Photovoltaic power station cleaning robot path planning method, medium and system

Abstract

The invention provides a photovoltaic power station cleaning robot path planning method, medium and system, and belongs to the technical field of photovoltaic power station operation and maintaining.The photovoltaic power station cleaning robot path planning method includes the steps that firstly, a photovoltaic panel grid division model is established, and stain distribution information is obtained through the image processing technology to construct a stain distribution matrix; secondly, calculating a cleaning water demand matrix based on stain distribution, and meanwhile, considering an influence relationship among vertical column grid units; determining the cleaning duration by combining the efficiency of the cleaning device; constructing a robot energy consumption model; calculating an optimal cleaning path by applying a resource optimization distribution function; integrating the environmental parameters and inputting the environmental parameters into a photovoltaic cleaning neural network model to predict the actual cleaning difficulty; a cleaning strategy is dynamically adjusted accordingly; and executing the final path and recording data to update the related model. According to the method, through real-time sensing and analysis of the stain distribution characteristics, intelligence of path planning of the cleaning robot is realized, and the technical problem that the cleaning robot cannot perform intelligent path planning according to the surface stain distribution characteristics of the photovoltaic panel is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of photovoltaic power station operation and maintenance, and in particular relates to a path planning method, medium and system for a photovoltaic power station cleaning robot. Background Art

[0002] Photovoltaic power generation, a key form of clean energy, has a power generation efficiency that is closely linked to the surface cleanliness of photovoltaic panels. Studies have shown that dust and dirt accumulated on the surface of photovoltaic panels can reduce power generation efficiency by 5% to 30%. Traditional photovoltaic power station cleaning relies primarily on manual labor or simple mechanized cleaning, which is labor-intensive, inefficient, and poses numerous safety risks. With the advancement of automation technology, photovoltaic cleaning robots are increasingly being used in photovoltaic power station maintenance, primarily performing regular cleaning of photovoltaic panels along pre-set, fixed paths.

[0003] However, existing photovoltaic cleaning robots generally use simple linear or spiral fixed-path planning methods, which are unable to perceive and identify the actual distribution of stains on the photovoltaic panel surface. The distribution of stains on the photovoltaic panel surface is significantly affected by environmental factors such as wind direction, rainfall, and the surrounding environment, and exhibits significant unevenness. Fixed-path cleaning strategies lead to irrational allocation of cleaning resources, with some areas being overly cleaned and wasted, while others being under-cleaned and impacting the cleaning effect.

[0004] Currently, photovoltaic cleaning robots lack the ability to intelligently plan paths based on soil distribution characteristics. They are unable to dynamically adjust cleaning paths and parameters based on real-time soil distribution information. This severely impacts cleaning efficiency and the overall power generation performance of photovoltaic systems, becoming a technical bottleneck restricting the development of intelligent operation and maintenance for photovoltaic power plants. In other words, existing technologies hinder the ability of photovoltaic power plant cleaning robots to intelligently plan paths based on soil distribution characteristics on photovoltaic panel surfaces. Summary of the Invention

[0005] In view of this, the present invention provides a path planning method, medium and system for a photovoltaic power station cleaning robot, which can solve the technical problem in the prior art that the photovoltaic power station cleaning robot cannot perform intelligent path planning based on the distribution characteristics of stains on the photovoltaic panel surface.

[0006] The present invention is implemented as follows: The first aspect of the present invention provides a path planning method for a photovoltaic power station cleaning robot, including: establishing a photovoltaic panel grid division model, dividing each photovoltaic panel into m×n grid units, and numbering each grid unit; obtaining the degree of stain of each grid unit through image processing technology, and constructing a stain distribution matrix; calculating the cleaning water requirement of each grid unit based on the stain distribution matrix, considering the water requirement correlation index between grid units in the same vertical column, and constructing a cleaning water requirement matrix; calculating the cleaning time matrix based on the cleaning water requirement of each grid unit and the cleaning efficiency per unit time of the cleaning device; constructing a robot energy consumption model; applying a resource optimization allocation function to perform water distribution and path planning; integrating photovoltaic panel temperature sensor data and environmental parameters, inputting a photovoltaic cleaning neural network model, and predicting the actual cleaning difficulty coefficient of each grid unit; dynamically adjusting the cleaning path and parameters based on the output results of the photovoltaic cleaning neural network model; and executing the final cleaning path.

[0007] The stain distribution matrix refers to a two-dimensional matrix constructed according to the degree of stains on the surface of the photovoltaic panel. The element value of the stain distribution matrix ranges from 0 to 1, where 0 indicates no stains and 1 indicates the most serious stains. It is obtained through image processing technology.

[0008] Among them, the water demand correlation index refers to the coefficient that describes the degree of correlation between the stain cleaning between adjacent grid cells in the same vertical column. The value range is 0 to 1. When the water demand correlation index is 0, it means that the grid cells are completely independent, and when the water demand correlation index is 1, it means that they are completely correlated. The cleaning water demand matrix is a two-dimensional matrix. The calculation of the elements of the cleaning water demand matrix takes into account the influencing factors of the stain degree of the current grid cell and the stain cleaning of the upper grid cell.

[0009] Among them, the total energy consumption function takes into account the energy consumption of the robot motor, the working energy consumption of the water pump, and the energy consumption of the control system, and is expressed as the sum of three parts, of which the first part is proportional to the travel distance, the second part is proportional to the total water consumption, and the third part is a fixed value.

[0010] Among them, the resource optimization allocation function is used to achieve global optimization of cleaning paths and water distribution under limited resource constraints. The input of the resource optimization allocation function includes the cleaning water demand matrix, the cleaning time matrix, the remaining water volume of the robot, the remaining power of the robot, and the photovoltaic panel layout information. The output of the resource optimization allocation function is the optimal cleaning path sequence and the actual water distribution of each grid unit.

[0011] Among them, the specific structure of the photovoltaic cleaning neural network model is a hybrid architecture combining a multi-layer convolutional neural network and an attention mechanism. The convolution layer is responsible for extracting the stain distribution characteristics, the attention mechanism is used to capture the stain correlation between different grid cells, and the fully connected layer outputs the cleaning difficulty prediction value; the photovoltaic cleaning neural network model adopts a dynamic attention threshold mechanism, and the dynamic attention threshold is jointly determined by the inclination angle of the photovoltaic panel, the current ambient temperature and the light intensity.

[0012] Among them, the steps of establishing the training data set in the photovoltaic cleaning neural network model training process specifically include collecting images of photovoltaic panel stains under different weather conditions in various types of photovoltaic power stations, recording the actual cleaning water consumption and cleaning time for each type of stain, and establishing labeled data based on the stain distribution and actual cleaning effect.

[0013] The training steps for the photovoltaic cleaning neural network model include first training basic visual feature extraction capabilities on multiple data sets, then performing transfer learning for the photovoltaic panel stain classification task, then introducing environmental factor input channels for joint training, and finally performing reinforcement learning optimization through actual cleaning robot feedback data.

[0014] A second aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores program instructions. When the program instructions are run in a computer, the program instructions are used to execute the above-mentioned photovoltaic power station cleaning robot path planning method.

[0015] The third aspect of the present invention provides a photovoltaic power station cleaning robot path planning system, which includes the above-mentioned computer-readable storage medium. The system is any one of a computer, a server, and a single-chip microcomputer. The computer-readable storage medium is arranged in the system, and the system is provided with a microprocessor for executing the program instructions stored in the computer-readable storage medium.

[0016] This invention establishes a photovoltaic panel grid model and a stain distribution matrix, combining image processing technology to obtain real-time stain distribution information, enabling intelligent planning and dynamic adjustment of cleaning paths. Based on the stain distribution matrix, the invention calculates the cleaning requirements of each grid cell and uses the degree of stain as the decision-making basis for path planning, aligning cleaning resource allocation with actual needs. Simultaneously, a photovoltaic cleaning neural network model analyzes the relationship between surface stain characteristics and environmental parameters, predicting the actual cleaning difficulty coefficient for different areas and further optimizing cleaning strategies. Furthermore, the system continuously updates the stain distribution model based on actual performance data obtained during the cleaning process, forming a closed-loop optimization mechanism.

