A dust real-time monitoring method for intelligent operation and maintenance of a photovoltaic power station

Abstract

The application discloses a kind of dust real-time monitoring methods of photovoltaic station intelligent operation and maintenance, and the application relates to photovoltaic panel technical field.The method steps include: collecting the first photovoltaic panel dust image and diffuse reflection standard plate image under different dust density;Remove noise by Scharr operator and gradient weighted filtering method, obtain the second photovoltaic panel dust image;Calculate the image gray ratio of second photovoltaic panel dust image and diffuse reflection standard plate image;Through the second photovoltaic panel dust image, image gray ratio and dust density, train improved DenseNet169 network model;Establish the illumination-power generation power curve of photovoltaic panel under clean state;Real-time collection second photovoltaic panel dust image and image gray ratio input to trained DenseNet169 network model, output dust density prediction value;Real-time collection ambient light intensity and power generation power input illumination-power generation power curve, calculate difference to get photovoltaic panel power generation coefficient;Combine dust density prediction value and photovoltaic panel power generation coefficient to calculate photovoltaic panel dust index.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic panel technology, specifically to a method for real-time dust monitoring in intelligent operation and maintenance of photovoltaic power plants. Background Technology

[0002] As the global energy structure shifts towards clean energy, the photovoltaic industry, as an important component of renewable energy, continues to expand its installed capacity and scale of power plants. Dust accumulation on the surface of photovoltaic panels significantly reduces light absorption efficiency, leading to a decrease in power generation. Therefore, real-time monitoring of dust on photovoltaic panels has become a crucial link in ensuring efficient operation and maintenance of photovoltaic power plants and improving power generation efficiency.

[0003] Currently, dust monitoring at photovoltaic power plants generally adopts traditional image analysis methods. Specifically, it involves collecting images of photovoltaic panels and combining information such as the grayscale characteristics of the photovoltaic panel images to assess the dust accumulation status and determine the degree of dust accumulation.

[0004] However, in existing technologies, traditional image analysis methods are difficult to eliminate the interference of ambient light intensity fluctuations on the accuracy of dust monitoring. Light changes can cause unstable image grayscale values, resulting in large errors in the quantitative assessment of dust density, and thus failing to provide accurate and reliable decision-making basis for photovoltaic power plant operation and maintenance. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a real-time dust monitoring method for intelligent operation and maintenance of photovoltaic power plants, thereby resolving the problems existing in the background technology.

[0006] To achieve the above objectives, the present invention provides a real-time dust monitoring method for intelligent operation and maintenance of photovoltaic power plants, comprising the following steps:

[0007] Step S1: Install a diffuse reflection standard plate on the photovoltaic panel and collect images of the photovoltaic panel and the diffuse reflection standard plate under different dust accumulation densities to obtain the first dust accumulation image of the photovoltaic panel and the diffuse reflection standard plate image;

[0008] Step S2: Calculate the gradient magnitude of the first photovoltaic panel dust accumulation image using the Scharr operator. Based on the gradient magnitude, remove noise from the first photovoltaic panel dust accumulation image using a gradient weighted filtering method to obtain the second photovoltaic panel dust accumulation image.

[0009] Step S3: The grayscale ratio of the images is calculated by comparing the grayscale values ​​of the second photovoltaic panel dust accumulation image and the diffuse reflection standard plate image;

[0010] Step S4: Input the second photovoltaic panel dust accumulation image, image grayscale ratio, and dust accumulation density into the improved DenseNet169 network model for training to obtain the trained DenseNet169 network model; collect the ambient light intensity and clean power generation of the photovoltaic panel in the clean state, and perform curve fitting to obtain the light intensity-power generation curve in the clean state.

[0011] Step S5: Collect the image of the second photovoltaic panel with dust accumulation and the image grayscale ratio in real time and input it into the trained DenseNet169 network model to output the predicted value of dust accumulation density; collect the ambient light intensity and power generation of the photovoltaic panel in real time, input the ambient light intensity into the light intensity-power generation curve under clean conditions, and output the clean power generation; calculate the difference between the clean power generation and the power generation to obtain the photovoltaic panel power generation coefficient.

[0012] Step S6: By combining the predicted dust density value and the photovoltaic power generation coefficient, the photovoltaic panel dust index is calculated. The photovoltaic panel dust index is compared with a preset threshold and a risk level is established to achieve real-time monitoring of dust.

[0013] Preferably, the step of installing a diffuse reflection standard plate on the photovoltaic panel and collecting images of the photovoltaic panel and the diffuse reflection standard plate under different dust accumulation densities to obtain a first dust accumulation image of the photovoltaic panel and an image of the diffuse reflection standard plate includes the following specific steps:

[0014] The photovoltaic panel is laid flat on the ground with its front side facing up. A sealed plexiglass cover of the same size is placed on the photovoltaic panel to provide a closed and controllable experimental environment. A diffuse reflection standard plate is installed at a fixed position on each photovoltaic panel to be monitored. The diffuse reflection standard plate is closely attached to the outside of the photovoltaic panel frame and is parallel to the photovoltaic panel surface and within the same imaging angle range.

[0015] Use a blower to continuously blow air through the only air inlet on the side of the plexiglass cover, carrying dust particles, until the dust particles are completely lifted in the sealed container; then stop; let the photovoltaic panel stand for 10 hours in a dry environment to allow the dust to settle naturally and adhere evenly to the surface of the photovoltaic panel; repeat the above steps to evenly sprinkle dust particles with mass densities of 0, 5, 10, 15, 20, 25 and 30 g / m² on the surface of photovoltaic panels of the same size.

[0016] After the dust accumulation is completed, the plexiglass cover is removed, and the camera position is calibrated so that a single frame of the camera simultaneously and completely covers the photovoltaic panel and the attached diffuse reflection standard panel. Based on the template matching algorithm of SIFT feature points, the diffuse reflection standard panel area is automatically located when the photovoltaic panel image is acquired, and the located diffuse reflection standard panel area is cropped to obtain the diffuse reflection standard panel image. The cropped image is the first photovoltaic panel dust accumulation image. Finally, the first photovoltaic panel dust accumulation image and the diffuse reflection standard panel image under different dust accumulation densities are obtained.

[0017] Preferably, the step of calculating the gradient magnitude of the first photovoltaic panel dust accumulation image using the Scharr operator includes the following steps:

[0018] The gradient magnitude of the dust accumulation image of the first photovoltaic panel is calculated using the Scharr operator, and based on the gradient magnitude:

[0019]

[0020] in, Let be the gradient value of the dust accumulation image of the first photovoltaic panel in the horizontal direction at coordinates (x, y). Let be the gradient value of the dust accumulation image of the first photovoltaic panel in the vertical direction at coordinates (x, y). Let (x, y) be the pixel value of the first photovoltaic panel's dust accumulation image at coordinates (x, y). For the horizontal gradient operator of the Scharr operator, The vertical gradient operator for the Scharr operator;

[0021] Calculate the gradient magnitude of the first photovoltaic panel's dust accumulation image based on the horizontal and vertical gradient values:

[0022]

[0023] in, Let be the gradient magnitude of the dust accumulation image of the first photovoltaic panel at coordinates (x, y). Let be the gradient value of the dust accumulation image of the first photovoltaic panel in the horizontal direction at coordinates (x, y). The gradient value of the first photovoltaic panel's dust accumulation image in the vertical direction at coordinates (x, y).

