Self-adaptive intelligent cleaning control system and method for photovoltaic panel

Abstract

The invention discloses a self-adaptive intelligent cleaning control system and method for a photovoltaic panel, and relates to the technical field of photovoltaic power generation, and the system comprises an image collection module which is used for dynamically collecting visible light and near-infrared images of the photovoltaic panel; the grade division module is used for realizing self-adaptive division of dust pollution grades through image analysis; the cleaning decision module analyzes and judges whether cleaning is executed or not based on the power generation income and the cleaning cost; and the parameter regulation and control module generates control parameters such as mechanical arm speed and water pressure through deep reinforcement learning. Intelligent and precise control over photovoltaic panel cleaning is achieved, the cleaning efficiency is improved, and the operation and maintenance cost is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic power generation, and in particular to a photovoltaic panel adaptive intelligent cleaning control system and method. Background Art

[0002] Dust accumulation on the surface of photovoltaic panels can significantly reduce their power generation efficiency. Existing intelligent cleaning technologies have the following drawbacks in practical applications:

[0003] Existing detection methods often rely on single visible light cameras, which can only capture surface color and texture information. They lack the ability to identify deeper features such as dust thickness. This is especially true in scenarios where the color of dust is similar to that of photovoltaic panels, which can easily lead to missed detections or misjudgments. Furthermore, dust pollution levels are often classified based on fixed thresholds, without dynamically adjusting multi-dimensional features such as coverage and grayscale differences. This makes it difficult to adapt to differences in dust distribution in different environments, resulting in a mismatch between classification and actual pollution levels.

[0004] Existing methods often use fixed cycles for cleaning decisions, without considering real-time electricity prices, cleaning costs and other economic factors. This can lead to excessive cleaning or delayed cleaning, increasing operation and maintenance costs.

[0005] The control of cleaning execution parameters lacks an adaptive mechanism and relies heavily on manual experience or preset parameters. It cannot be optimized in real time based on pollution levels and equipment status, resulting in low cleaning efficiency.

[0006] Therefore, there is an urgent need for an adaptive intelligent cleaning control system and method to achieve precise and intelligent cleaning of photovoltaic panels. Summary of the Invention

[0007] In response to the deficiencies in the prior art, the present invention provides a photovoltaic panel adaptive intelligent cleaning control system and method, which solves the problems of unintelligent detection, decision-making and parameter regulation in photovoltaic panel cleaning.

[0008] To achieve the above objectives, the present invention is implemented through the following technical solutions: a photovoltaic panel adaptive intelligent cleaning control system, comprising:

[0009] The image acquisition module uses dual cameras to collect visible light and near-infrared images of the photovoltaic panel at a frequency f within a dynamic time window;

[0010] The grading module averages multiple visible light images to obtain an average visible light image, and averages multiple near-infrared images to obtain an average near-infrared image. The average visible light image is used to calculate the coverage R, and the average absolute grayscale difference ΔG is used to calculate the average near-infrared image. A dynamic feature space mapping method is constructed to achieve adaptive classification of dust pollution levels.

[0011] The cleaning decision module calculates power generation benefits Sg and cleaning cost Sr of the photovoltaic panel group under different pollution levels, and determines whether to perform the cleaning operation;

[0012] The parameter regulation module constructs a track type mechanical arm on the top of the photovoltaic panel array, and carries a high-pressure water gun, generates three continuous control parameters of the mechanical arm speed, water pressure and water volume based on a deep deterministic policy gradient algorithm DDPG, and regulates the mechanical arm and the high-pressure water gun in real time.

[0013] As a further scheme of the present application, the double-camera selects an industrial RGB camera and a near-infrared camera, the industrial RGB camera is additionally provided with an infrared cut filter, and the near-infrared camera is additionally provided with an 850nm narrowband filter.

[0014] As a further scheme of the present application, the specific steps of setting the dynamic time window are as follows:

[0015] The longitude and latitude of the photovoltaic power station, the current date, and the orientation of the photovoltaic panel are obtained;

[0016] An astronomical algorithm library is called to calculate the time period in which the solar elevation angle is less than 30 degrees on the current day;

[0017] The ambient light is monitored in real time through a light sensor to obtain the time period in which the light intensity is greater than 3000lux;

[0018] The intersection period of the solar elevation angle less than 30 degrees and the light greater than 3000lux is obtained to form two candidate windows, and the first candidate window is selected as the effective window;

[0019] If the shadow length is greater than 1 / 3 of the panel length for three consecutive days, the window is automatically advanced or delayed by 15 minutes. The collection frequency f of the double camera is set, wherein T is the length of the dynamic time window.