[0017] Through the above-mentioned technical means, the present invention solves the core problem that photovoltaic cleaning robots are unable to perform intelligent path planning based on the distribution characteristics of stains, so that the cleaning process can not only strengthen treatment of key polluted areas, but also avoid wasting resources in slightly polluted areas, significantly improving cleaning efficiency and photovoltaic power generation performance, and providing effective technical support for the intelligent operation and maintenance of photovoltaic power stations. 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 path planning method for a photovoltaic power station cleaning robot provided by the first aspect of the present invention. The method includes the following steps:

[0021] S01. Establish a photovoltaic panel grid division model, divide each photovoltaic panel into m×n grid units, and number and identify each grid unit;

[0022] S02. Obtaining the degree of stain of each grid unit by image processing technology, and constructing a stain distribution matrix, wherein each element in the stain distribution matrix represents the degree of stain of the corresponding grid unit;

[0023] S03. Calculating the cleaning water requirement of each grid unit based on the stain distribution matrix, while considering the water requirement correlation index between the grid units in the same vertical column, to construct a cleaning water requirement matrix;

[0024] S04. Calculating a cleaning time matrix based on the cleaning water requirement of each grid unit and the cleaning efficiency per unit time of the cleaning device;

[0025] S05. Construct a robot energy consumption model, consider the impact of the amount of water carried by the robot on its travel speed, and calculate the total energy consumption function under different paths;

[0026] S06. Applying a resource optimization allocation function to perform water allocation and path planning, wherein the input of the resource optimization allocation function includes the cleaning water demand matrix, the cleaning time matrix, the remaining water volume of the robot, the remaining battery power of the robot, and the photovoltaic panel layout information, and the output includes the optimal cleaning path sequence and the actual allocated water volume of each grid cell;

[0027] S07, integrating photovoltaic panel temperature sensor data and environmental parameters, inputting into a photovoltaic cleaning neural network model, and predicting the actual cleaning difficulty coefficient of each grid unit;

[0028] S08. Based on the output results of the photovoltaic cleaning neural network model, dynamically adjust the cleaning path and parameters to implement an adaptive cleaning strategy;

[0029] S09. Execute the final determination of the cleaning path, optionally including recording actual cleaning effect data, updating the stain distribution matrix, the cleaning water requirement matrix and the robot energy consumption model parameters for the next path planning optimization.

[0030] The stain distribution matrix refers to a two-dimensional matrix constructed according to the degree of stains on the surface of the photovoltaic panel. The element value of the stain distribution matrix ranges from 0 to 1, where 0 indicates no stains and 1 indicates the most serious stains. It is obtained through image processing technology.

[0031] Among them, the water demand correlation index refers to a coefficient that describes the degree of correlation between the stain cleaning between adjacent grid units in the same vertical column, and the value range is 0 to 1. When the water demand correlation index is 0, it means that the grid units are completely independent, and when the water demand correlation index is 1, it means that they are completely correlated.

[0032] The cleaning water requirement matrix is a two-dimensional matrix, and the calculation of the elements of the cleaning water requirement matrix takes into account the influence factors of the current degree of stain of the grid unit and the stain cleaning of the grid unit above.

[0033] Among them, the total energy consumption function takes into account the robot motor energy consumption, water pump working energy consumption and control system energy consumption, and is expressed as the sum of three parts, where the first part is proportional to the travel distance, the second part is proportional to the total water consumption, and the third part is a fixed value.

[0034] Among them, the resource optimization allocation function is used to achieve global optimization of cleaning paths and water distribution under limited resource constraints. The input of the resource optimization allocation function includes the cleaning water demand matrix, the cleaning time matrix, the remaining water volume of the robot, the remaining power of the robot and the photovoltaic panel layout information. The output of the resource optimization allocation function is the optimal cleaning path sequence and the actual allocated water volume of each grid unit.

[0035] Among them, the specific structure of the photovoltaic cleaning neural network model is a hybrid architecture combining a multi-layer convolutional neural network and an attention mechanism, wherein the convolution layer is responsible for extracting the stain distribution characteristics, the attention mechanism is used to capture the stain correlation between different grid units, and the fully connected layer outputs the cleaning difficulty prediction value. The photovoltaic cleaning neural network model adopts a dynamic attention threshold mechanism, and the dynamic attention threshold is jointly determined by the inclination angle of the photovoltaic panel, the current ambient temperature and the light intensity.

[0036] Among them, the steps of establishing a training data set in the photovoltaic cleaning neural network model training process specifically include collecting images of photovoltaic panel stains under different weather conditions in various types of photovoltaic power stations, recording the actual cleaning water consumption and cleaning time for each type of stain, establishing labeled data based on the stain distribution and actual cleaning effect, dividing the data into training set, validation set and test set in a ratio of 7:2:1, and performing data enhancement processing according to the characteristics of different regional environments to ensure the generalization ability of the photovoltaic cleaning neural network model.

[0037] Among them, the steps of training the photovoltaic cleaning neural network model specifically include first training basic visual feature extraction capabilities on multiple data sets, then performing transfer learning for the photovoltaic panel stain classification task, then introducing environmental factor input channels for joint training, and finally performing reinforcement learning optimization through actual cleaning robot feedback data. During the training process, a dynamic learning rate strategy and early stopping mechanism are used to prevent overfitting, and finally a model suitable for photovoltaic panel cleaning tasks under various environmental conditions is formed.

[0038] The specific implementation of the above steps is described in detail below.

[0039] The specific implementation method of step S01 is to establish a standardized grid division model based on the geometric dimensions and cleaning accuracy requirements of the photovoltaic panels. First, the physical dimensions of the photovoltaic panels of the photovoltaic power station are obtained. The size of the photovoltaic panels is usually 107×78 cm. The grid unit size is determined according to the cleaning accuracy requirements, with a typical value of 5×5 to 10×10 cm. Subsequently, each photovoltaic panel is divided into m×n grid units using a uniform division method, where m and n are integer values obtained by dividing the length and width of the photovoltaic panel by the grid unit size. Each grid unit is uniquely identified using a row-first encoding rule, and the numbering format is "board number-row number-column number" to ensure that each grid unit can be accurately located during the path planning process. The purpose of this step is to discretize the continuous photovoltaic panel surface into computable units, providing a basic data structure for subsequent stain analysis and path planning.

[0040] The specific implementation of step S02 is to use computer vision technology to analyze the distribution of stains on the surface of photovoltaic panels. First, a high-resolution camera array is deployed to ensure that all photovoltaic panel surfaces are covered, and the camera resolution is not less than 1080p, while collecting images of the photovoltaic panel surface. Then, image preprocessing is performed, including denoising, contrast enhancement and geometric correction, to eliminate the influence of uneven ambient lighting. Then, the YOLOv5 target detection algorithm is used to identify and segment each photovoltaic panel area to ensure the accurate positioning of each photovoltaic panel in the image. Subsequently, the improved HSV color space analysis algorithm is applied to extract stain features, and the stain area is identified by setting the stain color threshold (typical H channel threshold range is 30-45, S channel threshold range is 0.25-0.75, and V channel threshold range is 0.15-0.65). Then, the stain coverage and depth in each grid unit are calculated by grayscale statistical method, and the stain degree value (range 0-1) is obtained by normalization. Finally, the stain degree values of all grid units are organized into a two-dimensional matrix to construct a stain distribution matrix. The purpose of this step is to accurately quantify the surface contamination of the photovoltaic panels and provide a data basis for the subsequent calculation of cleaning water requirements.

