Intelligent cleaning method for photovoltaic module

Abstract

This invention relates to an intelligent cleaning method for photovoltaic modules, comprising: multi-source data acquisition and preprocessing to form an input matrix containing spatiotemporal correlation features; extracting image features from the input matrix using a convolutional neural network architecture to identify dust density, stain type, and coordinates of obstructions on the photovoltaic modules; calculating a pollution index and using a sliding window mechanism to adjust weight coefficients in real time to respond to environmental changes; setting cleaning priorities based on the pollution index using a fuzzy logic system with triangular membership functions combined with array layout; using an improved A* algorithm for cleaning path planning to generate an energy-efficient obstacle avoidance path; a cleaning robot working along the generated obstacle avoidance path and dynamically switching multimodal actuators based on stain identification results; and constructing a closed-loop evaluation system based on the power generation efficiency improvement rate and the difference in image cleanliness. Compared with existing technologies, this invention has advantages such as improved cleaning efficiency and reduced operation and maintenance costs.

Description

Technical Field

[0001] This invention relates to a method for cleaning photovoltaic modules, and more particularly to an intelligent method for cleaning photovoltaic modules. Background Technology

[0002] In actual operation, photovoltaic (PV) modules face challenges such as surface contamination (e.g., dust, bird droppings, oil) and environmental fluctuations (e.g., temperature, wind speed, light intensity), leading to decreased power generation efficiency and performance degradation. Current PV module cleaning technologies primarily employ control methods based on fixed-period triggering, simple environmental thresholds (e.g., cumulative number of rainless days), or basic sensor signals (e.g., light transmittance from dust sensors). These systems typically consist of environmental monitoring units deployed at PV sites, simple dust sensors, and actuators (e.g., fixed spray systems or track-mounted cleaning robots). The workflow is as follows: the environmental monitoring unit continuously collects data; when preset conditions are met, the actuator is activated to perform a uniform cleaning of all or a specific area of ​​the PV modules. Some improved solutions introduce simple image recognition technology, analyzing overall grayscale changes in module images to determine whether cleaning is necessary. The entire system operates on a linear logic of "monitoring-threshold judgment-execution," failing to adequately consider differences in contamination between modules, types of dirt, and the dynamic impact of changing environments.

[0003] A search of Chinese Patent Publication No. CN117593675A reveals an intelligent system for a photovoltaic cleaning robot, specifically comprising a control terminal, an intelligent control center, a stain detection database, and intelligent devices. The intelligent devices include a camera module, a motion flight module, and a cleaning module. Both the intelligent devices and the intelligent control center are integrated onto the photovoltaic cleaning robot. The intelligent control center is connected to the control terminal via wireless communication, and the intelligent control center is also connected to the stain detection database via wireless communication. However, while this existing patent uses a camera module to capture images and employs GhostNet convolutional computation for recognition, its core is based on sensor data and artificial intelligence algorithms to achieve an intelligent upgrade from "perception" to "decision-making." However, it focuses on identifying the location of photovoltaic panels and route planning, resulting in limited dimensions and low precision in stain recognition. Furthermore, it relies on static thresholds and fails to clearly define the dynamic adaptability of its model, leading to lag or inaccurate judgments in changing environments. Additionally, it lacks comprehensive and quantitative evaluation indicators.

[0004] Therefore, the existing technology has the following problems that need to be addressed: 1) Existing photovoltaic module cleaning methods usually rely on fixed cleaning cycles or single environmental parameter thresholds, which makes it impossible to accurately perceive the actual type, distribution density and dynamic changes of dirt on the module surface. There are common problems of "over-cleaning" or "under-cleaning". They lack the ability to integrate and analyze multi-source information and cannot construct a dynamic index that reflects the real pollution status. 2) The cleaning strategy lacks prioritization, making it impossible to prioritize the most polluted or most power-affecting components; the path planning of the cleaning robot does not fully consider the photovoltaic array layout and energy consumption optimization, resulting in low movement efficiency. 3) The existing technology system is an open-loop system with rigid decision-making logic. It fails to form a closed-loop intelligent optimization loop that integrates stain identification, strategy formulation, path planning, precise execution, and effect evaluation. This results in unreasonable allocation of cleaning resources, high operation and maintenance costs, and limited effect on restoring the power generation efficiency of photovoltaic power plants. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the existing technology and provide a smart cleaning method for photovoltaic modules that improves cleaning efficiency and reduces operation and maintenance costs.