[0020] As a further scheme of the present application, the specific steps of calculating the coverage R according to the average visible light image are as follows:

[0021] The average visible light image is converted into a gray scale image, and a 3*3 median filter is used to remove salt and pepper noise;

[0022] The gray scale image is converted into an HSV color space, the brightness V channel is extracted, the Otsu adaptive threshold method is used to binarize the V channel, and the image is divided into a dust area and a clean area;

[0023] An opening operation is performed on the binary image to remove small area noise points, and a closing operation is performed to fill small holes in the dust area, the number of dust area pixels Ndust and the total number of image pixels Ntotal are counted, and the coverage R is calculated according to the formula R=Ndust / Ntotal.

[0024] As a further solution of the present invention, the specific steps for calculating ΔG based on the average near-infrared image are as follows:

[0025] Apply a 5 × 5 window median filter to the averaged NIR image;

[0026] The filtered image is divided into several sub-blocks, and the local mean threshold is calculated for each sub-block. Based on the local threshold, the image is binarized into dust area and background area;

[0027] Perform the above two steps on the reference infrared image of the same group of photovoltaic panels in a clean state and the current image to be detected to obtain Iref and Icurr respectively;

[0028] Calculate the absolute grayscale difference between Icurr and Iref pixel by pixel according to the formula ΔI(x,y)=|Icurr(x,y)-Iref(x,y|);

[0029] According to the formula Calculate the mean absolute grayscale difference for the dust area.

[0030] As a further solution of the present invention, the specific steps of constructing a dynamic feature space mapping method for level classification are:

[0031] Normalize R and ΔG and map them to the polar coordinate system: According to the formula Calculate the polar diameter ρ according to the formula Calculate the polar angle θ;

[0032] When ρ<ρ0×(1+k1×var) and θ∈[θ1,θ2], it is classified as light dust pollution;

[0033] When ρ0×(1+k1×var)≤ρ<ρ1×(1+k2×var), it is classified as moderate dust pollution;

[0034] When ρ ≥ ρ1×(1+k2×var), it is classified as heavy dust pollution;

[0035] Among them, ρ0 and ρ1 are the reference radius thresholds, k1 and k2 are dynamic coefficients, θ1 and θ2 are the upper and lower limits of the polar angle when there is light dust pollution, and var is the variance of historical pollution data.

[0036] As a further embodiment of the present invention, the power generation income Sg is calculated according to the formula Sg = ∑Q0 × R × ΔG × η × P, where Q0 is the daily power generation per unit area of ​​the clean photovoltaic panel, η is the dust power generation loss conversion coefficient, and P is the real-time electricity price;

[0037] The cleaning cost Sr is calculated according to the formula Sr=Sr0*(1+k*alpha(R,DeltaG)), wherein Sr0 is a reference cleaning cost, k is a dust pollution level adjustment factor, k is 0 if the dust pollution is light, k is 0.5 if the dust pollution is moderate, and k is 1 if the dust pollution is heavy, and alpha(R,DeltaG) is a level coefficient.

[0038] As a further scheme of the present application, the level coefficient alpha(R,DeltaG) is calculated according to the formula

[0039] A photovoltaic panel adaptive intelligent cleaning control method, comprising:

[0040] S1, using a double camera to collect visible light and near-infrared images of the photovoltaic panel in a dynamic time window at a frequency f;

[0041] S2, averaging a plurality of collected visible light images to obtain an average visible light image, averaging a plurality of collected near-infrared light images to obtain an average near-infrared light image, using the average visible light image to calculate the coverage R, using the average near-infrared light image to calculate the average absolute gray difference DeltaG, and constructing a dynamic feature space mapping method to realize adaptive division of the dust pollution level;

[0042] S3, calculating the power generation income Sg and the cleaning cost Sr of the photovoltaic panel group under different pollution levels, and determining whether to perform a cleaning operation;

[0043] S4, constructing a track type mechanical arm on the top of the photovoltaic panel array and carrying a high-pressure water gun, generating three continuous control parameters of the mechanical arm speed, water pressure and water volume based on the deep deterministic policy gradient algorithm DDPG, to real-time control the mechanical arm and the high-pressure water gun.

[0044] The present application provides a photovoltaic panel adaptive intelligent cleaning control system and method, which has the following advantages compared with the prior art:

[0045] (1) The present application uses a dynamic time window and a double camera to collect, dynamically adjusts the sun elevation angle and the light intensity, avoids shadow and light interference, improves the image quality, and provides reliable data basis for subsequent pollution detection;

[0046] (2) The present application divides the dust pollution level by constructing a dynamic feature space mapping method based on the multi-dimensional features of visible light and near-infrared images, adaptively matches different pollution scenes, and improves the accuracy and adaptability of dust pollution detection.