[0024] Preferably, the step of removing noise from the first photovoltaic panel dust accumulation image based on the gradient magnitude using a gradient-weighted filtering method to obtain the second photovoltaic panel dust accumulation image includes the following steps:

[0025] For each point (x, y) on the first photovoltaic panel dust accumulation image, calculate the vertical relative gradient intensity of point (x, y):

[0026]

[0027] in, Let be the vertical relative gradient intensity at point (x, y). Let be the gradient magnitude of the dust accumulation image of the first photovoltaic panel at coordinates (x, y+1). The gradient magnitude of the dust accumulation image of the first photovoltaic panel at coordinates (x, y);

[0028] For each point (x, y) on the first photovoltaic panel dust accumulation image, calculate the horizontal relative gradient intensity of point (x, y):

[0029]

[0030] in, Let (x, y) be the horizontal relative gradient intensity. Let be the gradient magnitude of the dust accumulation image of the first photovoltaic panel at coordinates (x+1, y). The gradient magnitude of the dust accumulation image of the first photovoltaic panel at coordinates (x, y);

[0031] Calculate the symmetrical gradient change rate of all adjacent pixels in the vertical and horizontal directions in the first photovoltaic panel dust accumulation image:

[0032]

[0033]

[0034] in, This represents the degree of abrupt change in gradient at point (x, y) in the vertical direction. The larger the value, the more drastic the gradient change between (x, y) and the point below it (x, y+1). The value represents the degree of gradient abrupt change at point (x,y) in the horizontal direction. The larger the value, the more drastic the gradient change between (x,y) and its rightmost point (x+1,y).

[0035] calculate , The degree of abrupt change in the total directional gradient is given by the degree of abrupt change in the total directional gradient in the image of the first photovoltaic panel with dust accumulation. The total directional gradient abrupt change in the first photovoltaic panel's dust accumulation image is sorted in ascending order, and its 90th percentile is taken. Calculate the basic threshold , ;

[0036] Introducing tolerance coefficient Generate dynamic thresholds:

[0037]

[0038] Where d is the gradient intensity threshold, This is the tolerance factor, which defaults to 1.1. Basic threshold;

[0039] By comparing the relative gradient strengths and gradient strength thresholds in the horizontal and vertical directions, When the gradient changes between the two pixels are relatively flat, the pixel (x,y) remains unchanged. This indicates a significant gradient abrupt change between the two pixels. Therefore, a gradient-weighted filtering method is applied to the pixel at coordinates (x, y). For the 3x3 neighborhood window of this pixel at coordinates (x, y), the gradient weight between each neighboring point in the 3x3 neighborhood window and the pixel at coordinates (x, y) is calculated.

[0040]

[0041] in, Let be the gradient weight of the j-th neighboring pixel of the pixel with coordinates (x, y). Let be the gradient magnitude of the j-th neighboring pixel of the pixel with coordinates (x, y). To prevent division by zero, the value is taken as... ;

[0042] Based on the gradient magnitude, noise removal is performed on the first photovoltaic panel dust accumulation image using a gradient weighted filtering method:

[0043] when and When that happens, the pixels are preserved. = ;

[0044] when hour:

[0045]

[0046] in, Let be the pixel value of the second photovoltaic panel's dust accumulation image at coordinates (x, y), j be the index of the j-th neighboring pixel at coordinates (x, y), and m be the number of all neighboring pixels of the pixel at coordinates (x, y). Let be the gradient weight of the j-th neighboring pixel of the pixel with coordinates (x, y). This represents the pixel value of the j-th neighboring pixel of the (x, y) pixel point in the image of the first photovoltaic panel with dust accumulation.

[0047] Preferably, the step of calculating the image grayscale ratio by comparing the grayscale values ​​of the second photovoltaic panel's dust accumulation image and the diffuse reflection standard panel image includes the following steps:

[0048] By converting the grayscale value of the second photovoltaic panel's dust accumulation image, the grayscale value of the second photovoltaic panel's dust accumulation image is obtained:

[0049]

[0050] in, Let (x, y) be the grayscale value of the image of the second photovoltaic panel with dust accumulation. The red channel value of the second photovoltaic panel's dust accumulation image at coordinates (x, y) is... The green channel value of the second photovoltaic panel's dust accumulation image at coordinates (x, y) is given. The blue channel value of the second photovoltaic panel's dust accumulation image at coordinates (x, y);

[0051] Calculate the average grayscale value of the image of dust accumulation on the second photovoltaic panel:

[0052]

[0053] in, Let W be the average grayscale value of the second photovoltaic panel's dust accumulation image, W be the width of the second photovoltaic panel's dust accumulation image, and H be the height of the second photovoltaic panel's dust accumulation image. The grayscale value of the image of the second photovoltaic panel with accumulated dust is given at coordinates (x, y).

[0054] The diffuse reflection standard plate image is converted to grayscale, and the average grayscale value is calculated to obtain the average grayscale value of the diffuse reflection standard plate. ;

[0055] The image grayscale ratio is calculated by combining the average grayscale value of the second photovoltaic panel and the average grayscale value of the diffuse reflection standard panel:

[0056]

[0057] in, Grayscale ratio, The average grayscale value of the image showing dust accumulation on the second photovoltaic panel. This represents the average grayscale value of the diffuse reflection standard plate.

[0058] Preferably, the step of inputting the second photovoltaic panel's dust accumulation image, image grayscale ratio, and dust accumulation density into the improved DenseNet169 network model for training to obtain the trained DenseNet169 network model includes the following specific steps:

[0059] The image of the second photovoltaic panel with accumulated dust and the normalized grayscale ratio were used as the input to the main and auxiliary channels of the improved DenseNet169 network model, and the dust accumulation density was used as the output layer of the improved DenseNet169 network model. The data set size was expanded by using methods such as random flipping, rotation, simulation of different illuminations, affine transformation, and Gaussian blur on the image of the second photovoltaic panel with accumulated dust.

[0060] The improved DenseNet169 network model includes: a dual-channel input interface, a backbone network, auxiliary branches, and a gated residual fusion module;

[0061] Dual-channel input interface: The main channel receives a 224×224 pixel image of the second photovoltaic panel's dust accumulation; the auxiliary channel receives the grayscale ratio.

[0062] The backbone network consists of an initial convolutional layer, a max pooling layer, four dense blocks, and three transition layers.

[0063] The initial convolutional layer uses a 7*7 convolutional kernel with a stride of 2; the max pooling layer uses a 3*3 convolutional layer with a stride of 2.