[0006] The objective of this invention can be achieved through the following technical solutions: According to one aspect of the present invention, a smart cleaning method for photovoltaic modules is provided, comprising: Multi-source data acquisition and preprocessing form an input matrix containing spatiotemporal correlation features; A convolutional neural network architecture is used to extract image features from the input matrix to identify the dust density, stain type and coordinates of obstructions on photovoltaic modules, and the model recognition accuracy is optimized by transfer learning and cross-entropy loss function. A dynamic threshold model is constructed, which integrates ash density, temperature difference and wind speed ratio in a weighted linear combination to calculate the pollution index, and a sliding window mechanism is used to adjust the weight coefficients in real time to respond to environmental changes. Based on the pollution index, a fuzzy logic system using triangular membership functions is combined with array layout to set cleaning priorities; An improved A* algorithm is used for cleaning path planning to generate an obstacle avoidance path with optimal energy consumption; The cleaning robot operates using a generated obstacle avoidance path and dynamically switches between multimodal actuators based on stain recognition results. A closed-loop evaluation system was constructed by comparing the improvement rate of power generation efficiency with the difference in image cleanliness to verify the cleaning effect and provide feedback on the data.

[0007] As a preferred technical solution, the multi-source data acquisition and preprocessing specifically includes: By deploying cameras, infrared thermal imagers, light intensity sensors, and meteorological monitoring units on the surface of photovoltaic modules, multi-source data on the surface of photovoltaic modules are collected in real time, including images, temperature distribution, light intensity, and environmental parameters. After normalizing the image data, the multi-source data are aligned by timestamps to form an input matrix containing spatiotemporal correlation features.

[0008] As a preferred technical solution, the convolutional neural network architecture extracts image features for stain recognition as follows: After the input image undergoes convolution, the size of the output feature map is calculated. : in The height of the input image. For the number of fill pixels, The kernel size is the convolution kernel size. Step size; After global average pooling, the feature map is mapped to a high-dimensional feature vector through a fully connected layer. Finally, the Softmax function is used to classify and output the ash density, stain type, and coordinates of the occluder.

[0009] As a preferred technical solution, the optimization of model recognition accuracy through transfer learning and cross-entropy loss function specifically involves: The model is fine-tuned based on ImageNet pre-trained weights, and the cross-entropy loss function is used to measure the difference between the model's predictions and the true values. in, The labels are the real image labels, and n is the number of labels. To predict probabilities, we find the point where the loss function is minimized, which improves the model's generalization ability, and then determine the model parameters based on this.

[0010] As a preferred technical solution, the calculation of the pollution index is specifically as follows: in, Pollution index; The dust density obtained from image recognition is obtained by fitting pixel gray values ​​with a calibration curve. The difference between the component surface temperature and the ambient temperature is obtained by subtracting the ambient temperature from the weather station temperature from the surface temperature measured by the infrared thermal imager. The ratio of real-time wind speed to a preset threshold wind speed; coefficient. , , Based on historical data, the desired result was determined using gradient descent optimization, satisfying... + + =1.

[0011] As a preferred technical solution, the coefficient , , The sliding window mechanism updates in real time, automatically increasing the window size when sandstorms are frequent. The weighting of dust accumulation density is increased, while the weighting is increased during high temperatures and strong winds. , The values ​​are used to reinforce the corresponding influence weights; the mango peeling strategy is formulated by changing the weight values ​​of different parameters.