[0047] ​(3) The application dynamically analyzes the power generation income and cleaning cost, generates the mechanical arm speed, water pressure and water volume by combining the DDPG algorithm, realizes self-adaptive regulation and control of the cleaning strategy execution parameter, improves the cleaning efficiency and reduces the operation and maintenance cost. BRIEF DESCRIPTION OF DRAWINGS

[0048] Figure 1 The system principle frame of the application is shown in the figure;

[0049] Figure 2 The step flow chart of the application is shown in the figure. DETAILED DESCRIPTION

[0050] The technical solutions in the embodiments of the application will be clearly and completely described in combination with the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the application.

[0051] As Figure 1 The application provides a photovoltaic panel adaptive intelligent cleaning control system, which comprises:

[0052] An image acquisition module acquires the visible light and near-infrared image of the photovoltaic panel through a camera;

[0053] One set of double cameras is arranged for every 50 photovoltaic panels, and is installed obliquely above the photovoltaic panel array, with a distance of 3 meters and a depression angle of 45°, so as to avoid the shadow of the photovoltaic panel from shielding the lens;

[0054] The double cameras can be industrial RGB cameras and near-infrared cameras;

[0055] The industrial RGB camera is additionally provided with an infrared cutoff filter, so as to ensure that the RGB channel only receives visible light;

[0056] The near-infrared camera is additionally provided with an 850nm narrowband filter, so as to block visible light interference and make the light transmittance attenuation characteristics of thick dust more obvious;

[0057] A light sensor is installed on the side of the camera support and is at the same height as the lens, and is used for measuring the light intensity in real time;

[0058] When the real-time photovoltaic panel image acquisition is performed, the time period for acquisition needs to be determined first. The traditional method often adopts a fixed time window, however, the fixed time window cannot cover the shadow law of different scenes, because the solar elevation angle changes significantly with latitude, season and date. Therefore, a dynamic time window can be adopted;

[0059] The specific steps of adopting the dynamic time window are as follows:

[0060] Enter the latitude and longitude of the photovoltaic power station (can be automatically obtained through GPS), current date, and photovoltaic panel orientation;

[0061] Call the astronomical algorithm library to calculate the time period when the solar altitude angle is less than 30°;

[0062] The key to choosing a time period with a sun altitude angle of less than 30° is to strike a balance between light quality and shadow control. The specific reasons are as follows:

[0063] When the sun's altitude is low, light hits the photovoltaic panel surface at an oblique angle, which significantly reduces specular reflection compared to direct sunlight at noon, allowing the camera to clearly capture subtle defects on the panel surface, such as hidden cracks and dust accumulation.

[0064] Photovoltaic panels are typically arranged in rows. When the sun's altitude is too high, the shadows of the front row of panels may be cast onto the rear row. However, when the sun's altitude is less than 30°, a reasonable array spacing design can ensure that the self-shadow of a single panel (such as the bracket shadow) and the shadow length of adjacent panels are less than 1 / 3 of the panel's length, avoiding shadow obstruction and interference with detection, such as misinterpreting shadows as stains.

[0065] The ambient light is monitored in real time through a light sensor, and only the time periods with light intensity > 3000 lux are retained;

[0066] The core logic behind retaining periods with light intensities greater than 3000 lux is to find the optimal balance in engineering practice between image acquisition quality and environmental interference control. Periods below 3000 lux may contain stray light, such as reflections from the ground and surrounding buildings, which can lead to uneven image brightness, similar to the flickering lights on and off stage. This can easily lead to light and shadow changes being mistakenly interpreted as defects. Sufficient natural light can unify the lighting direction and reduce background interference.

[0067] The intersection of the solar altitude angle < 30° and the illumination > 3000 lux is selected to form two candidate windows (morning + afternoon). The first valid window in the morning is preferred to avoid possible cloud cover in the afternoon to ensure stable data collection once a day.

[0068] If the shadow length is greater than 1 / 3 of the board length during three consecutive days of collection, the window will be automatically advanced / delayed by 15 minutes;

[0069] If the shadow is too long for three consecutive days, it is essentially a sign that the sun's altitude angle continues to be lower than expected. This may be due to seasonal changes in the sunrise / sunset azimuth, or the local latitude causing the sun's altitude angle corresponding to the time window to gradually deviate from the design threshold.

