deionized water at a concentration of 0.2% and a spraying pressure of 0.7MPa. Solid particle stains are rinsed with a combination of dry scrubbing and low-pressure water rinsing, with a scrubbing speed of 90 rpm, a contact pressure of 0.15kPa, and a water pressure of 0.3MPa. Crystalline stains are rinsed with an additive, adding a weak acid detergent to deionized water at a concentration of 0.3%, a spraying pressure of 0.8MPa, and an action time of 90 seconds.

[0173] Based on the environmental parameter matrix, stain distribution data, cleaning resource constraints, and robot performance parameters, the stain cleaning efficiency optimization function is called for optimization. The resulting cleaning resource allocation is shown in Table 5:

[0174] Table 5 Cleaning resource allocation table

[0175] Resource Type Total Water-soluble stains Grease stains Solid particle stains Crystalline stains Water volume (L) 1200 350 220 380 250 Cleaner (L) 15 0 4.5 0 10.5 Energy (kWh) 25 5 6 8 6 Time (min) 180 45 30 65 40

[0176] Finally, the environmental parameter matrix, stain distribution data, historical cleaning records, and photovoltaic panel layout information are input into the photovoltaic cleaning planning network model to generate the optimal cleaning path. The model is based on a graph neural network and a multi-head adaptive attention mechanism, which includes an environmental encoding module, a stain representation module, a path planning module, and a parameter generation module. The number of attention heads in the multi-head adaptive attention mechanism is calculated according to the formula N. head =max(4,log2(N panel )) Determine, where N panel is the number of photovoltaic panels. For the 162,000 photovoltaic panels in this power station, we can calculate N head =18.

[0177] The comparison of power generation efficiency before and after cleaning is shown in Table 6:

[0178] Table 6 Comparison of power generation efficiency before and after cleaning

[0179] index Before cleaning After cleaning Improvement Cleanliness Index 0.78 0.96 23.1％ Average daily power generation (MWh) 215 262 21.9％ Conversion efficiency (%) 15.6 19.2 23.1％ Stain coverage (%) 68.0 5.2 92.4% reduction

[0180] Traditional photovoltaic panel cleaning systems typically rely on manual or simple mechanical cleaning at fixed intervals, failing to account for variations in soil type, changing environmental conditions, and the impact of power generation periods. This results in low cleaning efficiency, significant resource waste, and a significant impact on power generation. Traditional methods rely primarily on manual experience to determine cleaning times and methods, lacking a scientific basis and unable to optimize cleaning strategies based on real-time environmental and soiling conditions.

[0181] In contrast, the present invention builds a complete intelligent cleaning system that acquires environmental data through multiple sensors, employs multispectral technology to identify stain types, establishes a mapping relationship between stains and power generation efficiency losses, comprehensively considers weather adaptability, the impact of power generation time periods, and resource constraints, and uses a multi-objective optimization algorithm and a photovoltaic cleaning planning network model to generate an optimal cleaning strategy. This invention achieves precise allocation of cleaning resources, significantly improves cleaning efficiency, reduces interference with power generation, and reduces resource consumption. Furthermore, the present invention's closed-loop feedback system can adjust parameters in real time based on cleaning results, achieving adaptive optimization and further enhancing the system's intelligence and adaptability.

[0182] It should be noted that the variables involved in the present invention are explained in detail as shown in Tables 7, 8 and 9 below.

[0183] Table 7 Variable Explanation Table (Part 1)

[0184]

[0185]

[0186] Table 9 Variable Explanation Table (Part 3)

[0187]

[0188]

[0189] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field can easily think of changes or replacements within the technical scope disclosed by the present invention, which should be covered by the scope of protection of the present invention.

Claims

1. A method for determining a photovoltaic panel cleaning solution taking the environment into consideration, characterized in that: include: Real-time weather data and forecast data for the photovoltaic panel monitoring area are obtained to form an environmental parameter matrix and establish an environmental stain association model; a multispectral camera is used to scan the surface of the photovoltaic panel, identify the surface stain type, and calculate the stain coverage area ratio and thickness distribution map; the photovoltaic panel cleaning index is calculated based on the real-time power generation data of the photovoltaic power station, and a mapping relationship between the stain coverage rate and the power generation efficiency loss rate is established; a weather adaptability scoring system is constructed and the robot operation threshold is set; the optimal cleaning time window is calculated based on the sunshine intensity time distribution curve; the corresponding cleaning plan is matched for the identified stain type; the stain cleaning efficiency optimization function is called to optimize the cleaning plan; the environmental parameter matrix, stain distribution data, historical cleaning records and photovoltaic panel layout information are input into the photovoltaic cleaning planning network model to generate the optimal cleaning path and parameter configuration.