[0012] As a preferred technical solution, the setting of cleaning priority specifically involves: based on the contamination index. The fuzzy logic decision system takes three pollution levels as input variables: low, medium, and high, and the corresponding output actions are delayed cleaning, regular cleaning, and emergency cleaning, respectively. The triangular membership function is used to describe the membership changes of fuzzy sets: in For membership function, for The threshold value for changing from low pollution level to medium pollution level. for The threshold between medium and high pollution levels.

[0013] If, under the same pollution level, the photovoltaic array is further divided into sub-priorities based on its layout location, the photovoltaic modules located at the edge of the power station have a higher priority than those in the central area. The cleaning priority of the photovoltaic modules is then ranked according to this influencing factor, and the ranking result is transmitted to the cleaning robot control system.

[0014] As a preferred technical solution, the method of using the improved A* algorithm for cleaning path planning specifically includes: First, a two-dimensional grid map of the photovoltaic panel array is obtained by drone aerial scanning, and areas with obstacles such as brackets and cables are marked as impassable areas; Then the heuristic function of the energy consumption optimization objective. The design is the sum of the Euclidean distance from the current node to the target node and the turning angle penalty term: in, , The coordinates of the current node. , The coordinates of the target node. For steering angle, For the steering weighting coefficient; Finally, during the traversal, the path with the lowest total cost is selected first: in, The actual cost from the starting point to the current node. This is a heuristic function.

[0015] As a preferred technical solution, the dynamic switching of the multimodal actuator based on the stain recognition result specifically refers to: When the identification result is a low pollution level, only the centrifugal fan is activated to remove loose particles using the negative pressure adsorption force generated by Bernoulli's equation. When the pollution level is identified as medium, the high-pressure water gun and rotating brush roller are activated to work together, using the impact force of the water flow. Friction with the brush roller The combined effect improves cleaning efficiency; When the identification result is a heavy pollution level, each actuator is started in sequence according to the order of fan pre-purging, water gun flushing, and brush roller fine washing.

[0016] As a preferred technical solution, the construction of the closed-loop evaluation system specifically includes: The effect of module cleaning on power generation is reflected by calculating the difference in maximum power point current under the same illumination conditions before and after cleaning: in, This represents the maximum power point current value under the same illumination conditions after cleaning. This represents the maximum power point current value under the same illumination conditions before cleaning. To improve power generation efficiency; For the component surface images before and after cleaning, divide them into blocks and calculate the sum of the absolute values ​​of the mean grayscale differences between each block: in, The normalized grayscale value after cleaning. M represents the normalized grayscale value before cleaning, and M and N represent the pixel values ​​of each image block after block processing. After setting the cleaning... The value determines the cleaning effect; The evaluation data is transmitted to the backend platform as a reference for subsequent cleaning strategy optimization.

[0017] Compared with the prior art, the present invention has the following advantages: 1) This invention achieves accurate identification of the type of dirt and the density of dust accumulation on the surface of photovoltaic modules through multi-source data fusion and convolutional neural network (ResNet-50). Combined with dynamic threshold model, it adaptively adjusts the pollution index weight. Compared with traditional fixed-cycle or single-index cleaning methods, it can more accurately determine the timing and intensity of cleaning, effectively avoiding insufficient cleaning or waste of resources. 2) By introducing a fuzzy logic decision system and an improved A* path planning algorithm, this invention achieves intelligent prioritization of cleaning and energy-optimized cleaning path planning, which not only ensures the timely treatment of highly polluting components, but also significantly improves the operating efficiency and energy utilization of the cleaning robot, and is especially suitable for large-scale, complex photovoltaic arrays. 3) This invention combines a multimodal cleaning actuator with PID real-time control, which can dynamically switch cleaning modes according to the type of stains, and evaluate the closed-loop effect after cleaning by the improvement rate of power generation efficiency and the difference in image cleanliness, thus realizing full-link intelligent closed-loop management from stain identification, strategy formulation, path planning to execution feedback. 4) This invention can significantly improve the power generation efficiency of photovoltaic modules, reduce operation and maintenance costs and water consumption, and at the same time avoid damage to the modules during the cleaning process through precise control, which helps to improve the overall operational stability and economic benefits of photovoltaic power plants. Attached Figure Description