[0070] A single prolonged shadow may be due to temporary weather conditions (such as cloud cover) or occasional obstruction (such as a bird flying by), but if it persists for three consecutive days, it must be due to a mismatch between the time window and the sun's position. If no adjustment is made, the shadow will continue to cover more than one-third of the PV panel surface, increasing the probability of misjudgment during detection.

[0071] Advancing / delaying the sun by 15 minutes can change the solar altitude angle by about 3.75° (the sun moves 1° every 4 minutes). This angle change is enough to shorten the shadow length by about 1 / 4. The shadow length is inversely proportional to the solar altitude angle, so it is likely that the shadow can be controlled within 1 / 3 of the board length.

[0072] For example, a power station is located at 114° east longitude and 37° north latitude on October 1, 2024. The calculated time periods when the solar angle is less than 30° are 6:30-8:30 a.m. and 15:30-17:30 p.m. The actual light intensity measured by the light sensor is stable at >3000 lux after 7:00 a.m., begins to decrease at 15:30 p.m., and is <3000 lux after 16:00. Therefore, the time period with light intensity greater than 3000 lux is 7:00-16:00. The intersection of the two results shows that the morning period is 7:00-8:30 and the afternoon period is 15:30-16:00, making the final selected window 7:00-8:30.

[0073] After determining the dynamic time window, according to the formula Set the acquisition frequency f of the dual cameras, where T is the length of the dynamic time window;

[0074] Numerator 100-40×e -0.05T Indicates the total number of acquired frames, integrating the exponential decay characteristics and the basic frame number;

[0075] e -0.05T Reflects the time decay of the dust deposition rate. The denominator term 60T converts the time window into seconds, ensuring that the frequency unit is frames per second;

[0076] When T is less than 10 minutes, the frequency decays quickly, matching the high-frequency sampling requirements of sudden pollution;

[0077] When T ≥ 10 minutes, the frequency attenuation slows down to adapt to the slow-changing characteristics of dust deposition and avoid data redundancy.

[0078] The grading module averages multiple visible light images collected by the industrial RGB camera in a dynamic time window to obtain an average visible light image, and averages multiple near-infrared images collected by the near-infrared camera in a dynamic time window to obtain an average near-infrared image;

[0079] The core purpose of averaging multiple images is to suppress random noise. Because factors such as sensor noise and environmental interference are random, when averaging multiple images of the same scene, random noise cancels out due to superposition, while the true scene signal is preserved and enhanced, thereby improving the image signal-to-noise ratio and stability, providing more reliable input for subsequent processing.

[0080] Dust typically appears as a grayscale or color different from the background in visible light images. The dust region extracted through threshold segmentation and morphological processing can directly quantify the "extent" of surface contamination based on its coverage area.

[0081] The specific steps for calculating the coverage R based on the average visible light image are:

[0082] The average visible light image was converted to grayscale and a 3×3 median filter was used to remove salt and pepper noise.

[0083] Grayscale is used to reduce color dimension interference and focus on brightness features (dust usually appears as grayscale differences). Median filtering can eliminate light spots and noise while preserving edges, preventing noise from being mistakenly identified as dust during subsequent segmentation.

[0084] Convert the grayscale image to HSV color space, extract the brightness V channel, use Otsu adaptive threshold method to binarize the V channel, and segment the image into dust area (foreground, low grayscale value) and clean area (background, high grayscale value);

[0085] In HSV space, the brightness channel V is more robust to lighting changes and can reduce the impact of reflections or shadows on the photovoltaic panel surface. The Otsu threshold method automatically determines the optimal segmentation threshold by statistically analyzing the grayscale histogram, adapting to the brightness differences between dust and clean areas under different lighting conditions.

[0086] Perform an opening operation (erode first, then dilate) on the binary image to remove small noise points. Then perform a closing operation (dilate first, then erode) to fill the small holes in the dust area. Count the number of pixels in the dust area Ndust and the total number of pixels in the image Ntotal, and calculate the coverage using the formula R = Ndust / Ntotal.

[0087] Morphological operations can eliminate isolated noise points generated during segmentation and fill holes in the dust area, ensuring that the dust area is a continuous and complete connected domain, making coverage calculation more accurate;

[0088] Near-infrared light has different penetrating and reflective properties from visible light. Dust can change the intensity of a surface's reflection of near-infrared light, causing changes in grayscale values. Calculating the absolute grayscale difference can quantify this change, and its value directly reflects the "depth" of surface contamination.