[0041] The specific implementation method of step S03 is to construct a water demand calculation model based on the structural characteristics of the photovoltaic panel and the principles of water flow dynamics. First, a basic water demand mapping function is established to map the degree of contamination value to the theoretical water demand. The mapping function adopts a piecewise linear model. The degree of contamination of 0 to 0.3 corresponds to the low water demand range (0 to 100 ml / m2), 0.3 to 0.7 corresponds to the medium water demand range (100 to 250 ml / m2), and 0.7 to 1 corresponds to the high water demand range (250 to 400 ml / m2). Then, the water demand correlation index between the vertical grid cells is calculated, which is determined according to the inclination angle and surface roughness of the photovoltaic panel. The typical value is 0.6 when the photovoltaic panel inclination angle is 15°, 0.8 when the inclination angle is 30°, and 0.9 when the inclination angle is 45°. Next, a hydrodynamic model is used to correct the actual water demand for each grid cell, taking into account the pre-cleaning effect of water flow on the current grid cell while cleaning the grid cells above. This correction formula combines the stain distribution matrix and the water demand-related index to achieve a cumulative calculation of water demand from top to bottom. Finally, a cleaning water demand matrix is constructed, with the matrix elements representing the actual amount of water required for cleaning each grid cell. This step aims to optimize water resource allocation, avoid water waste, and ensure cleaning results.

[0042] The specific implementation method of step S04 is to determine the cleaning time in combination with the parameters of the cleaning device and the characteristics of the stain. First, the specification parameters of the cleaning device are obtained, including the nozzle type, the water spray pressure (typical value is 0.3-0.5MPa), the detergent concentration (typical value is 0.5-2%) and the brush head speed (typical value is 200-300 rpm). Then, a cleaning efficiency model is established to define the degree of stain that can be cleaned under a unit amount of water per unit time, with a typical value of 0.2 degrees of stain / (minute·liter). Then, in combination with the cleaning water requirement and the degree of stain of each grid unit, a nonlinear mapping function is applied to calculate the cleaning time. For lightly polluted areas (stain degree <0.3), the cleaning time is approximately linearly related to the water requirement; for moderately polluted areas (0.3≤stain degree <0.7), the cleaning time growth rate is accelerated; for heavily polluted areas (stain degree ≥0.7), the cleaning time increases exponentially. Finally, the calculation results are organized into a cleaning time matrix, and each element in the matrix corresponds to the time required to clean the grid unit. The purpose of this step is to provide time dimension constraints for path planning and optimize cleaning task scheduling.

[0043] The specific implementation of step S05 involves constructing a mathematical model that reflects the robot's energy consumption characteristics. First, a basic dynamic model of the robot is established, including mass parameters (typically 25-35 kg), friction coefficient (typically 0.2-0.3), and motor efficiency (typically 75-85%). The influence of water weight on travel speed is then measured. A full-water state (typically 10-15 liters) reduces travel speed by 15-20%, while a half-water state reduces travel speed by 7-10%. Next, a three-part energy consumption model is constructed: the motor's energy consumption is calculated as the distance multiplied by the energy consumption per unit distance (typically 12-18 watt-hours / meter) plus the additional load factor (adding 0.5-0.8 watt-hours / meter per liter of water); the pump's energy consumption is calculated as the water consumption multiplied by the energy consumption per unit water volume (typically 5-8 watt-hours / liter); and the control system's energy consumption is a fixed value (typically 5-8 watt-hours / hour). Finally, a total energy consumption function is constructed. The sum of these three energy consumption components, combined with variations in distance and water volume under different paths, forms the objective function for path optimization. The purpose of this step is to establish energy consumption evaluation standards and provide energy dimension constraints for path planning.

[0044] The specific implementation of step S06 is to apply a mixed integer linear programming algorithm to achieve optimal resource allocation. First, the decision variables are defined, including the path selection variable (a 0-1 variable indicating whether a certain path is selected) and the water allocation variable (a continuous variable indicating the amount of water allocated to each grid cell). Then, a set of constraints is constructed, including a total water constraint (the total amount of allocated water does not exceed the remaining water amount of the robot), a total power constraint (the total energy consumption does not exceed the remaining power of the robot), a minimum cleaning effect constraint (the amount of water allocated to each grid cell is not less than 80% of the minimum water requirement), and a path continuity constraint (ensuring that the path is continuous and non-repeated). Next, an objective function is constructed, using a weighted combination form, including three sub-goals: minimizing total energy consumption, minimizing total cleaning time, and maximizing cleaning effect. The weights are set according to actual application requirements, with typical values of 0.4, 0.3, and 0.3, respectively. Subsequently, an improved branch and bound algorithm is applied to solve the mixed integer programming problem. The algorithm parameter settings include a maximum number of iterations of 1000 times, a convergence threshold of 0.01, and a branching strategy using the maximum variable estimate. Finally, the optimal cleaning path sequence and the actual water allocation plan for each grid cell are output. The purpose of this step is to achieve global optimization under resource constraints and improve cleaning efficiency and resource utilization.

[0045] The specific implementation of step S07 is to predict the actual cleaning difficulty by combining multi-source sensor data. First, a sensor network is deployed to collect environmental parameters, including a photovoltaic panel surface temperature sensor array (accuracy ±0.5°C), an ambient temperature sensor (range -20 to 60°C), a relative humidity sensor (range 0 to 100%), and a light intensity sensor (range 0 to 2000W / m 2 ) and wind speed sensor (range 0-30m / s). The sensor data is then organized and combined with the stain distribution matrix to form a feature vector. The feature vector contains parameters such as stain degree, stain type, photovoltaic panel surface temperature, ambient temperature, relative humidity, light intensity and wind speed. Then the pre-trained photovoltaic cleaning neural network model is input for forward calculation, the spatial characteristics of the stain distribution are extracted through the convolution layer, the attention mechanism is applied to identify the key cleaning areas, and the actual cleaning difficulty coefficient of each grid unit is predicted through the fully connected layer in combination with the environmental parameters. Finally, the cleaning difficulty coefficient matrix is output for subsequent path adjustment. The purpose of this step is to enhance the model's adaptability to the actual cleaning environment and improve the accuracy of the cleaning strategy.

[0046] The specific implementation method of step S08 is to realize the adaptive cleaning strategy based on the reinforcement learning framework. First, the state space is constructed, including the current position of the robot, the remaining water volume, the remaining power, the distribution of cleaning difficulty of the completed cleaning area and the unfinished area. Then the action space is defined, including the selection of the moving direction and the adjustment of the water distribution. Then the reward function is designed. The reward function comprehensively considers the three factors of cleaning effect, resource consumption and cleaning time. The cleaning effect weight is 0.5, the resource consumption weight is 0.3, and the cleaning time weight is 0.2. The deep Q learning algorithm is then applied to calculate the optimal strategy. The algorithm parameters include a learning rate of 0.01, a discount factor of 0.9, an initial value of the exploration rate of 0.3, and linear decay to 0.05 with training. Finally, the cleaning path and parameters are dynamically adjusted according to the calculation results, including path replanning, water distribution adjustment, and cleaning time optimization. The purpose of this step is to achieve real-time optimization of the cleaning strategy to cope with environmental changes and uncertainties in the cleaning process.

[0047] The specific implementation method of step S09 is to execute the cleaning task and update the model. First, the actual cleaning operation is performed according to the final cleaning path. The robot moves according to the planned path and cleans each grid unit according to the allocation plan. Optionally, it also includes deploying a cleaning effect evaluation system, including post-cleaning image acquisition, cleaning effect scoring (range 0 to 100) and resource consumption records. Then calculate the error between the model prediction value and the actual observation value, including the cleaning difficulty prediction error, water demand prediction error and energy consumption prediction error. Subsequently, the stain distribution matrix, cleaning water demand matrix and robot energy consumption model parameters are updated, and the weighted moving average method is used to update the parameters, with the new value weight 0.3 and the historical value weight 0.7. Finally, the updated model parameters are stored in the system database for the next path planning optimization. The purpose of this step is to achieve continuous optimization and self-learning of the system and improve long-term operation efficiency.