[0018] Figure 1 This is an overall flowchart of the method of the present invention; Figure 2 This is a flowchart illustrating the specific process of stain identification and classification in this invention. Figure 3 This is a flowchart illustrating the specific process for calculating the pollution index in this invention. Figure 4 This is a flowchart illustrating the specific process of prioritizing the cleaning process in this invention. Figure 5 This is a flowchart illustrating the path planning process of the present invention. Figure 6 This is a flowchart illustrating the specific process of multimodal cleaning execution control according to the present invention. Figure 7 This is a flowchart illustrating the closed-loop effect evaluation of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0020] This invention is primarily applied to the field of photovoltaic power plant operation and maintenance cleaning. Through multi-source data fusion sensing, convolutional neural network stain recognition, dynamic pollution index modeling, fuzzy logic priority decision-making, improved A* algorithm path planning, and multimodal adaptive cleaning execution, it significantly improves the accuracy, adaptability, and resource utilization efficiency of photovoltaic module cleaning. It is particularly suitable for photovoltaic power plants with complex pollution types, frequent environmental fluctuations, or large array scales. Its AI-based intelligent recognition and dynamic threshold mechanism effectively distinguishes stain types and pollution levels, overcoming the problems of misoperation, missed cleaning, or resource waste caused by traditional fixed-cycle or simple threshold cleaning. Combined with energy-optimized path planning and a multi-mode collaborative cleaning mechanism, it achieves fully automated and refined control of the cleaning process. This method is applicable to the operation and maintenance management of centralized ground-mounted power plants, distributed rooftop photovoltaics, and agricultural-photovoltaic complementary systems, and has the potential for widespread adoption in the photovoltaic intelligent operation and maintenance industry. It can be integrated into power plant centralized control systems or robot scheduling platforms. Through precise cleaning and closed-loop effect evaluation, it effectively improves system power generation efficiency, reduces operation and maintenance costs, and provides core algorithmic support for the digital operation and maintenance and intelligent upgrading of photovoltaic power plants.

[0021] Compared to existing technologies, this invention offers higher dimensionality and precision in stain recognition. It not only identifies the presence and location of stains but also accurately quantifies and calculates "dust density" and "coordinates of obstructions." In contrast, existing technologies focus on identifying the location of photovoltaic panels and route planning, while this invention innovates in the depth and accuracy of stain perception. Furthermore, it innovatively constructs a dynamically weighted pollution index decision model. This invention introduces a dynamic pollution index model that integrates multi-dimensional parameters and uses a sliding window mechanism to adjust the weights of each parameter in real time. Existing technologies rely on static thresholds and do not clearly define the dynamic adaptability of their models. This invention effectively solves the problem of lag or inaccurate judgment in traditional methods under changing environments. In addition, this invention establishes a complete intelligent closed loop of "perception-decision-execution-evaluation." This invention designs a closed-loop evaluation system for cleaning effects, verifying cleaning quality through two types of quantitative indicators: "power generation efficiency improvement rate" and "image cleanliness difference." Existing technologies generally do not mention this type of closed loop based on effect feedback. The evaluation indicators of this invention are more comprehensive and quantitative, forming a complete optimization loop.