[0089] The specific steps for calculating the average absolute gray difference ΔG according to the average near-infrared image are as follows:

[0090] Median filtering is applied to the average near-infrared image using a 5*5 window;

[0091] The filtered image is divided into sub-blocks, and the local mean threshold is calculated for each sub-block. Based on the local threshold, the image is binarized into a dust area (high gray level) and a background area (low gray level);

[0092] The surface temperature distribution of the photovoltaic panel may be uneven due to uneven light and internal circuit differences. Global thresholding may lead to false segmentation. Dynamic adjustment of the threshold according to the local gray characteristics of each sub-block can more accurately segment the high-temperature area covered by dust;

[0093] The reference infrared image and the current image to be detected are obtained by performing the above two steps on the same group of photovoltaic panels in a clean state;

[0094] Iref refers to the near-infrared image collected when the photovoltaic panel surface is free of dust and the environmental conditions (such as light intensity, ambient temperature, and shooting angle) are consistent with the state to be detected;

[0095] The absolute gray difference is calculated pixel by pixel for Icurr and Iref according to the formula ΔI(x, y) = |Icurr(x, y) - Iref(x, y)|;

[0096] According to the formula The average absolute gray difference is calculated for the dust area;

[0097] Direct comparison of the gray difference between the clean and dirty states can eliminate common-mode interference such as environmental temperature and light. The average absolute gray difference of the dust area can directly reflect the thickness - the thicker the thickness, the stronger the heat insulation effect, the higher the current image gray relative to the reference image, and the larger the ΔG value;

[0098] A dynamic feature space mapping method is constructed to map the area coverage ratio R and the average absolute gray difference ΔG to the polar coordinate system. The adaptive division of the dust pollution level is realized by dynamically expanding the sector area. The specific operation is as follows:

[0099] After normalization, R and ΔG are mapped to the polar coordinate system:

[0100] The polar radius ρ represents the severity of pollution, which is directly calculated from the geometric mean of R and ΔG, i.e.

[0101] The polar angle θ represents the pollution type bias, which is determined by the ratio of R and ΔG, i.e.

[0102] When θ ≈ 0 (R >> ΔG), it represents large-area thin dust;

[0103] when When , it indicates local thick dust;

[0104] Three dynamically expanding fan-shaped areas are defined in the polar coordinate system, corresponding to the pollution levels:

[0105] When ρ<ρ0×(1+k1×var) and θ∈[θ1,θ2], it is classified as light dust pollution;

[0106] If only the polar diameter is used to determine light dust pollution, local thick dust with small ρ but large θ may be missed. By setting the polar angle range, local thick dust with large θ can be filtered out, leaving only large areas of thin dust as light pollution, avoiding "small areas of thick pollution" being mistakenly judged as not requiring treatment;

[0107] When ρ0×(1+k1×var)≤ρ<ρ1×(1+k2×var), it is classified as moderate dust pollution;

[0108] When ρ ≥ ρ1×(1+k2×var), it is classified as heavy dust pollution;

[0109] Among them, ρ0 and ρ1 are the reference radius thresholds, k1 and k2 are dynamic coefficients, θ1 and θ2 are the upper and lower limits of the polar angle when there is light dust pollution, and var is the variance of historical pollution data, which reflects the degree of environmental fluctuation.

[0110] The cleaning decision module calculates the power generation income Sg and cleaning cost Sr of the photovoltaic panel group under different pollution levels, and determines whether to perform the cleaning operation;

[0111] The cleaning instruction is triggered by comparing the power generation income Sg with the cleaning cost Sr. This is essentially an economic decision-making logic based on dynamic cost-benefit analysis. Its core lies in converting the operation and maintenance strategy into a quantifiable economic trade-off.

[0112] Dust covering photovoltaic panels reduces light transmittance, directly leading to power loss. Sg converts this loss into calculable economic value: the power generation revenue that can be recovered after cleaning. Furthermore, long-term dust accumulation can cause hidden damage such as hot spot effects. Sg also offers preventative protection for equipment lifespan.

[0113] Traditional fixed-cycle cleaning (such as monthly or quarterly) has significant limitations: when pollution is light, over-cleaning may occur, wasting resources; when pollution is heavy, cleaning may be delayed, exacerbating power generation losses. By comparing Sg and Sr in real time, the strategy can be dynamically adjusted according to the actual pollution level. When the power generation benefits brought by cleaning exceed the costs, the operation will be started; otherwise, it will be postponed, achieving precise management of on-demand maintenance.

[0114] The power generation revenue Sg is calculated using the formula Sg = ∑Q0 × R × ΔG × η × P, where Q0 is the daily power generation per unit area of ​​the clean photovoltaic panel, calibrated using historical data, η is the dust power generation loss conversion coefficient, representing the power generation loss caused by one unit (R × ΔG), and P is the real-time electricity price.