[0048] The detailed structure of the photovoltaic cleaning neural network model consists of five main components: the input layer, feature extraction layer, attention layer, fully connected layer, and output layer. The input layer accepts three types of input: stain distribution features (a tensor of size m×n×3, containing stain coverage, depth, and type); photovoltaic panel parameters (including tilt angle, surface material, and aging); and environmental parameters (including temperature, humidity, light intensity, and wind speed). The feature extraction layer consists of three convolutional blocks, each of which contains two convolutional layers and one max pooling layer. The first convolutional block uses 32 3×3 convolutional kernels, the second convolutional block uses 64 3×3 convolutional kernels, and the third convolutional block uses 128 3×3 convolutional kernels. All convolutional layers use the ReLU activation function. The attention layer uses an improved spatial attention mechanism to calculate the correlation between stains across different grid cells. The attention weight is calculated using the scaled dot product method. The attention threshold is dynamically adjusted based on the photovoltaic panel tilt angle, current ambient temperature, and light intensity, with a typical threshold range of 0.15 to 0.35. The fully connected layer consists of three layers, with 256, 128, and 64 neurons, respectively. Dropout regularization is used to prevent overfitting, with a dropout rate of 0.3. The output layer uses a sigmoid activation function and outputs the cleaning difficulty coefficient of each grid cell, ranging from 0 to 1.

[0049] The detailed steps for establishing a training dataset for the PV cleaning neural network model include: First, 10 representative PV power plants were selected from diverse locations, encompassing desert, plain, coastal, and mountainous environments. Images of PV panel stains were then collected under varying weather conditions (sunny, cloudy, and after rain), with at least 100 images collected from each plant under each weather condition. The collected images were then annotated, including stain type (dust, bird droppings, leaves, etc.), stain distribution, and severity. Actual cleaning experiments were then conducted, recording actual water consumption and cleaning time for different stain types under varying environmental conditions. The image data was then linked to cleaning records to establish a correlation between input features and cleaning difficulty. The data was then randomly partitioned into training, validation, and test sets in a 7:2:1 ratio. Finally, data augmentation was performed, including image rotation (±15°), brightness adjustment (±20%), contrast variation (±15%), and the addition of Gaussian noise (σ = 0.02), expanding the training dataset to three times the original size and improving the model's generalization capabilities.

[0050] A second aspect of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores program instructions. When the program instructions are run in a computer, the program instructions are used to execute the above-mentioned photovoltaic power station cleaning robot path planning method.

[0051] The third aspect of the present invention provides a photovoltaic power station cleaning robot path planning system, which includes the above-mentioned computer-readable storage medium. The system is any one of a computer, a server, and a single-chip microcomputer. The computer-readable storage medium is arranged in the system, and the system is provided with a microprocessor for executing the program instructions stored in the computer-readable storage medium.

[0052] The mathematical model or calculation process involved in the present invention is described in detail below.

[0053] In step S01, the photovoltaic panel grid division model needs to specify the number and identification of grid cells. The photovoltaic panel is divided into m×n grid cells, and the calculation formula is:

[0054]

[0055] Where L is the length of the photovoltaic panel, in centimeters; W is the width of the photovoltaic panel, in centimeters; l is the length of the grid unit, in centimeters; w is the width of the grid unit, in centimeters; Represents the floor function.

[0056] A unique identifier for each grid cell is generated using the following formula:

[0057] ID i，j =P·10000+i·100+j;

[0058] Where, ID i，j is the grid unit identifier; P is the photovoltaic panel number; i is the row number, i∈[1,m]; j is the column number, j∈[1,n].

[0059] In step S02, the stain distribution matrix D m×n Constructed by image processing technology. For each grid cell (i, j), its stain degree d i，j The calculation is as follows:

[0060] d i，j =α·C i，j +β·I i，j +γ·T i，j +ε d ;

[0061] Where, d i，j is the degree of contamination of grid unit (i, j), and its value range is [0, 1]; C i，j is the stain coverage, which indicates the area ratio covered by stains in the grid unit; I i，j T is the stain intensity, reflecting the color depth of the stain; i，jis the stain type factor, and different types of stains have different values; α, β, and γ are weight coefficients, satisfying α+β+γ=1, and the typical values are 0.4, 0.4, and 0.2 respectively; ε d is the error term, ranging from [-0.05, 0.05].

[0062] Among them, the stain coverage C i，j Calculate after image segmentation:

[0063] C i，j =A 污渍 / A 单元 ;

[0064] Where A 污渍 is the stain coverage area within the grid unit; A 单元 is the total area of the grid cells.

[0065] Stain Intensity I i，j Calculated by the V channel value in the HSV color space:

[0066]

[0067] Where N is the number of pixels in the stained area; V k is the V channel value of the kth stain pixel, ranging from [0, 1].

[0068] Stain type factor T i，j Classify and assign values based on the HSV features of the stains:

[0069]

[0070] In step S03, the relationship between the water demand related index λ and the photovoltaic panel inclination angle θ can be expressed as:

[0071] λ=λ0+k λ sinθ+ε λ ;

[0072] Where, λ is the water demand correlation index, which ranges from [0, 1]; λ0 is the basic correlation coefficient, which is 0.2; k λ is the inclination angle influence coefficient, which is 0.8; θ is the inclination angle of the photovoltaic panel, in radians; ε λ is the error term, ranging from [-0.05, 0.05].

[0073] Cleaning water demand matrix W m×n The calculation takes into account the pollution distribution matrix and water flow. For grid cell (i, j), the water demand w i，j The calculation is as follows:

[0074]

[0075] Where w i，j is the cleaning water requirement of grid unit (i, j), in milliliters; w 基础 (d i，j ) is the basic water demand mapping function; λ is the water demand related index; μ is the water flow attenuation coefficient, with a value range of [0.1, 0.3]; ε w is the error term, ranging from [-5, 5] ml.

[0076] Basic water demand mapping function w 基础 (d) is defined as a piecewise function:

[0077]

[0078] Where k1, k2, and k3 are the slope coefficients of the three intervals, respectively. Typical values are k1 = 333.3 ml / m2, k2 = 375 ml / m2, and k3 = 500 ml / m2.

[0079] In step S04, the cleaning time matrix T m×n The calculation of is based on the cleaning water demand matrix and the stain distribution matrix. For grid cell (i, j), its cleaning time t i，j The calculation is as follows:

[0080]

[0081] Where, t i，j is the cleaning time of grid cell (i, j), in seconds; w i，j is the cleaning water requirement of grid unit (i, j), in milliliters; d i，j is the degree of contamination of the grid unit (i, j); α is the contamination influence coefficient, which is 2; β is the nonlinear index, which is 2.5; E c is the cleaning device efficiency parameter, the unit is ml / s, and the value range is [5, 10]; ε t is the error term, ranging from [-2, 2] seconds.

[0082] This formula reflects the relationship between cleaning time and water required for cleaning, while also taking into account the nonlinear effect of soiling severity on cleaning efficiency. The higher the soiling severity, the longer it takes to clean a unit of water, as stubborn stains require longer action to be removed.

[0083] In step S05, the robot's total energy consumption function E 总 It consists of three parts:

[0084] E 总 =E 移动 +E 水泵 +E 控制 +εE ;

[0085] Where, E 总 is the total energy consumption in watt-hours; E 移动 is the energy consumption of movement; E 水泵 E is the energy consumption of the water pump; 控制 is the energy consumption of the control system; E is the error term, ranging from [-5, 5] watt-hours.

[0086] Mobile energy consumption E 移动 The effects of path length and water load are considered:

[0087]

[0088] Where P is the total number of grid cells in the path; l p，p+1 is the distance from the pth grid cell to the p+1th grid cell, in meters; k d k is the energy consumption coefficient per unit distance, ranging from [12, 18] watt-hours / meter; w is the water load influence coefficient, and its value range is [0.05, 0.08] l -1 ; V p is the amount of water carried by the robot when it passes through the pth grid unit, in liters.

[0089] Water pump working energy consumption E 水泵 Proportional to total water consumption:

[0090]

[0091] Where k p is the energy consumption coefficient per unit water volume, and its value range is [5, 8] watt-hours / liter; is the actual amount of water distributed to the grid cell (i, j), in liters.

[0092] Control system energy consumption E 控制 It is the fixed power multiplied by the working time:

[0093]

[0094] Where, P 控制 is the system power, with a value range of [5, 8] watts; is the actual cleaning time of grid cell (i, j), in hours.