[0022] like Figure 1 As shown, the specific implementation process of the method of the present invention is as follows: 1) Multi-source data acquisition and preprocessing: This invention uses high-definition cameras, infrared thermal imagers, light intensity sensors, and meteorological monitoring units deployed on the surface of photovoltaic modules to acquire real-time images of the module surface, temperature distribution, light intensity, and environmental parameters. The acquired image data is then normalized. in, These are the original image data values ​​acquired from the red channel. These represent the maximum and minimum values ​​of the images acquired from each channel. Z-score normalization is used to eliminate dimensional differences in the acquired environmental parameters. After aligning the multi-source data with timestamps, an input matrix containing spatiotemporal correlation features is formed, providing a foundation for subsequent AI analysis.

[0023] 2) Convolutional Neural Network Feature Extraction and Stain Recognition. A ResNet-50 network architecture is used for image feature extraction. This network contains four residual block layers, each extracting edge, texture, and shape features using a 3×3 convolutional kernel. After the input image undergoes convolution operations, the output feature map size is calculated. in, The height of the input image. For the number of fill pixels, The kernel size is the convolution kernel size. The step size is used. After global average pooling, the feature map is mapped to a high-dimensional feature vector through a fully connected layer. Finally, the Softmax function is used to classify and output the dust density, stain type (dust / bird droppings / oil stains, etc.) and coordinates of the occluder. in, This represents the i-th element of the input vector. Represents an exponential function. This represents the sum of the exponents of all elements, ensuring that the total output sum is 1.

[0024] The model training employs a transfer learning strategy, fine-tuning the model based on ImageNet pre-trained weights, and using the cross-entropy loss function to measure the difference between the model's predictions and the true values. in, The labels are the real image labels, and n is the number of labels. To predict probabilities, we find the point where the loss function is minimized, which improves the model's generalization ability, and then determine the model parameters based on this.

[0025] 3) Dynamic threshold model and pollution index calculation. For example... Figure 3 As shown, a pollution index model is constructed using a weighted linear combination formula combined with environmental parameters: in, Pollution index; The ash density is obtained from image recognition, with units of g / m², and is obtained by fitting pixel gray values ​​to a calibration curve. The difference between the component surface temperature and the ambient temperature is obtained by subtracting the ambient temperature from the weather station temperature from the surface temperature measured by the infrared thermal imager. The ratio of real-time wind speed to a preset threshold wind speed; coefficient. , , Based on historical data, the desired result was determined using gradient descent optimization, satisfying... + + =1. This model updates coefficient weights in real time through a sliding window mechanism, automatically increasing them when sandstorms are frequent. The weighting of dust accumulation density is increased, while the weighting is increased during high temperatures and strong winds. , The values ​​are used to reinforce the corresponding influence weights. A corresponding cleaning strategy is formulated by varying the weight values ​​of different parameters.

[0026] 4) Prioritize fuzzy logic cleaning. For example... Figure 2 As shown, based on the pollution index The fuzzy logic decision-making system takes three pollution levels as input variables: low, medium, and high, and the corresponding output actions are delayed cleaning, regular cleaning, and emergency cleaning, respectively. A triangular membership function is used to describe the changes in the membership degrees of the fuzzy sets. in For membership function, for The threshold value for changing from low pollution level to medium pollution level. for The threshold value for changing from medium pollution level to high pollution level. The rule base can be defined as: if... If 8, then output "Delayed cleaning"; if If so, then output "Regular cleaning"; if If the value is 15, then the output will be "Emergency Cleaning". Under this rule, when... At 3 PM, the pollution level was determined to be "high," with a membership output of 1; after cleaning... Once the pollution level drops to 12 or below, it is classified as "medium pollution" and conventional cleaning methods are used. Within the same pollution level, sub-priority is further determined based on the layout of the photovoltaic array. Modules located at the edge of the power station are more susceptible to surrounding dust, so their sub-priority is higher than that of modules in the central area. The cleaning priority of the modules is ranked according to this influencing factor or other factors, and the ranking result is transmitted to the cleaning robot control system via the Modbus protocol. Figure 4 As shown.