[0064] Four dense blocks: The first dense block contains a structure of 6 fully connected convolutional layers. Each layer is processed by the sequence "batch normalization-ReLU activation function-1*1 convolution-batch normalization-ReLU activation function-3*3 convolution". The 3*3 convolution in the third layer is replaced by an ACGBlock module. The output channels are expanded to 256 through a dense connection mechanism. The first dense block is followed by the first transition layer. The first transition layer compresses the channels to 128 through 1*1 convolution and uses 2*2 average pooling to achieve spatial downsampling, outputting a 28*28*128 feature map.

[0065] ACGBlock module: Replaces the 1*3 and 3*1 convolutions in the asymmetric convolution module with grouped convolutions; replaces the 3*3 convolutions in the original asymmetric convolution module with conditional convolutions of the same size;

[0066] The second dense block contains 12 fully connected convolutional layers. Each layer also uses the sequence of "batch normalization-ReLU activation function-1*1 convolution-batch normalization-ReLU activation function-3*3 convolution", but the 3*3 convolutions in layers 5 and 9 are replaced with ACGBlock modules. The second dense block is followed by a second transition layer. The second transition layer compresses the channels to 256 through 1*1 convolution and uses 2*2 average pooling to achieve spatial downsampling, outputting a 14*14*256 feature map.

[0067] The third dense block contains 32 fully connected convolutional layers. Each layer also uses the sequence of "batch normalization - ReLU activation function - 1*1 convolution - batch normalization - ReLU activation function - 3*3 convolution". In layers 8, 16 and 24, the 3*3 convolutions are replaced with ACGBlock modules. The third dense block is followed by the third transition layer. The third transition layer compresses the channels to 256 through 1*1 convolution and uses 2*2 average pooling to achieve spatial downsampling, outputting a 7*7*512 feature map.

[0068] The fourth dense block contains 32 fully connected convolutional layers. Each layer also uses the sequence of "batch normalization-ReLU activation function-1*1 convolution-batch normalization-ReLU activation function-3*3 convolution", but the 3*3 convolutions in the 11th and 12th layers are replaced with ACGBlock modules.

[0069] An EMS module is embedded after the fourth dense block. The EMS module includes average pooling and standard deviation pooling. Finally, a 1024-dimensional visual feature vector is output through the Sigmoid function and spatial weighted aggregation.

[0070] Auxiliary branch: The normalized grayscale ratio is converted into a 16-dimensional feature vector through a two-layer fully connected network. The ReLU activation function is used to enhance the non-linear expressive power to obtain an auxiliary feature vector.

[0071] Gated residual fusion module: A dynamic gating mechanism is used to establish cross-modal association. The auxiliary feature vector is input into the Sigmoid function to generate 1024-dimensional gating weights. The visual feature vector is multiplied element by element with the gating weights. After transformation by a 128-dimensional fully connected layer, a 128-dimensional fused feature vector is output.

[0072] Regression output layer: The fused feature vectors are non-linearly mapped through a 128-unit fully connected layer; finally, the predicted ash density is output by a linearly activated neuron. Based on the predicted ash density and the input ash density, the improved DenseNet169 network model is trained using the cross-entropy loss function and optimized using the Ranger optimizer to obtain the trained DenseNet169 network model.

[0073] Preferably, the process of collecting the ambient light intensity and clean power generation of the photovoltaic panel under clean conditions, and performing curve fitting to obtain the light intensity-power generation curve under clean conditions includes the following specific steps:

[0074] Collect the ambient light intensity and corresponding clean power generation when the photovoltaic panels are in a clean state, ensuring that the timestamps of each set of data are consistent and establishing a one-to-one correspondence; preprocess the collected clean power generation data to remove outliers caused by equipment failure, measurement errors, etc., and retain valid data samples.

[0075] The light intensity-power generation curve under clean conditions was obtained by curve fitting using the least squares method with ambient light intensity as the independent variable and clean power generation as the dependent variable.

[0076]

[0077] Where P represents clean power generation, GZ represents ambient light intensity, and a, b, and c are the parameters of the light-power generation curve under clean conditions. The temperature coefficient is determined by the material of the photovoltaic panel. For ambient temperature, The reference temperature is 25°C by default.

[0078] Preferably, the step of collecting real-time images of the second photovoltaic panel's dust accumulation and its grayscale ratio, inputting them into a trained DenseNet169 network model, and outputting a predicted dust accumulation density value includes the following specific steps:

[0079] The system collects real-time images of dust accumulation on the second photovoltaic panel and their grayscale ratios, then inputs these images into a trained DenseNet169 network model to output a predicted dust accumulation density value.

[0080]

[0081] in, To predict ash accumulation density, DenseNet169() is the input function for the DenseNet169 model. This is an image of dust accumulation on the second photovoltaic panel. This represents the image grayscale ratio.

[0082] Preferably, the real-time collection of ambient light intensity and power generation of the photovoltaic panel, inputting the ambient light intensity into the light-power generation curve under clean conditions, outputting the clean power generation, calculating the difference between the clean power generation and the total power generation to obtain the photovoltaic panel's power generation coefficient, includes the following specific steps:

[0083] The ambient light intensity, ambient temperature, and actual power generation of the photovoltaic panels are collected in real time within a preset time window T. The ambient light intensity and ambient temperature are then input into the light intensity-power generation curve under clean conditions to obtain the clean power generation.

[0084]

[0085] in, Let be the clean power generation capacity at time t. Let be the ambient light intensity at time t, and a, b, and c be the parameters of the light intensity-historical clean power generation curve. The temperature coefficient is determined by the material of the photovoltaic panel. Let be the ambient temperature at time t. For reference temperature, the default value is 25°C;

[0086] The difference between the clean power generation and the actual power generation is calculated to obtain the photovoltaic power generation coefficient:

[0087]

[0088] Where C is the photovoltaic power generation coefficient, T is the length of the preset time window, and t is time t. Let be the clean power generation capacity at time t. Let t be the actual power generation at time t.

[0089] Preferably, the step of calculating the photovoltaic panel dust index by combining the predicted dust density value and the photovoltaic panel power generation coefficient includes the following specific steps:

[0090] By combining the predicted dust density value and the photovoltaic panel power generation coefficient, the photovoltaic panel dust index is calculated:

[0091]

[0092] Wherein, DI is the dust index of the photovoltaic panel. To predict ash accumulation density, This is a reference value for dust accumulation density, with a default value of 10. C represents the photovoltaic power generation coefficient. This is the adjustment factor, with a default value of 10.

[0093] This invention provides a real-time dust monitoring method for intelligent operation and maintenance of photovoltaic power plants, involving machine learning and deep learning technologies, which has the following beneficial effects:

[0094] (1) Based on the gradient magnitude, the gradient weighted filtering method is used to remove noise from the first photovoltaic panel's dust accumulation image. Its core significance lies in accurately distinguishing between the uniform dust accumulation area and the grid line noise area. When the gradient change of adjacent pixels is gradual, the original pixel value is retained to avoid blurring of the dust accumulation texture. When the gradient change is obvious, the grid line noise is suppressed by gradient weighted interpolation, which effectively eliminates the interference of the silver grid lines on the photovoltaic panel surface on image analysis.