[0095] The amount of water carried by the robot V p As the cleaning process continues to decrease:

[0096]

[0097] Where V初始 The initial amount of water carried by the robot, in liters; is the actual amount of water allocated to the kth grid cell in the path, in liters.

[0098] In step S06, the resource optimization allocation function uses a mixed integer linear programming method. Introducing the decision variable x i，j，k and y i，j :

[0099]

[0100] y i，j Represents the actual amount of water allocated to grid cell (i, j) in milliliters.

[0101] Constraints include:

[0102] 1. Total water volume constraints:

[0103] 2. Total power constraint: E 总 (x, y)≤E 剩余 ;

[0104] 3. Minimum cleaning effect constraints:

[0105] 4. Path continuity constraints:

[0106]

[0107] Where V 剩余 is the remaining water volume of the robot, in liters; E 剩余 is the remaining power of the robot, in watt-hours; K is the number of grid cells contained in the path.

[0108] The objective function F is the weighted sum of three sub-objectives:

[0109]

[0110] Where α1, α2 and α3 are weight coefficients, with typical values of 0.4, 0.3 and 0.3 respectively; E 总 (x, y) is the total energy consumption function; T 总 (x, y) is the total cleaning time function; Q 总 (x, y) is the total cleaning effect function; E 标准 、T 标准 and Q 标准 is the corresponding standardized coefficient.

[0111] Total cleaning time function T 总 (x, y) is calculated as follows:

[0112]

[0113] Where, t 移动 (p k , p k+1 ) is the path from point p k to p k+1 moving time.

[0114] Total cleaning effect function Q 总 (x, y) is calculated as follows:

[0115]

[0116] In the formula, q(d i，j ,y i，j ) is the cleaning effect function of the grid unit (i, j):

[0117]

[0118] The principle of using nonlinear functional relationships in the above equations is to more accurately reflect the physical phenomena in the actual cleaning process. For example, the power function (d i，j ) β , because the degree of staining has a nonlinear relationship with the difficulty of cleaning. The higher the degree of staining, the faster the difficulty of cleaning increases. The sine function sinθ is used in the calculation of the water demand related index because the flow of water along the surface of the photovoltaic panel is affected by gravity and is proportional to the sine value of the photovoltaic panel inclination angle. The exponential decay function e is used in the calculation of water flow effect. -μ(i-k) , because the influence of the water flow generated by the cleaning of the upper grid cell on the lower grid cell decays exponentially with the increase of distance.

[0119] Specifically, the principle of the present invention is: the technical principle of the present invention is based on the organic combination of the identification of the distribution characteristics of stains on the surface of photovoltaic panels and intelligent path planning, which is mainly reflected in the following aspects:

[0120] First, the present invention uses grid division and image processing techniques to establish an accurate contamination distribution matrix. Each matrix element corresponds to the degree of contamination of a grid cell, with values ranging from 0 to 1. This quantitatively characterizes the distribution of contamination on the photovoltaic panel surface. This method overcomes the limitation of traditional technologies that cannot perceive contamination distribution and provides a data foundation for intelligent path planning. Unlike simple binary processing, this invention uses continuous values to describe contamination degree, which can more accurately reflect the actual contamination situation.

[0121] Secondly, the present invention innovatively introduces a water demand correlation index between grid cells in the same vertical column, scientifically describing the vertical influence of water flow. This aligns with the physical characteristics of photovoltaic panel cleaning: during the cleaning of the upper grid cells, water flow will flush the area below. By establishing a cleaning water demand matrix, the system can comprehensively consider the degree of contamination of the current grid cell and the impact of cleaning the grid cells above, avoiding the waste of resources caused by a simple linear path.

[0122] Finally, the present invention uses a photovoltaic cleaning neural network model that combines a multi-layer convolutional neural network with an attention mechanism. This model integrates photovoltaic panel temperature sensor data with environmental parameters to predict the actual cleaning difficulty coefficient for each grid cell. This model can capture the correlation between stains in different grid cells and adapt to different environmental conditions through a dynamic attention threshold mechanism, making path planning more consistent with actual cleaning needs.

[0123] Finally, the present invention constructs a complete resource optimization allocation function, incorporating multidimensional factors such as stain distribution characteristics, cleaning requirements, and robot resource constraints into the decision-making process, and calculates the optimal cleaning path sequence, realizing the transition from simple fixed paths to intelligent and personalized path planning, greatly improving cleaning efficiency and resource utilization.

[0124] By integrating and applying the above principles, the present invention realizes intelligent path planning based on the stain distribution characteristics, solving the technical problem in the prior art that photovoltaic power station cleaning robots cannot perform intelligent path planning based on the stain distribution characteristics on the photovoltaic panel surface.

[0125] 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.

[0126] The specific implementation of step S01 is to establish a standardized grid division model based on the geometric dimensions and cleaning accuracy requirements of the photovoltaic panels. First, the physical dimensions of the photovoltaic panels in the photovoltaic power station are obtained. The typical size of the photovoltaic panels is 107 × 78 cm. The grid unit size is determined based on the cleaning accuracy requirements, with a typical value of 10 × 10 cm. Then, each photovoltaic panel is divided into m × n grid units using a uniform division method, where m and n are calculated using the following formulas:

[0127]

[0128]

[0129] Where L is the length of the photovoltaic panel, in centimeters; W is the width of the photovoltaic panel, in centimeters; l is the length of the grid unit, in centimeters; w is the width of the grid unit, in centimeters; Represents the floor function. Row-first encoding is used to uniquely identify each grid cell, and the number is generated using the following formula:

[0130] ID i，j =P·10000+i·100+j;

[0131] Where, ID i，j is the grid cell identifier; P is the photovoltaic panel number; i is the row number, i∈[1,m]; j is the column number, j∈[1,n]. The purpose of this step is to discretize the continuous photovoltaic panel surface into computable units, providing the basic data structure for subsequent contamination analysis and path planning.

[0132] The specific implementation of step S02 is to use computer vision technology to analyze the distribution of stains on the surface of photovoltaic panels. First, a high-resolution camera array is deployed to ensure that all photovoltaic panel surfaces are covered, and the camera resolution is not less than 1080p, while collecting images of the photovoltaic panel surface. Then, image preprocessing is performed, including denoising, contrast enhancement, and geometric correction, to eliminate the influence of uneven ambient lighting. The YOLOv5 target detection algorithm is then used to identify and segment each photovoltaic panel area to ensure accurate positioning of each photovoltaic panel in the image. Subsequently, the improved HSV color space analysis algorithm is applied to extract stain features, and the stain area is identified by setting the stain color threshold (the typical H channel threshold range is 30-45, the S channel threshold range is 0.25-0.75, and the V channel threshold range is 0.15-0.65). Based on the image processing results, the stain distribution matrix D is constructed. m×n , for each grid cell (i, j), its degree of contamination d i，j The calculation is as follows:

[0133] d i，j =α·C i，j +β·I i，j +γ·T i，j +ε d ;

[0134] Where, d i，j is the degree of contamination of grid unit (i, j), and its value range is [0, 1]; C i，j is the stain coverage, which indicates the area ratio covered by stains in the grid unit; I i，j T is the stain intensity, reflecting the color depth of the stain; i，j is the stain type factor, and different types of stains have different values; α, β, and γ are weight coefficients, satisfying α+β+γ=1, and the typical values are 0.4, 0.4, and 0.2 respectively; ε d is the error term, ranging from [-0.05, 0.05]. i，j Calculate after image segmentation:

[0135] C i，j =A 污渍 / A 单元 ;

[0136] Stain Intensity I i，j Calculated by the V channel value in the HSV color space:

[0137]

[0138] Stain type factor T i，j Classify and assign values based on the HSV features of the stains:

[0139]

[0140] The purpose of this step is to accurately quantify the surface contamination of the photovoltaic panels and provide a data basis for the subsequent calculation of cleaning water requirements.