[0027] 5) Improve the adaptive path planning of the A* algorithm. For example... Figure 5As shown, the improved A* algorithm used in this invention introduces photovoltaic array layout constraints and energy consumption optimization objectives on the basis of traditional heuristic search. During path planning, a two-dimensional grid map of the photovoltaic array is first obtained through UAV aerial scanning, marking areas with obstacles such as supports and cables as impassable areas. The heuristic function of the energy consumption optimization objective is then applied... The design is the sum of the Euclidean distance from the current node to the target node and the turning angle penalty term: in, , The coordinates of the current node. , The coordinates of the target node. For steering angle, These are the turning weight coefficients. During the algorithm's traversal, the path with the lowest total cost is selected first. in, The actual cost from the starting point to the current node. This serves as the heuristic function. Considering the parallel arrangement of the photovoltaic array, the steering logic for lateral movement is optimized. When the distance between adjacent components is detected to be less than the robot's turning radius, the steering weight coefficient is adjusted. An obstacle avoidance path with a specific trajectory is generated, and the coordinate sequence of the output path and the action commands are finally sent to the robot motion controller via the CAN bus. ① Inverter efficiency diagnosis 6) Multimodal cleaning execution system. For example... Figure 6 As shown, the cleaning robot is equipped with three independent actuators: a high-pressure water gun, a rotating brush roller, and a centrifugal fan. Its control logic dynamically switches based on the stain type identification level results. Dust Removal Mode: When the identification result is a low pollution level, only the centrifugal fan is activated, using the negative pressure adsorption force generated by Bernoulli's equation to remove loose particles. in, For static pressure, air density, For wind speed, It is a constant.

[0028] Stubborn Stain Mode: When the stain level is identified as medium, the high-pressure water gun and rotating brush roller work together to apply water pressure. Friction with the brush roller The combined effect improves cleaning efficiency: in, For the density of water, For jet velocity, This represents the nozzle area. The coefficient of friction, For contact pressure.

[0029] Mixed Stain Mode: When the identification result indicates a heavily polluted level, each actuator is activated sequentially in the order of "fan pre-purging → water gun rinsing → brush roller fine washing". The parameters of the actuators are adjusted in real time through a PID controller. For example, when the infrared thermal imager detects that the component temperature exceeds 45°C, it will automatically reduce the water gun pressure through PID adjustment to prevent thermal stress damage.

[0030] 7) A closed-loop evaluation system for cleaning effectiveness. For example... Figure 7 As shown, the photovoltaic module cleaning effect evaluation module verifies the cleaning quality through two sets of quantitative indicators: Power generation efficiency improvement rate: The effect of module cleaning on power generation improvement is reflected by calculating the difference in maximum power point current under the same illumination conditions before and after cleaning. in, This represents the maximum power point current value under the same illumination conditions after cleaning. This represents the maximum power point current value under the same illumination conditions before cleaning. To improve power generation efficiency.

[0031] Image cleanliness difference: The absolute sum of the differences in the mean grayscale values ​​of each block is calculated when the component surface images before and after cleaning are segmented. in, The normalized grayscale value after cleaning. M represents the normalized grayscale value before cleaning, and M and N represent the pixel values ​​of each image block after block processing. By setting the values ​​after cleaning... The value determines the cleaning effect. The evaluation data is transmitted wirelessly to the backend platform via LoRa as a reference for subsequent cleaning strategy optimization.

[0032] The main key points of this invention include: (1) This invention achieves accurate identification and classification of surface stains on photovoltaic modules through multi-source data fusion and a ResNet-50-based convolutional neural network, effectively distinguishing dust density, stain type, and spatial distribution. This method breaks through the limitations of traditional single sensor or fixed threshold judgment, and utilizes transfer learning and cross-entropy loss to optimize the model, significantly improving the perception accuracy and adaptability to complex pollution conditions, and providing a reliable data foundation for intelligent cleaning decisions.