[0095] (2) The key significance of calculating the image grayscale ratio by combining the average grayscale value of the second photovoltaic panel and the average grayscale value of the diffuse reflection standard panel lies in eliminating the influence of ambient light intensity fluctuations. The reflectivity of the diffuse reflection standard panel is constant, and its average grayscale value only changes with light intensity. However, the average grayscale value of the photovoltaic panel is affected by both dust accumulation and light intensity. The ratio of the two can offset the light intensity interference and only reflect the change in reflectivity of the photovoltaic panel surface caused by dust accumulation, thus providing a stable characteristic parameter for the accurate assessment of dust accumulation density.

[0096] (3) The significance of calculating the photovoltaic panel dust index by combining the predicted dust density and the photovoltaic panel power generation coefficient lies in achieving a comprehensive quantitative assessment of the impact of dust. This index reflects both the physical density of dust accumulation (reflecting the degree of dust accumulation) and the power generation coefficient (reflecting the actual impact of dust on power generation efficiency), avoiding the one-sidedness of a single indicator. By comparing with preset thresholds to establish risk levels, different operation and maintenance responses can be accurately triggered, providing a scientific and timely basis for dust monitoring and operation and maintenance decisions for photovoltaic power plants. Attached Figure Description

[0097] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0098] Figure 1 This is a flowchart of the steps of a real-time dust monitoring method for intelligent operation and maintenance of photovoltaic power plants proposed in this invention;

[0099] Figure 2 This is a step hierarchy diagram of obtaining the image grayscale ratio in a real-time dust monitoring method for intelligent operation and maintenance of photovoltaic power stations proposed in this invention;

[0100] Figure 3 This is a step hierarchy diagram of obtaining the dust index of photovoltaic panels in a real-time dust monitoring method for intelligent operation and maintenance of photovoltaic power stations proposed in this invention. Detailed Implementation

[0101] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0102] Please see Figures 1-3 This invention provides a technical solution: a method for real-time dust monitoring in intelligent operation and maintenance of photovoltaic power plants.

[0103] Step S1: Install a diffuse reflection standard plate on the photovoltaic panel and collect images of the photovoltaic panel and the diffuse reflection standard plate under different dust accumulation densities to obtain the first image of the photovoltaic panel dust accumulation and the image of the diffuse reflection standard plate.

[0104] To achieve light-independent monitoring of dust accumulation on photovoltaic (PV) panels, a diffuse reflection standard panel must be installed at a fixed position on each PV panel to be monitored. This standard panel should be tightly fitted to the outer edge of the PV panel frame, ensuring its plane is parallel to the PV panel surface and within the same imaging angle. The standard panel is made of a highly stable ceramic-based composite material with a matte finish and a self-cleaning coating. Its calibrated reflectivity is constant at 20% ± 1% (or a specific value selected according to the scenario requirements). During installation, it is mechanically secured using stainless steel clips and anti-loosening screws. The center of the standard panel should be 5-8 cm from the edge of the PV panel, and the overall size should be controlled to 3 cm × 3 cm to avoid module shading. An installation positioning diagram must be created during on-site deployment, recording the coordinate number and corresponding reflectivity parameters of each diffuse reflection standard panel within the PV array.

[0105] Lay the photovoltaic panel face up on the ground and cover it with a sealed plexiglass cover of identical size (providing a closed and controlled experimental environment). Blow air through the single air inlet on the side of the plexiglass cover, carrying a certain mass of sieved dust particles. The total mass of dust particles blown into the container should be slightly higher than the expected dust cover mass to avoid experimental errors caused by dust settling on the plexiglass cover surface. Continuously blow air with a blower until the dust is completely agitated in the sealed container. Let the photovoltaic panel stand in a dry environment for 10 hours to allow the dust to settle naturally and adhere evenly to the surface. To prevent dust particles from falling off, humidify the photovoltaic panel surface using an air humidifier. During the experiment, the air humidifier should be kept at a safe distance from the dust-covered photovoltaic panel surface to prevent damage. Repeat the above steps, evenly sprinkling dust particles with mass densities of 0, 5, 10, 15, 20, 25, and 30 g / m² onto the surface of identical photovoltaic panels.

[0106] After dust accumulation, the plexiglass cover was removed, and the camera position was calibrated so that a single frame simultaneously and completely covered the target photovoltaic panel and the associated diffuse reflection standard panel. Based on the SIFT feature point template matching algorithm, the diffuse reflection standard panel area was automatically located and cropped out when acquiring the photovoltaic panel image. Finally, the first photovoltaic panel dust accumulation image and the diffuse reflection standard panel image under different dust accumulation densities were obtained.

[0107] Step S2: Calculate the gradient magnitude of the first photovoltaic panel dust accumulation image using the Scharr operator. Based on the gradient magnitude, remove noise from the first photovoltaic panel dust accumulation image using a gradient weighted filtering method to obtain the second photovoltaic panel dust accumulation image.

[0108] The surface of a photovoltaic panel is covered with parallel metal lines, called grid lines. These grid lines are silver and evenly and neatly arranged on the surface of the photovoltaic panel, interfering with the subsequent calculation of the grayscale values ​​of the image. In order to minimize the interference of the silver grid lines on the image analysis results, when analyzing the grayscale values ​​of photovoltaic panels under dust conditions, the silver grid lines on the photovoltaic panel should be removed first.

[0109] The gradient magnitude of the dust accumulation image of the first photovoltaic panel is calculated using the Scharr operator, and based on the gradient magnitude:

[0110]

[0111] in, Let be the gradient value of the dust accumulation image of the first photovoltaic panel in the horizontal direction at coordinates (x, y). Let be the gradient value of the dust accumulation image of the first photovoltaic panel in the vertical direction at coordinates (x, y). Let (x, y) be the pixel value of the first photovoltaic panel's dust accumulation image at coordinates (x, y). For the horizontal gradient operator of the Scharr operator, This is the vertical gradient operator for the Scharr operator.

[0112] Calculate the gradient magnitude of the first photovoltaic panel's dust accumulation image based on the horizontal and vertical gradient values:

[0113]

[0114] in, Let be the gradient magnitude of the dust accumulation image of the first photovoltaic panel at coordinates (x, y). Let be the gradient value of the dust accumulation image of the first photovoltaic panel in the horizontal direction at coordinates (x, y). The gradient value of the first photovoltaic panel's dust accumulation image in the vertical direction at coordinates (x, y).

[0115] For each point (x, y) on the first photovoltaic panel dust accumulation image, calculate the vertical relative gradient intensity of point (x, y):

[0116]

[0117] in, Let be the vertical relative gradient intensity at point (x, y). Let be the gradient magnitude of the dust accumulation image of the first photovoltaic panel at coordinates (x, y+1). The gradient magnitude of the first photovoltaic panel's dust accumulation image at coordinates (x, y).