[0115] The cleaning cost Sr is calculated according to the formula Sr = Sr0 × (1 + k × α (R, ΔG)), where Sr0 is the baseline cleaning cost;

[0116] k is the dust pollution level adjustment factor. If it is light dust pollution, k is 0; if it is moderate dust pollution, k is 0.5; if it is heavy dust pollution, k is 1;

[0117] α(R,ΔG) is the grade coefficient, and its specific calculation formula is:

[0118] Where w1+w2=1 represents the weight coefficient of coverage R and average absolute grayscale difference ΔG, ΔGmin and ΔGmax are the minimum and maximum grayscale differences in historical data, and Rmask is the coverage mask of the non-detection area, with a value of 0-1.

[0119] The coverage R directly determines the physical area to be cleaned. The larger R is, the wider the area to be cleaned is, and the consumables and time costs increase linearly.

[0120] The average absolute grayscale difference ΔG reflects the dust thickness and adhesion strength. The larger the ΔG, the stronger the dust insulation effect, which usually corresponds to a thicker or more adherent contamination layer, requiring higher cleaning pressure, resulting in a nonlinear increase in energy consumption and equipment loss costs.

[0121] 1-Rmask is used to remove areas in the image that do not need to be cleaned, such as borders and bracket shadows, so that the grade coefficient only reflects the pollution of the effective detection area, where Rmask is less than 0.9;

[0122] W1 focuses on breadth cost: It is suitable for scenarios where dust is evenly distributed but thin, such as floating dust. In this case, the cleaning cost is positively correlated with the area.

[0123] W2 focuses on depth cost: suitable for locally thick dust, such as bird droppings and scab pollution, where the thickness has a greater impact on cleaning difficulty;

[0124] When Sg≥1.2×Sr, the cleaning operation is triggered, otherwise, the cleaning operation is not triggered;

[0125] Cleaning costs Sr usually only include direct investment, while the actual power generation benefits Sg must cover implicit costs, such as downtime losses during the cleaning process and equipment temporary commissioning costs. Setting Sg to 1.2 times Sr essentially ensures that the benefits not only cover the explicit cleaning costs but also offset the implicit losses through the proportional coefficient, avoiding decision-making losses due to incomplete cost accounting.

[0126] Pollution levels, equipment status, environmental factors, and other factors can all cause Sg and Sr to change dynamically. If only Sg = Sr is used as the trigger condition, when data fluctuates, the benefits may just cover the costs but there may be no actual surplus. The 1.2x coefficient essentially introduces a risk buffer for decision-making. By reserving a 20% benefit space, the probability of decision-making errors caused by data uncertainty is reduced.

[0127] The parameter control module builds a track-mounted robotic arm on top of the photovoltaic array and is equipped with a high-pressure water cannon. It uses the deep deterministic policy gradient algorithm (DDPG) to construct the state space S, action space A, and reward function R. Through learning, it generates three continuous control parameters: robotic arm speed, water pressure, and water volume, which are used to control the robotic arm and high-pressure water cannon in real time.

[0128] The track layout is designed to fit the dense arrangement and inclination of photovoltaic panels. The robotic arm achieves full coverage through translation and joint rotation, solving the blind spot problem of manual / ground equipment spraying and ensuring that every panel is cleaned thoroughly.

[0129] High-pressure water guns use impact force to remove stubborn stains such as bird droppings and dust accumulation. This is more efficient than manual wiping and avoids hard scratches that damage the photovoltaic panel coating. Combined with flow control, it achieves water-saving cleaning.

[0130] The specific operations for generating continuous control parameters based on DDPG are:

[0131] Define the state space and construct a multidimensional state vector S that includes pollution characteristics, environmental parameters, and equipment status t =[R t ,ΔG t ,T t ,H t ,P t ,L t ], where R t is the dust coverage of the average visible light image, ΔG t is the average absolute grayscale difference of the average near-infrared image, T t is the real-time temperature (affects the efficiency of photovoltaic panels and the volatilization of cleaning fluid), H t is the air humidity (affects dust stickiness), P t is the current electricity price, L t The service life of the component;

[0132] Design the action space and define the discrete action set A = {robot arm speed, water pressure, water volume};

[0133] Construct a reward function, the specific formula is R(S t ,a t )=γ×(Su-λ×Sr)+(1-γ)×Q t , where Su is the incremental power generation income after cleaning, Sr is the cleaning cost, γ is the recent income weight, which can be 0.7, λ is the cost penalty coefficient, which can be 1.2, Q t To reward component lifespan protection, 10 points will be added for each light cleaning and 5 points will be deducted for heavy cleaning to avoid excessive wear and tear;