[0033] (2) This invention proposes a dynamically weighted pollution index model that integrates multi-dimensional parameters such as image-recognized dust density, temperature difference between component surface and environment, and real-time wind speed ratio. It also uses a sliding window mechanism to optimize the weight coefficients in real time, thereby achieving adaptive adjustment of the cleaning threshold. This model can automatically enhance the influence of corresponding parameters according to different weather conditions such as sandstorms, high temperatures, and strong winds, overcoming the problem of lag or inaccurate judgment in traditional fixed threshold models under changing environments.

[0034] (3) This invention constructs a cleaning priority decision system based on fuzzy logic, which uses a triangular membership function to map the contamination index to "delayed / regular / urgent" cleaning levels, and sorts sub-priorities within the same level by combining array layout. This method realizes the decision upgrade from "whether to clean" to "when to clean and who to clean first", which significantly improves the allocation efficiency of cleaning resources and the systematic nature of cleaning operations.

[0035] (4) The present invention establishes a closed-loop evaluation system for cleaning effect, verifies cleaning quality through two quantitative indicators: power generation efficiency improvement rate and image cleanliness difference, and feeds the evaluation data back to the strategy module to form a full-link optimization closed loop of "perception-decision-execution-evaluation" to continuously improve the system's adaptability and cleaning economy.

[0036] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for intelligent cleaning of photovoltaic modules, characterized in that, include: Multi-source data acquisition and preprocessing form an input matrix containing spatiotemporal correlation features; A convolutional neural network architecture is used to extract image features from the input matrix to identify the dust density, stain type and coordinates of obstructions on photovoltaic modules, and the model recognition accuracy is optimized by transfer learning and cross-entropy loss function. A dynamic threshold model is constructed, which integrates ash density, temperature difference and wind speed ratio in a weighted linear combination to calculate the pollution index, and a sliding window mechanism is used to adjust the weight coefficients in real time to respond to environmental changes. Based on the pollution index, a fuzzy logic system using triangular membership functions is combined with array layout to set cleaning priorities; An improved A* algorithm is used for cleaning path planning to generate an obstacle avoidance path with optimal energy consumption; The cleaning robot operates using a generated obstacle avoidance path and dynamically switches between multimodal actuators based on stain recognition results. A closed-loop evaluation system was constructed by comparing the improvement rate of power generation efficiency with the difference in image cleanliness to verify the cleaning effect and provide feedback on the data.

2. The intelligent cleaning method for photovoltaic modules according to claim 1, characterized in that, The multi-source data acquisition and preprocessing specifically includes: By deploying cameras, infrared thermal imagers, light intensity sensors, and meteorological monitoring units on the surface of photovoltaic modules, multi-source data on the surface of photovoltaic modules are collected in real time, including images, temperature distribution, light intensity, and environmental parameters. After normalizing the image data, the multi-source data are aligned by timestamps to form an input matrix containing spatiotemporal correlation features.

3. The intelligent cleaning method for photovoltaic modules according to claim 1, characterized in that, The convolutional neural network architecture extracts image features for stain recognition as follows: After the input image undergoes convolution, the size of the output feature map is calculated. : in The height of the input image. For the number of fill pixels, The kernel size is the convolution kernel size. Step size; After global average pooling, the feature map is mapped to a high-dimensional feature vector through a fully connected layer. Finally, the Softmax function is used to classify and output the ash density, stain type, and coordinates of the occluder.

4. The intelligent cleaning method for photovoltaic modules according to claim 1, characterized in that, The optimization of model recognition accuracy through transfer learning and cross-entropy loss function specifically involves: The model is fine-tuned based on ImageNet pre-trained weights, and the cross-entropy loss function is used to measure the difference between the model's predictions and the true values. in, The labels are the real image labels, and n is the number of labels. To predict probabilities, we find the point where the loss function is minimized, which improves the model's generalization ability, and then determine the model parameters based on this.