[0141] The specific implementation of step S03 is to build a water demand calculation model based on the structural characteristics of photovoltaic panels and the principles of water flow dynamics. First, a basic water demand mapping function is established to map the contamination level value to the theoretical water demand. The mapping function adopts a piecewise linear model:

[0142]

[0143] Where k1, k2, and k3 are the slope coefficients for the three intervals, with typical values of k1 = 333.3 ml / m2, k2 = 375 ml / m2, and k3 = 500 ml / m2. The water demand correlation index λ between the vertical grid cells is then calculated, determined based on the inclination angle and surface roughness of the photovoltaic panel:

[0144] λ=λ0+k λ sinθ+ε λ ;

[0145] Where, λ is the water demand correlation index, which ranges from [0, 1]; λ0 is the basic correlation coefficient, which is 0.2; k λ is the inclination angle influence coefficient, which is 0.8; θ is the inclination angle of the photovoltaic panel, in radians; ε λ is the error term, ranging from [-0.05, 0.05]. Then the water flow dynamics model is used to correct the actual water demand of each grid unit and construct the clean water demand matrix W m×n , for grid cell (i, j), its water demand w i，j The calculation is as follows:

[0146]

[0147] Where w i，j is the cleaning water requirement of grid unit (i, j), in milliliters; w基础 (d i，j ) is the basic water demand mapping function; λ is the water demand related index; μ is the water flow attenuation coefficient, with a value range of [0.1, 0.3]; ε w The error term is [-5, 5] ml. The purpose of this step is to optimize water resource allocation, avoid water waste, and ensure cleaning results.

[0148] The specific implementation method of step S04 is to determine the cleaning time in combination with the cleaning device parameters and the stain characteristics. First, obtain the specification parameters of the cleaning device, including the nozzle type, water spray pressure (typical value is 0.3-0.5MPa), detergent concentration (typical value is 0.5-2%) and brush head speed (typical value is 200-300 rpm). Then establish a cleaning efficiency model to define the degree of stain that can be cleaned per unit water volume per unit time, with a typical value of 0.2 degrees of stain / (minute·liter). Then, combine the cleaning water requirement and stain degree of each grid unit to calculate the cleaning time matrix T m×m , for grid cell (i, j), its cleaning time is t i，j The calculation is as follows:

[0149]

[0150] Where, t i，j is the cleaning time of grid cell (i, j), in seconds; w i，j is the cleaning water requirement of grid unit (i, j), in milliliters; d i，j is the degree of contamination of the grid unit (i, j); α is the contamination influence coefficient, which is 2; β is the nonlinear index, which is 2.5; E c is the cleaning device efficiency parameter, the unit is ml / s, and the value range is [5, 10]; ε t is the error term, ranging from [-2, 2] seconds. The purpose of this step is to provide time dimension constraints for path planning and optimize cleaning task scheduling.

[0151] The specific implementation of step S05 is to construct a mathematical model that reflects the energy consumption characteristics of the robot. First, a basic dynamic model of the robot is established, including mass parameters (typical values are 25 to 35 kilograms), friction coefficient (typical values are 0.2 to 0.3), and motor efficiency (typical values are 75 to 85%). Then the influence coefficient of water weight on the travel speed is measured. In the full water state (typical capacity is 10 to 15 liters), the travel speed is reduced by 15 to 20%, and in the half water state, the travel speed is reduced by 7 to 10%. Then a three-part energy consumption model is constructed, and the total energy consumption function E of the robot is: 总 It consists of three parts:

[0152] E 总 =E 移动 +E水泵 +E 控制 +ε E ;

[0153] Where, E 总 is the total energy consumption in watt-hours; E 移动 is the energy consumption of movement; E 水泵 E is the energy consumption of the water pump; 控制 is the energy consumption of the control system; E is the error term, ranging from [-5, 5] watt-hours.

[0154] Mobile energy consumption E 移动 The effects of path length and water load are considered:

[0155]

[0156] Where P is the total number of grid cells in the path; l p，p+1 is the distance from the pth grid cell to the p+1th grid cell, in meters; k d k is the energy consumption coefficient per unit distance, ranging from [12, 18] watt-hours / meter; w is the water load influence coefficient, and its value range is [0.05, 0.08] l -1 ; V p is the amount of water carried by the robot when it passes through the pth grid unit, in liters.

[0157] Water pump working energy consumption E 水泵 Proportional to total water consumption:

[0158]

[0159] Where k p is the energy consumption coefficient per unit water volume, and its value range is [5, 8] watt-hours / liter; is the actual amount of water distributed to the grid cell (i, j), in liters.

[0160] Control system energy consumption E 控制 It is the fixed power multiplied by the working time:

[0161]

[0162] Where, P 控制 is the system power, with a value range of [5, 8] watts; is the actual cleaning time of grid cell (i, j), in hours.

[0163] The amount of water carried by the robot V p As the cleaning process continues to decrease:

[0164]

[0165] Where V 初始 The initial amount of water carried by the robot, in liters; The actual amount of water allocated to the kth grid cell in the path, in liters. The purpose of this step is to establish an energy consumption evaluation standard and provide energy dimension constraints for path planning.

[0166] The specific implementation of step S06 is to use the mixed integer linear programming algorithm to achieve resource optimization allocation. First, define the decision variable x i，j，k and y i，j :

[0167]

[0168] y i，j Represents the actual amount of water allocated to grid cell (i, j) in milliliters.

[0169] Then build a set of constraints, including:

[0170] 1. Total water volume constraints:

[0171] 2. Total power constraint: E 总 (x, y)≤E 剩余 ;

[0172] 3. Minimum cleaning effect constraints:

[0173] 4. Path continuity constraints:

[0174] Then construct the objective function F as the weighted sum of the three sub-goals:

[0175]

[0176] Where α1, α2 and α3 are weight coefficients, with typical values of 0.4, 0.3 and 0.3 respectively; E 总 (x, y) is the total energy consumption function; T 总 (x, y) is the total cleaning time function; Q 总 (x, y) is the total cleaning effect function; E 标准 、T 标准 and Q 标准 is the corresponding standardized coefficient.

[0177] Total cleaning time function T 总 (x, y) is calculated as follows:

[0178]

[0179] Where, t 移动 (p k , p k+1 ) is the path from point p k to p k+1 moving time.

[0180] Total cleaning effect function Q 总 (x, y) is calculated as follows:

[0181] In the formula, q(d i，j ,y i，j ) is the cleaning effect function of the grid unit (i, j):

[0182] An improved branch-and-bound algorithm was then applied to solve the mixed integer programming problem. Algorithm parameters included a maximum number of iterations of 1000, a convergence threshold of 0.01, and a branching strategy using the maximum variable estimate. Ultimately, the optimal cleaning path sequence and the actual water allocation for each grid cell were output. This step aims to achieve global optimization within resource constraints, improving cleaning efficiency and resource utilization.

[0183] The specific implementation of step S07 is to predict the actual cleaning difficulty by combining multi-source sensor data. First, a sensor network is deployed to collect environmental parameters, including a photovoltaic panel surface temperature sensor array (accuracy ±0.5°C), an ambient temperature sensor (range -20 to 60°C), a relative humidity sensor (range 0 to 100%), and a light intensity sensor (range 0 to 2000W / m 2 ) and a wind speed sensor (range 0-30 m / s). The sensor data is then organized and combined with a dirt distribution matrix to form a feature vector. This feature vector includes parameters such as dirt level, dirt type, PV panel surface temperature, ambient temperature, relative humidity, light intensity, and wind speed. This feature vector is then fed into a pre-trained PV cleaning neural network model for forward computation. The convolutional layer extracts the spatial characteristics of the dirt distribution, an attention mechanism is applied to identify key cleaning areas, and a fully connected layer, combined with environmental parameters, predicts the actual cleaning difficulty coefficient for each grid cell. The PV cleaning neural network model utilizes a hybrid architecture combining a multi-layer convolutional neural network with an attention mechanism. The convolutional layer extracts the dirt distribution characteristics, the attention mechanism captures the correlation between dirt across different grid cells, and the fully connected layer outputs a predicted cleaning difficulty value. The attention mechanism employs a dynamic attention threshold mechanism, which is determined by the PV panel tilt angle, current ambient temperature, and light intensity. Finally, a cleaning difficulty coefficient matrix is output for subsequent path adjustments. This step aims to enhance the model's adaptability to actual cleaning environments and improve the accuracy of cleaning strategies.