[0118] For each point (x, y) on the first photovoltaic panel dust accumulation image, calculate the horizontal relative gradient intensity of point (x, y):

[0119]

[0120] in, Let (x, y) be the horizontal relative gradient intensity. Let be the gradient magnitude of the dust accumulation image of the first photovoltaic panel at coordinates (x+1, y). The gradient magnitude of the first photovoltaic panel's dust accumulation image at coordinates (x, y).

[0121] Calculate the symmetrical gradient change rate of all adjacent pixels in the vertical and horizontal directions in the first photovoltaic panel dust accumulation image:

[0122]

[0123]

[0124] in, This represents the degree of abrupt change in gradient at point (x, y) in the vertical direction. The larger the value, the more drastic the gradient change between (x, y) and the point below it (x, y+1). The value represents the degree of gradient abrupt change at point (x,y) in the horizontal direction. The larger the value, the more drastic the gradient change between (x,y) and its rightmost point (x+1,y).

[0125] calculate , This represents the degree of abrupt change in the total directional gradient. For the total directional gradient abrupt change in the first photovoltaic panel's dust accumulation image... The total directional gradient abrupt change in the first photovoltaic panel's dust accumulation image is sorted in ascending order, and its 90th percentile is taken. (This quantile covers 90% of the uniform dust accumulation area and excludes 10% of local noise points, such as minor scratches or abnormal dust accumulation points), calculate the base threshold. Introducing tolerance coefficient (Default 1.1, can be calibrated experimentally within ±20%) Generate dynamic threshold:

[0126]

[0127] Where d is the gradient intensity threshold, This is the tolerance factor, with a default value of 0.9. The base threshold.

[0128] By comparing the relative gradient strength in the horizontal and vertical directions with a gradient strength threshold, when and When the gradient change between the two pixels is gradual (in a uniform dust accumulation area), the pixel (x, y) remains unchanged; when the horizontal direction is ( or ) or vertical direction is ( or When the gradient change between two pixels is significant, indicating a grid line region, a gradient-weighted filtering method is applied to the pixel at coordinates (x, y). For the 3x3 neighborhood window of the pixel at coordinates (x, y), the gradient weight between each neighboring point in the 3x3 neighborhood window and the pixel at coordinates (x, y) is calculated.

[0129]

[0130] in, Let be the gradient weight of the j-th neighboring pixel of the pixel with coordinates (x, y). Let be the gradient magnitude of the j-th neighboring pixel of the pixel with coordinates (x, y). To prevent division by zero, the value is taken as... .

[0131] Based on the gradient magnitude, noise removal is performed on the first photovoltaic panel dust accumulation image using a gradient weighted filtering method:

[0132] when and In this case, the pixels are preserved: = ;

[0133] when hour:

[0134]

[0135] in, Let be the pixel value of the second photovoltaic panel's dust accumulation image at coordinates (x, y), j be the index of the j-th neighboring pixel at coordinates (x, y), and m be the number of all neighboring pixels of the pixel at coordinates (x, y). Let be the gradient weight of the j-th neighboring pixel of the pixel with coordinates (x, y). This represents the pixel value of the j-th neighboring pixel of the (x, y) pixel point in the image of the first photovoltaic panel with dust accumulation.

[0136] By removing noise from the first photovoltaic panel dust accumulation image, the second photovoltaic panel dust accumulation image is finally obtained.

[0137] Step S3: The grayscale ratio of the images is calculated by comparing the grayscale values ​​of the second photovoltaic panel's dust accumulation image and the diffuse reflection standard plate image.

[0138] By converting the grayscale value of the second photovoltaic panel's dust accumulation image, the grayscale value of the second photovoltaic panel's dust accumulation image is obtained:

[0139]

[0140] in, Let (x, y) be the grayscale value of the image of the second photovoltaic panel with dust accumulation. The red channel value of the second photovoltaic panel's dust accumulation image at coordinates (x, y) is... The green channel value of the second photovoltaic panel's dust accumulation image at coordinates (x, y) is given. The value of the blue channel in the image of the second photovoltaic panel with dust accumulation is shown in coordinates (x, y).

[0141] Calculate the average grayscale value of the image of dust accumulation on the second photovoltaic panel:

[0142]

[0143] in, Let W be the average grayscale value of the second photovoltaic panel's dust accumulation image, W be the width of the second photovoltaic panel's dust accumulation image, and H be the height of the second photovoltaic panel's dust accumulation image. The grayscale value of the second photovoltaic panel's dust accumulation image at coordinates (x, y).

[0144] The diffuse reflection standard plate image is converted to grayscale, and the average grayscale value is calculated to obtain the average grayscale value of the diffuse reflection standard plate. .

[0145] The image grayscale ratio is calculated by combining the average grayscale value of the second photovoltaic panel and the average grayscale value of the diffuse reflection standard panel:

[0146]

[0147] in, Grayscale ratio, The average grayscale value of the image showing dust accumulation on the second photovoltaic panel. This represents the average grayscale value of the diffuse reflection standard plate.

[0148] It should be noted that the grayscale ratio is calculated by comparing the average grayscale value of the second photovoltaic panel with the average grayscale value of the diffuse reflection standard panel. The calculation of the grayscale ratio aims to eliminate the interference of ambient light intensity on the grayscale measurement. The diffuse reflection standard panel is based on a fixed reflectance (20.0% ± 0.5%), ensuring that its average grayscale value only varies with light intensity. Under a certain light intensity, an increase in dust density leads to a decrease in the reflectance of the photovoltaic panel surface, thus reducing the average grayscale value of the dust accumulation image of the second photovoltaic panel. Since the average grayscale value of the diffuse reflection standard panel remains constant under a given light intensity, the grayscale ratio will ultimately decrease accordingly.

[0149] Step S4: Input the second photovoltaic panel's dust accumulation image, image grayscale ratio, and dust accumulation density into the improved DenseNet169 network model for training to obtain the trained DenseNet169 network model; collect the ambient light intensity and clean power generation of the photovoltaic panel in the clean state, and perform curve fitting to obtain the light intensity-power generation curve in the clean state.

[0150] The image of the second photovoltaic panel covered in dust (independent variable) and the normalized grayscale ratio (independent variable) are used as inputs to the main and auxiliary channels of the improved DenseNet169 network model, while the dust density (dependent variable) is used as the output layer of the improved DenseNet169 network model. The dataset size is expanded by applying methods such as random flipping, rotation, simulation of different lighting conditions, affine transformation, and Gaussian blur to the image of the second photovoltaic panel covered in dust.

[0151] The improved DenseNet169 network model includes: a dual-channel input interface, a backbone network, auxiliary branches, and a gated residual fusion module.

[0152] Dual-channel input interface: The main channel receives a 224×224 pixel image of the second photovoltaic panel with accumulated dust (with grid line noise removed by Scharr operator and gradient weighted filtering); the auxiliary channel receives the grayscale ratio, i.e., the average grayscale value of the second photovoltaic panel image with accumulated dust and the average grayscale value of the diffuse reflection standard panel.

[0153] The backbone network consists of an initial convolutional layer, a max-pooling layer, four dense blocks, and three transition layers. The initial convolutional layer uses a 7x7 kernel with a stride of 2; the max-pooling layer uses a 3x3 convolutional layer with a stride of 2.