5. The intelligent cleaning method for photovoltaic modules according to claim 1, characterized in that, The calculation of the pollution index is specifically as follows: in, Pollution index; The dust density obtained from image recognition is obtained by fitting pixel gray values ​​with a calibration curve. The difference between the component surface temperature and the ambient temperature is obtained by subtracting the ambient temperature from the weather station temperature from the surface temperature measured by the infrared thermal imager. The ratio of real-time wind speed to a preset threshold wind speed; coefficient. , , Based on historical data, the desired result was determined using gradient descent optimization, satisfying... + + =1.

6. The intelligent cleaning method for photovoltaic modules according to claim 5, characterized in that, The coefficient , , The sliding window mechanism updates in real time, automatically increasing the window size when sandstorms are frequent. The weighting of dust accumulation density is increased, while the weighting is increased during high temperatures and strong winds. , The values ​​are used to reinforce the corresponding influence weights; the mango peeling strategy is formulated by changing the weight values ​​of different parameters.

7. The intelligent cleaning method for photovoltaic modules according to claim 1, characterized in that, The setting of cleaning priority is specifically based on the contamination index. The fuzzy logic decision system takes three pollution levels as input variables: low, medium, and high, and the corresponding output actions are delayed cleaning, regular cleaning, and emergency cleaning, respectively. The triangular membership function is used to describe the membership changes of fuzzy sets: in For membership function, for The threshold value for changing from low pollution level to medium pollution level. for The threshold value from medium pollution level to high pollution level; If, under the same pollution level, the photovoltaic array is further divided into sub-priorities based on its layout location, the photovoltaic modules located at the edge of the power station have a higher priority than those in the central area. The cleaning priority of the photovoltaic modules is then ranked according to this influencing factor, and the ranking result is transmitted to the cleaning robot control system.

8. The intelligent cleaning method for photovoltaic modules according to claim 1, characterized in that, The specific steps for using the improved A* algorithm for cleaning path planning are as follows: First, a two-dimensional grid map of the photovoltaic panel array is obtained by drone aerial scanning, and areas with obstacles such as brackets and cables are marked as impassable areas; Then the heuristic function of the energy consumption optimization objective. The design is the sum of the Euclidean distance from the current node to the target node and the turning angle penalty term: in, , The coordinates of the current node. , The coordinates of the target node. For steering angle, For the steering weighting coefficient; Finally, during the traversal, the path with the lowest total cost is selected first: in, The actual cost from the starting point to the current node. This is a heuristic function.

9. The intelligent cleaning method for photovoltaic modules according to claim 1, characterized in that, The specific steps of dynamically switching the multimodal actuator based on the stain recognition result are as follows: When the identification result is a low pollution level, only the centrifugal fan is activated to remove loose particles using the negative pressure adsorption force generated by Bernoulli's equation. When the pollution level is identified as medium, the high-pressure water gun and rotating brush roller are activated to work together, using the impact force of the water flow. Friction with the brush roller The combined effect improves cleaning efficiency; When the identification result is a heavy pollution level, each actuator is started in sequence according to the order of fan pre-purging, water gun flushing, and brush roller fine washing.

10. The intelligent cleaning method for photovoltaic modules according to claim 1, characterized in that, The specific steps for constructing the closed-loop evaluation system are as follows: The effect of module cleaning on power generation is reflected by calculating the difference in maximum power point current under the same illumination conditions before and after cleaning: in, This represents the maximum power point current value under the same illumination conditions after cleaning. This represents the maximum power point current value under the same illumination conditions before cleaning. To improve power generation efficiency; For the component surface images before and after cleaning, divide them into blocks and calculate the sum of the absolute values ​​of the mean grayscale differences between each block: in, The normalized grayscale value after cleaning. M represents the normalized grayscale value before cleaning, and M and N represent the pixel values ​​of each image block after block processing. After setting the cleaning... The value determines the cleaning effect; The evaluation data is transmitted to the backend platform as a reference for subsequent cleaning strategy optimization.

Images