[0134] Build an Actor-Critic network. The specific parameters are:

[0135] For the Actor network, the input layer is a 6-dimensional vector St, the fully connected layer 1 is set to 256 neurons, the fully connected layer 2 is set to 128 neurons, the dropout is 0.2, the activation function is ReLU, the output layer is set to 4 neurons, corresponding to 3 actions, and the learning rate is 1e-4;

[0136] For the Critic network, the input layer is a 9-dimensional vector, i.e., state St + action A. The fully connected layer 1 is set with 384 neurons, the fully connected layer 2 is set with 192 neurons, the dropout is 0.3, the activation function is LeakyReLU, the output layer is set with 1 neuron, and the learning rate is 3e-4.

[0137] Collect three years of cleaning records, acquire new states in real time through edge computing, and store them in the experience pool D. Each time, randomly extract several samples from D and use the mean square error (MSE) to optimize network parameters.

[0138] In addition, γ and λ in the reward function can be updated every month based on the latest cleansing data.

[0139] like Figure 2 , a photovoltaic panel adaptive intelligent cleaning control method, the specific steps are:

[0140] S1, using dual cameras to collect visible light and near-infrared images of the photovoltaic panel at a frequency f within a dynamic time window;

[0141] S2. Averaging the multiple collected visible light images to obtain an average visible light image, averaging the multiple collected near-infrared images to obtain an average near-infrared image, using the average visible light image to calculate the coverage R, and using the average near-infrared image to calculate the average absolute grayscale difference ΔG, constructing a dynamic feature space mapping method to achieve adaptive classification of dust pollution levels;

[0142] S3. Calculate the power generation income Sg and cleaning cost Sr of the photovoltaic panel group under different pollution levels, and determine whether to perform the cleaning operation;

[0143] S4. Build a track-mounted robotic arm on top of the photovoltaic panel array and equip it with a high-pressure water gun. Use the deep deterministic policy gradient algorithm (DDPG) to generate three continuous control parameters: robotic arm speed, water pressure, and water volume. This is used to control the robotic arm and high-pressure water gun in real time.

[0144] Some of the data in the above formulas are dimensionless and numerically calculated. Meanwhile, the contents not described in detail in this specification belong to the prior art known to those skilled in the art.

[0145] The above embodiments are only used to illustrate the technical method of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical method of the present invention.

Claims

1. A photovoltaic panel adaptive intelligent cleaning control system, characterized in that: include: The image acquisition module uses dual cameras to collect visible light and near-infrared images of the photovoltaic panel at a frequency f within a dynamic time window; The grading module averages multiple visible light images to obtain an average visible light image, and averages multiple near-infrared images to obtain an average near-infrared image. The average visible light image is used to calculate the coverage R, and the average absolute grayscale difference ΔG is used to calculate the average near-infrared image. A dynamic feature space mapping method is constructed to achieve adaptive classification of dust pollution levels. The cleaning decision module calculates the power generation income Sg and cleaning cost Sr of the photovoltaic panel group under different pollution levels, and determines whether to perform the cleaning operation; The parameter control module builds a track-type robotic arm on top of the photovoltaic panel array and is equipped with a high-pressure water gun. Based on the deep deterministic policy gradient algorithm (DDPG), it generates three continuous control parameters: robotic arm speed, water pressure, and water volume, which are used to control the robotic arm and high-pressure water gun in real time.

2. The photovoltaic panel adaptive intelligent cleaning control system according to claim 1, characterized in that: The dual cameras include an industrial RGB camera and a near-infrared camera. The industrial RGB camera is equipped with an infrared cutoff filter, and the near-infrared camera is equipped with an 850nm narrow-band filter.

3. The photovoltaic panel adaptive intelligent cleaning control system according to claim 1, characterized in that: The specific steps for setting a dynamic time window are: Get the latitude and longitude of the photovoltaic power station, current date, and photovoltaic panel orientation; Call the astronomical algorithm library to calculate the time period when the solar altitude angle is less than 30°; The ambient light is monitored in real time by the light sensor to obtain the time period when the light intensity is greater than 3000 lux; Take the intersection period of the solar altitude angle < 30° and the illumination > 3000 lux to form two candidate windows, and select the first candidate window as the valid window; If the shadow length is greater than 1 / 3 of the board length during three consecutive days of collection, the window will be automatically advanced / delayed by 15 minutes.

4. The photovoltaic panel adaptive intelligent cleaning control system according to claim 1, characterized in that: After determining the dynamic time window, according to the formula Set the acquisition frequency f of the dual cameras, where T is the length of the dynamic time window.