[0154] Four dense blocks:

[0155] The first dense block contains six fully connected convolutional layers. Each layer uses a sequence of "batch normalization - ReLU activation function - 1*1 convolution (128 channels) - batch normalization - ReLU activation function - 3*3 convolution (32 channels)". The 3*3 convolution in the third layer is replaced with an ACGBlock module, which expands the output channels to 256 through a dense connection mechanism. The first dense block is followed by the first transition layer. The first transition layer compresses the channels to 128 through 1*1 convolution and uses 2*2 average pooling to achieve spatial downsampling, outputting a 28*28*128 feature map.

[0156] The ACGBlock module replaces the 1x3 and 3x1 convolutions in the asymmetric convolution module with grouped convolutions; secondly, it replaces the 3x3 convolutions in the original asymmetric convolution module with conditional convolutions of the same size.

[0157] The second dense block contains 12 fully connected convolutional layers. Each layer also uses the sequence "batch normalization - ReLU activation function - 1*1 convolution (128 channels) - batch normalization - ReLU activation function - 3*3 convolution (32 channels)," but the 3*3 convolutions in layers 5 and 9 are replaced with ACGBlock modules. The second dense block is followed by a second transition layer. This transition layer compresses the channels to 256 using 1*1 convolutions and performs spatial downsampling using 2*2 average pooling, outputting a 14*14*256 feature map.

[0158] The third dense block contains 32 fully connected convolutional layers. Each layer also employs the sequence of "batch normalization - ReLU activation function - 1*1 convolution (128 channels) - batch normalization - ReLU activation function - 3*3 convolution (32 channels)". In layers 8, 16, and 24, the 3*3 convolutions are replaced with ACGBlock modules. The third dense block is followed by a third transition layer. This transition layer compresses the channels to 256 using 1*1 convolutions and performs spatial downsampling using 2*2 average pooling, outputting a 7*7*512 feature map.

[0159] The fourth dense block contains 32 fully connected convolutional layers. Each layer also uses the sequence of "batch normalization-ReLU activation function-1*1 convolution (128 channels)-batch normalization-ReLU activation function-3*3 convolution (32 channels)," but the 3*3 convolutions in the 11th and 12th layers are replaced with ACGBlock modules.

[0160] An EMS module is embedded after the fourth dense block. The EMS module includes average pooling (1*1 convolution with 512 channels) and standard deviation pooling (1*1 convolution with 512 channels). Finally, a 1024-dimensional visual feature vector is output through the Sigmoid function and spatial weighted aggregation.

[0161] Auxiliary branch: The normalized grayscale ratio is converted into a 16-dimensional feature vector through a two-layer fully connected network. The ReLU activation function is used to enhance the non-linear expressive power to obtain an auxiliary feature vector.

[0162] Gated residual fusion module: A dynamic gating mechanism is used to establish cross-modal association. The auxiliary feature vector is input into the Sigmoid function to generate 1024-dimensional gating weights. The visual feature vector is multiplied element by element with the gating weights. After transformation by a 128-dimensional fully connected layer, a 128-dimensional fused feature vector is output.

[0163] Regression Output Layer: The fused feature vectors are non-linearly mapped through a 128-unit fully connected layer. Finally, linearly activated neurons output predicted ash density values. Based on these predicted values ​​and the input ash density, the improved DenseNet169 network model is trained using the cross-entropy loss function and optimized using the Ranger optimizer (initial learning rate 0.001), resulting in the trained DenseNet169 network model.

[0164] The ambient light intensity and clean power generation of photovoltaic (PV) panels under clean conditions were collected, and curve fitting was performed to obtain the light intensity-power generation curve under clean conditions. The specific steps are as follows: The "clean condition" of the PV panels refers to a clean state without dust accumulation, ensuring that there is no dust adhering to the PV panel surface when data is collected. A sufficiently long time period was selected (covering different seasons, weather conditions, and time periods). While the PV panels were in a clean state, the ambient light intensity (collected in real time through a light sensor) and the corresponding clean power generation (extracted from the PV power generation monitoring system) were recorded simultaneously, ensuring that the timestamps of each set of data were consistent and establishing a one-to-one correspondence. The collected clean power generation data was preprocessed to remove outliers caused by equipment failure, measurement errors, etc., retaining valid data samples.

[0165] The light intensity-power generation curve under clean conditions was obtained by curve fitting using the least squares method with ambient light intensity as the independent variable and clean power generation as the dependent variable.

[0166]

[0167] Where P represents clean power generation, GZ represents ambient light intensity, and a, b, and c are the parameters of the light-power generation curve under clean conditions. The temperature coefficient is determined by the material of the photovoltaic panel. For ambient temperature, The reference temperature is 25°C by default.

[0168] Step S5: Collect the image of the second photovoltaic panel with dust accumulation and the image grayscale ratio in real time and input it into the trained DenseNet169 network model to output the predicted value of dust accumulation density; collect the ambient light intensity and power generation of the photovoltaic panel in real time, input the ambient light intensity into the light-power generation curve under clean conditions, and output the clean power generation; calculate the difference between the clean power generation and the power generation to obtain the photovoltaic panel power generation coefficient.

[0169] The system collects real-time images of dust accumulation on the second photovoltaic panel and their grayscale ratios, then inputs these images into a trained DenseNet169 network model to output a predicted dust accumulation density value.

[0170]

[0171] in, To predict ash accumulation density, DenseNet169() is the input function for the DenseNet169 model. This is an image of dust accumulation on the second photovoltaic panel. This represents the image grayscale ratio.

[0172] The ambient light intensity, ambient temperature, and actual power generation of the photovoltaic panels are collected in real time within a preset time window T. The ambient light intensity is then input into the light intensity-power generation curve under clean conditions to obtain the clean power generation.

[0173]

[0174] in, Let be the clean power generation capacity at time t. Let be the ambient light intensity at time t, and a, b, and c be the parameters of the light intensity-historical clean power generation curve. The temperature coefficient is determined by the material of the photovoltaic panel. Let be the ambient temperature at time t. The reference temperature is 25°C by default.

[0175] The difference between the clean power generation and the actual power generation is calculated to obtain the photovoltaic power generation coefficient:

[0176]

[0177] Where C is the photovoltaic power generation coefficient, T is the length of the preset time window, and t is time t. Let be the clean power generation capacity at time t. Let t be the actual power generation at time t.

[0178] It should be noted that the calculation of the photovoltaic panel's power generation coefficient C quantifies the power generation efficiency loss caused by dust obstruction. The coefficient is based on the actual power generation within a preset time window T. The average ratio of C to clean power generation P(t) (calculated from the light intensity-historical cleanliness curve) isolates the influence of ambient light intensity fluctuations. For example, a C value of 0.85 indicates that dust causes a 15% power generation loss (since the actual power is only 85% of the theoretical clean power generation), thus directly reflecting the actual damage to power generation performance caused by dust accumulation and providing a key basis for subsequent dust index performance loss.