[0184] The specific implementation of step S08 is to realize the adaptive cleaning strategy based on the reinforcement learning framework. First, the state space is constructed, including the current position of the robot, the remaining water volume, the remaining power, the cleaning difficulty distribution of the completed cleaning area and the unfinished area. Then the action space is defined, including the selection of the movement direction and the adjustment of the water volume distribution. Then the reward function is designed. The reward function comprehensively considers the three factors of cleaning effect, resource consumption and cleaning time. The cleaning effect weight is 0.5, the resource consumption weight is 0.3, and the cleaning time weight is 0.2. The reward function R can be expressed as:

[0185]

[0186] Where R(s, a) is the reward obtained by executing action a in state s; Q 效果 (s, a) is the cleaning effect; C 资源 (s, a) is resource consumption; T 时间 (s, a) is the cleaning time; Q 最大 、C 最大 and T 最大 is the corresponding normalization coefficient. A deep Q-learning algorithm is then applied to calculate the optimal strategy. Algorithm parameters include a learning rate of 0.01, a discount factor of 0.9, and an initial exploration rate of 0.3, which decays linearly to 0.05 during training. Finally, the cleaning path and parameters are dynamically adjusted based on the calculated results, including path replanning, water allocation adjustment, and cleaning duration optimization. This step aims to achieve real-time optimization of the cleaning strategy to address environmental changes and uncertainties in the cleaning process.

[0187] The specific implementation method of step S09 is to execute the cleaning task and update the model. First, the actual cleaning operation is performed according to the final cleaning path. The robot moves according to the planned path and cleans each grid unit according to the allocation plan. Then the cleaning effect evaluation system is deployed, including post-cleaning image acquisition, cleaning effect scoring (range 0 to 100) and resource consumption records. Then calculate the error between the model prediction value and the actual observation value, including the cleaning difficulty prediction error, water demand prediction error and energy consumption prediction error. The error calculation formula is as follows:

[0188]

[0189] Where, e 难度 、e 水量 and e 能耗 are the prediction errors of cleaning difficulty, water demand, and energy consumption, respectively; and are the predicted and actual soiling levels, respectively; and are the predicted and actual water demands, respectively; and are the predicted and actual total energy consumption, respectively.

[0190] Then, the stain distribution matrix, cleaning water demand matrix and robot energy consumption model parameters are updated, and the weighted moving average method is used to update the parameters:

[0191] θ 新 =(1-α)·θ 旧 +α·θ 观测 ;

[0192] Where θ 新 and θ 旧 are the model parameters before and after updating respectively; θ 观测 is the current observation value; α is the update weight, typically 0.3. Finally, the updated model parameters are stored in the system database for use in the next path planning optimization. This step aims to achieve continuous optimization and self-learning of the system, improving long-term operational efficiency.

[0193] To better understand and implement the present invention, Example 2, a specific application scenario, is provided below. Researchers conducted a practical application at a photovoltaic power station covering approximately 5 square kilometers, with an installed capacity of 50 MW and 16,000 photovoltaic panels. The region's environmental characteristics include strong winds and sandstorms, large temperature swings between day and night, and the primary source of pollution from the photovoltaic panels, dust. This pollution is also uneven, placing high demands on cleaning path planning.

[0194] First, the photovoltaic panels are gridded. The photovoltaic panels used in this power plant measure 166 x 83 cm. Based on the cleaning precision requirements and the distribution of surface stains, each panel is divided into 10 x 5 grid cells, each measuring 16.6 x 16.6 cm. Each grid cell is assigned a unique identifier using the row-first numbering format of "panel number-row number-column number." For example, the grid cell in row 3, column 4 on photovoltaic panel 125 is identified as "125-03-04."

[0195] Then, a high-resolution camera array is used to obtain the distribution of stains on the surface of the photovoltaic panel. Taking photovoltaic panel No. 125 as an example, the distribution of stains collected is shown in Table 1:

[0196] Table 1. Distribution matrix of stains on photovoltaic panel No. 125

[0197] Row / Column 1 column 2 columns 3 columns 4 columns 5 columns 1 row 0.35 0.42 0.56 0.63 0.72 2 rows 0.33 0.38 0.45 0.58 0.65 3 rows 0.28 0.32 0.39 0.47 0.55 4 rows 0.25 0.27 0.35 0.42 0.48 5 lines 0.22 0.24 0.31 0.38 0.43 6 lines 0.19 0.21 0.25 0.33 0.38 7 lines 0.15 0.18 0.22 0.28 0.32 8 lines 0.12 0.15 0.18 0.24 0.28 9 lines 0.08 0.12 0.15 0.20 0.25 10 lines 0.05 0.08 0.11 0.16 0.21

[0198] The photovoltaic panels exhibit a typical distribution pattern of gradually decreasing dirt levels from top to bottom and from right to left, consistent with the prevailing wind direction and dust deposition patterns in the area. Image analysis shows that the dirt is primarily of the dust type, with a dirt type factor of 0.3.

[0199] The water demand correlation index was then calculated. The photovoltaic panels in this power station were installed at a 35° inclination angle. The water demand correlation index, λ, was calculated to be 0.69, indicating that cleaning the upper grid cells has a strong impact on the lower grid cells.

[0200] Based on the stain distribution matrix and water demand correlation index, the cleaning water demand of each grid unit is calculated, as shown in Table 2:

[0201] Table 2 Matrix of water requirements for cleaning of photovoltaic panels No. 125 (unit: ml)

[0202] Row / Column 1 column 2 columns 3 columns 4 columns 5 columns 1 row 116.7 140.0 186.7 220.0 256.7 2 rows 105.2 122.1 150.0 193.3 227.5 3 rows 83.3 98.1 123.8 156.7 183.3 4 rows 72.2 78.8 106.3 132.5 160.0 5 lines 61.9 67.6 90.8 117.0 138.3 6 lines 51.7 57.3 70.2 98.8 115.7 7 lines 39.5 47.6 58.3 79.2 93.8 8 lines 30.4 38.3 46.2 65.2 78.3 9 lines 19.5 29.0 37.1 52.1 67.9 10 lines 11.7 18.7 26.0 39.0 54.8

[0203] It can be seen that the amount of water required for cleaning decreases from top to bottom, and the decrease is greater than the decrease in the degree of staining, which reflects the cumulative effect of water flow.

[0204] The cleaning time matrix was calculated based on the cleaning device parameters. The cleaning device efficiency parameter was 7.5 ml / s, and the dust pollution nonlinear index was 2.0. After calculation, the cleaning time matrix shown in Table 3 was obtained:

[0205] Table 3 Matrix of cleaning time of photovoltaic panel No. 125 (unit: seconds)

[0206] Row / Column 1 column 2 columns 3 columns 4 columns 5 columns 1 row 20.9 26.6 38.2 47.8 59.7 2 rows 18.5 22.5 29.3 41.4 50.9 3 rows 14.0 16.9 22.5 31.0 39.2 4 rows 11.7 13.0 18.5 24.8 31.7 5 lines 9.7 10.8 15.5 21.1 26.4 6 lines 7.9 8.8 11.5 17.2 21.3 7 lines 5.9 7.1 9.3 13.5 16.4 8 lines 4.4 5.6 7.1 10.6 13.3 9 lines 2.8 4.2 5.5 8.3 11.2 10 lines 1.6 2.6 3.8 6.0 8.7

[0207] The cleaning robot model is GC-R35, and its basic parameters include: empty mass 28kg, water tank capacity 12L, battery capacity 350Wh, moving speed 0.3m / s, water pump rated power 60W, control system power 6W. According to the energy consumption model, the energy consumption coefficient per unit distance is 15Wh / m, and the water load influence coefficient is 0.06L. -1 The energy consumption coefficient per unit water volume is 6.5Wh / L.