5. The photovoltaic panel adaptive intelligent cleaning control system according to claim 1, characterized in that: The specific steps for calculating the coverage R based on the average visible light image are: The average visible light image was converted to grayscale and a 3×3 median filter was used to remove salt and pepper noise. The grayscale image is converted to HSV color space, the brightness V channel is extracted, and the V channel is binarized using the Otsu adaptive threshold method to segment the image into dust area and clean area; An opening operation is performed on the binary image to remove small area noise points, and then a closing operation is performed to fill the small holes in the dust area. The number of pixels in the dust area Ndust and the total number of pixels in the image Ntotal are counted, and the coverage is calculated according to the formula R = Ndust / Ntotal.

6. The photovoltaic panel adaptive intelligent cleaning control system according to claim 1, characterized in that: The specific steps for calculating ΔG based on the average near-infrared image are: Apply a 5 × 5 window median filter to the averaged NIR image; The filtered image is divided into several sub-blocks, and the local mean threshold is calculated for each sub-block. Based on the local threshold, the image is binarized into dust area and background area; Perform the above two steps on the reference infrared image of the same group of photovoltaic panels in a clean state and the current image to be detected to obtain Iref and Icurr respectively; Calculate the absolute grayscale difference between Icurr and Iref pixel by pixel according to the formula ΔI(x,y)=|Icurr(x,y)-Iref(x,y|); According to the formula Calculate the mean absolute grayscale difference for the dust area.

7. The photovoltaic panel adaptive intelligent cleaning control system according to claim 1, characterized in that: The specific steps of constructing the dynamic feature space mapping method for level division are: Normalize R and ΔG and map them to the polar coordinate system: According to the formula Calculate the polar diameter ρ according to the formula Calculate the polar angle θ; When ρ<ρ0×(1+k1×var) and θ∈[θ1,θ2], it is classified as light dust pollution; When ρ0×(1+k1×var)≤ρ<ρ1×(1+k2×var), it is classified as moderate dust pollution; When ρ ≥ ρ1×(1+k2×var), it is classified as heavy dust pollution; Among them, ρ0 and ρ1 are the reference radius thresholds, k1 and k2 are dynamic coefficients, θ1 and θ2 are the upper and lower limits of the polar angle when there is light dust pollution, and var is the variance of historical pollution data.

8. The photovoltaic panel adaptive intelligent cleaning control system according to claim 1, characterized in that: The power generation income Sg is calculated according to the formula Sg = ∑Q0 × R × ΔG × η × P, where Q0 is the daily power generation per unit area of ​​the clean photovoltaic panel, η is the dust power generation loss conversion coefficient, and P is the real-time electricity price; The cleaning cost Sr is calculated according to the formula Sr=Sr0×(1+k×α(R,ΔG)), where Sr0 is the baseline cleaning cost, k is the dust pollution level adjustment factor, if it is light dust pollution, k is 0, if it is moderate dust pollution, k is 0.5, if it is heavy dust pollution, k is 1, and α(R,ΔG) is the level coefficient.

9. The photovoltaic panel adaptive intelligent cleaning control system according to claim 8, characterized in that: According to the formula The grade coefficient α(R,ΔG) is calculated, where w1 and w2 represent the weight coefficients of coverage R and average absolute grayscale difference ΔG, ΔGmin and ΔGmax are the minimum and maximum grayscale differences in historical data, and Rmask is the coverage mask of the non-detection area.

10. A photovoltaic panel adaptive intelligent cleaning control method, used to implement a photovoltaic panel adaptive intelligent cleaning control system according to any one of claims 1 to 9, characterized in that: include: S1, using dual cameras to collect visible light and near-infrared images of the photovoltaic panel at a frequency f within a dynamic time window; S2. Averaging the multiple collected visible light images to obtain an average visible light image, averaging the multiple collected near-infrared images to obtain an average near-infrared image, using the average visible light image to calculate the coverage R, and using the average near-infrared image to calculate the average absolute grayscale difference ΔG, constructing a dynamic feature space mapping method to achieve adaptive classification of dust pollution levels; S3. Calculate the power generation income Sg and cleaning cost Sr of the photovoltaic panel group under different pollution levels, and determine whether to perform the cleaning operation; S4. Build a track-mounted robotic arm on top of the photovoltaic panel array and equip it with a high-pressure water gun. Use the deep deterministic policy gradient algorithm (DDPG) to generate three continuous control parameters: robotic arm speed, water pressure, and water volume. This is used to control the robotic arm and high-pressure water gun in real time.

Images