[0179] Step S6: By combining the predicted dust density value and the photovoltaic power generation coefficient, the photovoltaic panel dust index is calculated. The photovoltaic panel dust index is compared with a preset threshold and a risk level is established to achieve real-time monitoring of dust.

[0180] By combining the predicted dust density value and the photovoltaic panel power generation coefficient, the photovoltaic panel dust index is calculated:

[0181]

[0182] Wherein, DI is the dust index of the photovoltaic panel. To predict ash accumulation density, Here is a reference value for dust density, and C is the photovoltaic power generation coefficient. This is the adjustment coefficient.

[0183] It should be noted that the dust accumulation density reference value The default value is 10 grams per square meter, and the dust accumulation density reference value can be dynamically adjusted experimentally according to the type of photovoltaic panel; adjustment coefficient. The default value is 10, and the adjustment coefficient can be dynamically adjusted according to the operation and maintenance costs of photovoltaic power plants and the tolerance for power generation losses. .

[0184] It should be noted that the photovoltaic panel dust index, through a composite indicator, achieves a comprehensive quantitative assessment of the impact of dust, thereby supporting real-time monitoring and risk decision-making. The normalized dust density term... The predicted dust density is scaled relative to a reference value (default 10 g / m²) to ensure the value falls between 0 and 1. The square root operation smooths the data, preventing exponential spikes caused by extremely high dust density values ​​and making the indicator more stable. For example, when... near The square root function slows down the growth rate, reflecting the nonlinear effect of dust accumulation. Performance loss term 1−C: C is the photovoltaic power generation coefficient, representing the ratio of actual power generation to theoretical power generation under clean conditions (C close to 1 indicates high efficiency, close to 0 indicates severe loss). 1−C quantifies the proportion of power generation efficiency loss due to dust (e.g., C=0.9 represents a 10% loss). The square root is also used to smooth performance loss data, ensuring that power generation fluctuations do not excessively amplify the exponent, while highlighting the actual operational impact. The two factors, dust density and power generation loss, are multiplied together and integrated into a single synergistic index. This index reflects the combined effect of the physical presence of dust (from image analysis) and the actual performance degradation (from power generation data), avoiding the limitations of a single index. (Adjustment coefficient) (Default is 10) is used to amplify the results, scaling the DI value to a range where thresholds and risk levels can be easily set (e.g., 0-10 or higher). For example: when When C is 5 and C is 0.9, The dust index of the photovoltaic panel is 2.236, which is relatively low.

[0185] In real-time monitoring applications, the preset dust index threshold system will automatically trigger a graded response: when the DI value is below 2, it is judged as low risk, only data is recorded and no alarm is triggered; when the DI value is in the range of 2 to 5, it is marked as medium risk, and an operation and maintenance work order is automatically generated and pushed to the monitoring screen, prompting that an inspection or planned cleaning be arranged within the next 72 hours; when the DI value reaches or exceeds 5, it is immediately upgraded to high risk, and an emergency SMS alarm is sent to the mobile phone of the operation and maintenance manager, and the site's automated cleaning device is linked to start the emergency dust removal program (if there is no automated equipment, the manual cleaning task is forcibly placed at the top), while dust data is collected every 30 minutes to track the deterioration trend.

[0186] This paper proposes a real-time dust monitoring method for intelligent operation and maintenance of photovoltaic power plants. By installing a diffuse reflection standard plate on the photovoltaic panels to collect images under different dust densities, the gradient magnitude is calculated using the Scharr operator and denoised using a gradient weighted filtering method. The grayscale ratio is obtained by combining the average grayscale values ​​of the photovoltaic panels and the diffuse reflection standard plate. An improved DenseNet169 network model is used to train and predict the dust density. At the same time, the power generation coefficient is calculated by combining the irradiance-historical clean power generation curve and real-time power generation data. Finally, the dust index is obtained by combining the predicted dust density value and the power generation coefficient, and a risk level is established, so as to achieve real-time and accurate monitoring of dust on photovoltaic panels.

[0187] Noise removal from the first photovoltaic panel's dust-accumulated image is achieved using gradient-weighted filtering based on gradient magnitude. The core significance lies in accurately distinguishing between uniform dust accumulation areas and grid line noise areas. When the gradient change between adjacent pixels is gradual, the original pixel value is retained to avoid blurring the dust accumulation texture. When there are significant gradient abrupt changes, gradient-weighted interpolation is used to suppress grid line noise, effectively eliminating the interference of the silver grid lines on the photovoltaic panel surface on image analysis. This provides a high-quality image data foundation for subsequent grayscale calculations and model training.

[0188] The key significance of calculating the image grayscale ratio by combining the average grayscale value of the second photovoltaic panel and the average grayscale value of the diffuse reflection standard panel lies in eliminating the influence of ambient light intensity fluctuations. The diffuse reflection standard panel has a constant reflectivity, and its average grayscale value only changes with illumination. However, the average grayscale value of the photovoltaic panel is affected by both dust accumulation and illumination. The ratio of the two can cancel out illumination interference and only reflect the change in reflectivity of the photovoltaic panel surface caused by dust accumulation, providing a stable characteristic parameter for accurate assessment of dust accumulation density.

[0189] The significance of calculating the photovoltaic panel dust index by combining the predicted dust density and the photovoltaic panel power generation coefficient lies in achieving a comprehensive quantitative assessment of the impact of dust. This index reflects both the physical density of dust accumulation (reflecting the degree of dust accumulation) and the power generation coefficient (reflecting the actual impact of dust on power generation efficiency), avoiding the one-sidedness of a single indicator. By comparing with preset thresholds to establish risk levels, different operation and maintenance responses can be precisely triggered, providing a scientific and timely basis for dust monitoring and operation and maintenance decisions for photovoltaic power plants.

[0190] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, the phrase "comprising an element defined as..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0191] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for real-time dust monitoring in intelligent operation and maintenance of photovoltaic power plants, characterized in that: Includes the following steps: Step S1: Install a diffuse reflection standard plate on the photovoltaic panel and collect images of the photovoltaic panel and the diffuse reflection standard plate under different dust accumulation densities to obtain the first dust accumulation image of the photovoltaic panel and the diffuse reflection standard plate image; Step S2: Calculate the gradient magnitude of the first photovoltaic panel dust accumulation image using the Scharr operator. Based on the gradient magnitude, remove noise from the first photovoltaic panel dust accumulation image using a gradient weighted filtering method to obtain the second photovoltaic panel dust accumulation image. Step S3: The grayscale ratio of the images is calculated by comparing the grayscale values ​​of the second photovoltaic panel dust accumulation image and the diffuse reflection standard plate image; Step S4: Input the second photovoltaic panel dust accumulation image, image grayscale ratio, and dust accumulation density into the improved DenseNet169 network model for training to obtain the trained DenseNet169 network model; collect the ambient light intensity and clean power generation of the photovoltaic panel in the clean state, and perform curve fitting to obtain the light intensity-power generation curve in the clean state. Step S5: Collect the image of the second photovoltaic panel with dust accumulation and the image grayscale ratio in real time and input it into the trained DenseNet169 network model to output the predicted value of dust accumulation density; collect the ambient light intensity and power generation of the photovoltaic panel in real time, input the ambient light intensity into the light intensity-power generation curve under clean conditions, and output the clean power generation; calculate the difference between the clean power generation and the power generation to obtain the photovoltaic panel power generation coefficient. Step S6: By combining the predicted dust density value and the photovoltaic power generation coefficient, the photovoltaic panel dust index is calculated. The photovoltaic panel dust index is compared with a preset threshold and a risk level is established to achieve real-time monitoring of dust.