[0208] During the resource optimization phase, the system inputs included the cleaning water demand matrix, the cleaning time matrix, the robot's remaining water level (12L), remaining battery level (330Wh), and photovoltaic panel layout information. A mixed-integer linear programming algorithm was used for calculations, with objective function weights set at 0.4 for minimizing energy consumption, 0.3 for minimizing cleaning time, and 0.3 for maximizing cleaning effectiveness. After 483 iterations, the optimization algorithm determined the optimal cleaning path and water allocation scheme, as shown in Table 4.

[0209] Table 4 Optimal cleaning path sequence for photovoltaic panel No. 125

[0210] Serial number Grid Cell Dispensing water volume (ml) Cleaning time (seconds) 1 125-01-05 256.7 59.7 2 125-01-04 220.0 47.8 3 125-01-03 186.7 38.2 4 125-01-02 140.0 26.6 5 125-01-01 116.7 20.9 6 125-02-01 84.2 14.8 7 125-02-02 97.7 18.0 8 125-02-03 120.0 23.4 … … … … 50 125-10-05 43.8 7.0

[0211] It is worth noting that this path is not a simple serpentine or spiral, but an irregular path that takes into account the influence of water flow. Cleaning starts from the heavily polluted areas above, using water flow to assist in pre-cleaning the areas below, reducing overall water consumption.

[0212] Combined with environmental monitoring data, including photovoltaic panel surface temperature (45°C), ambient temperature (38°C), relative humidity (15%), light intensity (950W / m 2 ) and wind speed (4.5 m / s) were input into the photovoltaic cleaning neural network model to predict the actual cleaning difficulty coefficient. The model output showed that due to the high temperature and low humidity environment, the actual cleaning difficulty coefficient increased by 15% compared to standard conditions. The system automatically increased water allocation to the heavily polluted areas above by 12% to ensure cleaning results.

[0213] During the cleaning process, the reinforcement learning algorithm continuously adjusts its strategy based on real-time cleaning results. For example, if it discovers that the actual cleaning difficulty of the grid cell in row 3 and column 4 is higher than expected, it immediately recalculates the path and water allocation for the remaining grid cells, reusing the saved water to the more challenging area.

[0214] After cleaning was completed, the evaluation system recorded the actual cleaning results. The average cleaning rate for the entire board reached 96.8%, higher than the 92.3% of the conventional fixed-path solution. Total water consumption was saved by 17.5% compared to the traditional method. Cleaning time was reduced by 12.3%, and energy consumption was reduced by 15.2%.

[0215] Traditional photovoltaic panel cleaning path planning typically uses a simple serpentine or spiral fixed path, with a uniformly set water distribution standard, without considering the uneven distribution of stains, the impact of water flow, and the impact of environmental factors on the cleaning effect. This method is inefficient in desert areas with limited water resources, wasting water resources and failing to guarantee cleaning results. The present invention significantly improves cleaning efficiency and resource utilization by constructing a sophisticated stain distribution matrix, considering the impact of water flow dynamics, introducing a robot energy consumption model, applying mixed integer linear programming for resource optimization, and combining neural network model prediction and reinforcement learning to achieve adaptive adjustment. Experimental data shows that under the same resource conditions, this method can increase the cleaning rate by 4.5 percentage points and save approximately 18% of water resources, which is of great significance for photovoltaic power stations in desert areas where water resources are scarce. In addition, by reducing cleaning time and energy consumption, the robot's working time and service life can be extended, reducing operation and maintenance costs.

[0216] It should be noted that the variables involved in the present invention are explained in detail as shown in Tables 5 and 6 below.

[0217] Table 5 Variable Explanation Table (Part 1)

[0218]

[0219] Table 6 Variable Explanation Table (Part 2)

[0220]

[0221]

[0222] 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 path planning method for a photovoltaic power station cleaning robot, characterized in that: include: Establish a photovoltaic panel grid division model, divide each photovoltaic panel into m×n grid units, and number each grid unit; The degree of contamination of each grid cell is obtained through image processing technology, and a contamination distribution matrix is constructed. The cleaning water requirement of each grid cell is calculated based on the contamination distribution matrix, and the water requirement correlation index between grid cells in the same vertical column is considered to construct a cleaning water requirement matrix. Calculate the cleaning time matrix based on the cleaning water requirement of each grid unit and the cleaning efficiency per unit time of the cleaning device; Build a robot energy consumption model; apply resource optimization allocation functions for water distribution and path planning; integrate photovoltaic panel temperature sensor data and environmental parameters, input them into the photovoltaic cleaning neural network model, and predict the actual cleaning difficulty coefficient of each grid cell; Based on the output results of the photovoltaic cleaning neural network model, the cleaning path and parameters are dynamically adjusted; and the final cleaning path is executed.

2. The photovoltaic power station cleaning robot path planning method according to claim 1, characterized in that: The stain distribution matrix refers to a two-dimensional matrix constructed according to the degree of stains on the surface of the photovoltaic panel. The element value of the stain distribution matrix ranges from 0 to 1, where 0 indicates no stains and 1 indicates the most serious stains. It is obtained through image processing technology.

3. The path planning method for a photovoltaic power station cleaning robot according to claim 2, characterized in that: The water demand correlation index refers to the coefficient that describes the degree of correlation between the stain cleaning of adjacent grid cells in the same vertical column. The value range is 0 to 1. When the water demand correlation index is 0, it means that the grid cells are completely independent, and when the water demand correlation index is 1, it means that they are completely correlated. The cleaning water demand matrix is a two-dimensional matrix. The calculation of the elements of the cleaning water demand matrix takes into account the influencing factors of the stain degree of the current grid cell and the stain cleaning of the grid cell above.

4. The photovoltaic power station cleaning robot path planning method according to claim 3, characterized in that: The total energy consumption function takes into account the energy consumption of the robot motor, the water pump working energy consumption and the control system energy consumption, and is expressed as the sum of three parts, where the first part is proportional to the travel distance, the second part is proportional to the total water consumption, and the third part is a fixed value.

5. The photovoltaic power station cleaning robot path planning method according to claim 4, characterized in that: The resource optimization allocation function is used to achieve global optimization of cleaning paths and water distribution under limited resource constraints. The input of the resource optimization allocation function includes the cleaning water demand matrix, the cleaning time matrix, the remaining water volume of the robot, the remaining battery power of the robot, and the photovoltaic panel layout information. The output of the resource optimization allocation function is the optimal cleaning path sequence and the actual water distribution of each grid unit.

6. The photovoltaic power station cleaning robot path planning method according to claim 5, characterized in that: The specific structure of the photovoltaic cleaning neural network model is a hybrid architecture that combines a multi-layer convolutional neural network with an attention mechanism. The convolution layer is responsible for extracting the stain distribution characteristics, the attention mechanism is used to capture the stain correlation between different grid cells, and the fully connected layer outputs the cleaning difficulty prediction value; the photovoltaic cleaning neural network model adopts a dynamic attention threshold mechanism, and the dynamic attention threshold is jointly determined by the inclination angle of the photovoltaic panel, the current ambient temperature and the light intensity.

7. The photovoltaic power station cleaning robot path planning method according to claim 6, characterized in that: The steps of establishing a training data set in the photovoltaic cleaning neural network model training process specifically include collecting images of photovoltaic panel stains under different weather conditions in various types of photovoltaic power stations, recording the actual cleaning water consumption and cleaning time for each type of stain, and establishing labeled data based on the stain distribution and actual cleaning effect.

8. The photovoltaic power station cleaning robot path planning method according to claim 7, characterized in that: The training steps for the photovoltaic cleaning neural network model include first training basic visual feature extraction capabilities on multiple data sets, then performing transfer learning on the photovoltaic panel stain classification task, then introducing environmental factor input channels for joint training, and finally performing reinforcement learning optimization through actual cleaning robot feedback data.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium stores program instructions, and when the program instructions are run in a computer, they are used to execute the photovoltaic power station cleaning robot path planning method according to any one of claims 1 to 8.

10. A photovoltaic power station cleaning robot path planning system, characterized in that: The computer-readable storage medium according to claim 9 is included, the system is any one of a computer, a server, and a single-chip microcomputer, the computer-readable storage medium is arranged in the system, and the system is provided with a microprocessor for executing program instructions stored in the computer-readable storage medium.

Images