2. The method for determining a photovoltaic panel cleaning solution taking the environment into consideration according to claim 1, characterized in that: The acquisition of real-time weather data and forecast data in the photovoltaic panel monitoring area includes: obtaining precipitation, wind speed, temperature, humidity, ultraviolet intensity and dust content, forming an environmental parameter matrix through environmental monitoring sensor networking, and establishing an environmental stain association model in combination with historical weather and stain data.

3. The method for determining a photovoltaic panel cleaning solution taking the environment into consideration according to claim 2, characterized in that: The use of a multispectral camera to scan the surface of a photovoltaic panel includes: identifying the type of surface stains and classifying them into four types: water-soluble stains, grease stains, solid particle stains, and crystalline stains, and calculating the coverage area ratio and thickness distribution map of each stain.

4. The method for determining a photovoltaic panel cleaning solution taking the environment into consideration according to claim 3, characterized in that: The photovoltaic panel cleanliness index is a dimensionless parameter that characterizes the cleanliness of the photovoltaic panel and is calculated by comparing theoretical power generation with actual power generation. The value range is 0 to 1, and the closer the value is to 1, the cleaner the photovoltaic panel.

5. The method for determining a photovoltaic panel cleaning solution taking the environment into consideration according to claim 4, characterized in that: The weather adaptability scoring system includes a robot passage difficulty index, which is a parameter that quantifies the impact of different weather conditions on the movement and operation of the cleaning robot, taking into account the factors affecting mechanical performance such as ground slipperiness, wind resistance and temperature.

6. The method for determining a photovoltaic panel cleaning solution taking the environment into consideration according to claim 5, characterized in that: The stain cleaning efficiency optimization function is used to calculate the optimal cleaning strategy parameter combination while considering multiple environmental factors and stain characteristics. The input includes a stain type distribution matrix, an environmental parameter matrix, a cleaning resource constraint vector, a power generation period weight coefficient, and a robot performance parameter. The output is a cleaning scheme matrix and a cleaning effect index including cleaning scheme selection, detergent ratio, operating pressure, operating speed, and expected cleaning effect.

7. The method for determining a photovoltaic panel cleaning solution taking the environment into consideration according to claim 6, characterized in that: The structure of the photovoltaic cleaning planning network model is a hybrid architecture based on graph neural network and multi-head adaptive attention mechanism, which includes four main components: environment encoding module, stain representation module, path planning module and parameter generation module. It adopts an encoder-decoder architecture. The encoder is responsible for processing the input environment and stain information, and the decoder generates the cleaning path sequence and the corresponding operation parameter configuration.

8. The method for determining a photovoltaic panel cleaning solution taking the environment into consideration according to claim 7, characterized in that: The steps for establishing the training dataset for the photovoltaic cleaning planning network model include: collecting environmental data, stain accumulation pattern data, cleaning operation records, and power generation efficiency change data from multiple photovoltaic power stations under different seasons and climatic conditions; using multispectral cameras and infrared imaging equipment to perform high-precision scanning of the photovoltaic panel surface to construct a true annotation set of stain type distribution; and recording the actual effects and resource consumption of different cleaning strategies under various environmental conditions.

9. The method for determining a photovoltaic panel cleaning solution taking the environment into consideration according to claim 8, characterized in that: The photovoltaic cleaning planning network model training process includes: pre-training the model infrastructure in a simulated environment; introducing real photovoltaic power station data for fine-tuning training; using reinforcement learning methods to train the path planning module; training the stain characterization module through comparative learning; using transfer learning technology to adapt the pre-trained model to photovoltaic power stations in different geographical locations and climatic conditions; and performing end-to-end joint optimization training.

10. The method for determining a photovoltaic panel cleaning solution taking the environment into consideration according to claim 9, characterized in that: High power generation period specifically refers to the daytime period when solar radiation intensity is the highest and lasts the longest.

Images