2. The method for real-time dust monitoring in intelligent operation and maintenance of photovoltaic power plants according to claim 1, characterized in that: The process of installing a diffuse reflection standard plate on the photovoltaic panel and collecting images of the photovoltaic panel and the diffuse reflection standard plate under different dust accumulation densities to obtain a first dust accumulation image of the photovoltaic panel and an image of the diffuse reflection standard plate includes the following specific steps: The photovoltaic panel is laid flat on the ground with its front side facing up. A sealed plexiglass cover of the same size is placed on the photovoltaic panel to provide a closed and controllable experimental environment. A diffuse reflection standard plate is installed at a fixed position on each photovoltaic panel to be monitored. The diffuse reflection standard plate is closely attached to the outside of the photovoltaic panel frame and is parallel to the photovoltaic panel surface and within the same imaging angle range. Use a blower to continuously blow air through the only air inlet on the side of the plexiglass cover, carrying dust particles, until the dust particles are completely lifted in the sealed container; then stop; let the photovoltaic panel stand for 10 hours in a dry environment to allow the dust to settle naturally and adhere evenly to the surface of the photovoltaic panel; repeat the above steps to evenly sprinkle dust particles with mass densities of 0, 5, 10, 15, 20, 25 and 30 g / m² on the surface of photovoltaic panels of the same size. After the dust accumulation is completed, the plexiglass cover is removed, and the camera position is calibrated so that a single frame of the camera simultaneously and completely covers the photovoltaic panel and the attached diffuse reflection standard panel. Based on the template matching algorithm of SIFT feature points, the diffuse reflection standard panel area is automatically located when the photovoltaic panel image is acquired, and the located diffuse reflection standard panel area is cropped to obtain the diffuse reflection standard panel image. The cropped image is the first photovoltaic panel dust accumulation image. Finally, the first photovoltaic panel dust accumulation image and the diffuse reflection standard panel image under different dust accumulation densities are obtained.

3. The method for real-time dust monitoring in intelligent operation and maintenance of photovoltaic power plants according to claim 2, characterized in that: The calculation of the gradient magnitude of the first photovoltaic panel dust accumulation image using the Scharr operator includes the following steps: The gradient magnitude of the dust accumulation image of the first photovoltaic panel is calculated using the Scharr operator, and based on the gradient magnitude: ； in, Let be the gradient value of the dust accumulation image of the first photovoltaic panel in the horizontal direction at coordinates (x, y). Let be the gradient value of the dust accumulation image of the first photovoltaic panel in the vertical direction at coordinates (x, y). Let (x, y) be the pixel value of the first photovoltaic panel's dust accumulation image at coordinates (x, y). For the horizontal gradient operator of the Scharr operator, The vertical gradient operator for the Scharr operator; Calculate the gradient magnitude of the first photovoltaic panel's dust accumulation image based on the horizontal and vertical gradient values: ； in, Let be the gradient magnitude of the dust accumulation image of the first photovoltaic panel at coordinates (x, y). Let be the gradient value of the dust accumulation image of the first photovoltaic panel in the horizontal direction at coordinates (x, y). The gradient value of the first photovoltaic panel's dust accumulation image in the vertical direction at coordinates (x, y).

4. The method for real-time dust monitoring in intelligent operation and maintenance of photovoltaic power plants according to claim 3, characterized in that: The step of removing noise from the first photovoltaic panel dust accumulation image based on the gradient magnitude and using a gradient-weighted filtering method to obtain the second photovoltaic panel dust accumulation image includes the following steps: For each point (x, y) on the first photovoltaic panel dust accumulation image, calculate the vertical relative gradient intensity of point (x, y): ； in, Let be the vertical relative gradient intensity at point (x, y). Let be the gradient magnitude of the dust accumulation image of the first photovoltaic panel at coordinates (x, y+1). The gradient magnitude of the dust accumulation image of the first photovoltaic panel at coordinates (x, y); For each point (x, y) on the first photovoltaic panel dust accumulation image, calculate the horizontal relative gradient intensity of point (x, y): ； in, Let (x, y) be the horizontal relative gradient intensity. Let be the gradient magnitude of the dust accumulation image of the first photovoltaic panel at coordinates (x+1, y). The gradient magnitude of the dust accumulation image of the first photovoltaic panel at coordinates (x, y); Calculate the symmetrical gradient change rate of all adjacent pixels in the vertical and horizontal directions in the first photovoltaic panel dust accumulation image: ； ； in, Characteristic points The degree of gradient abruptness in the vertical direction; the larger the value, the more drastic the gradient change between (x,y) and the point below it (x,y+1). The gradient abrupt change at the point (x, y) in the horizontal direction represents the degree of gradient change; the larger the value, the stronger the gradient. And its right point The more drastic the gradient change between them; calculate , The degree of abrupt change in the total directional gradient is given by the degree of abrupt change in the total directional gradient in the image of the first photovoltaic panel with dust accumulation. The total directional gradient abrupt change in the first photovoltaic panel's dust accumulation image is sorted in ascending order, and its 90th percentile is taken. Calculate the basic threshold , ; Introducing tolerance coefficient Generate dynamic thresholds: ； Where d is the gradient intensity threshold, This is the tolerance factor, which defaults to 1.

1. Basic threshold; By comparing the relative gradient strength and the gradient strength threshold, when and When the gradient change between two pixels is gradual, the pixel (x, y) remains unchanged; when the horizontal direction is... Or the vertical direction is If the gradient change between two pixels is significant, then a gradient-weighted filtering method is applied to the pixel at coordinates (x, y). For the 3x3 neighborhood window of the pixel at coordinates (x, y), the gradient weight between each neighboring point in the 3x3 neighborhood window and the pixel at coordinates (x, y) is calculated. ； in, Let be the gradient weight of the j-th neighboring pixel of the pixel with coordinates (x, y). Let be the gradient magnitude of the j-th neighboring pixel of the pixel with coordinates (x, y). To prevent division by zero, the value is taken as... ; Based on the gradient magnitude, noise removal is performed on the first photovoltaic panel dust accumulation image using a gradient weighted filtering method: when and In this case, the pixels are preserved: = ; when or hour: ； in, The image of dust accumulation on the second photovoltaic panel in coordinates The pixel value below, where j is the coordinate. The index of the j-th neighboring pixel, where m is the coordinate. The number of all neighboring pixels of a pixel. Let be the gradient weight of the j-th neighboring pixel of the pixel with coordinates (x, y). This represents the pixel value of the j-th neighboring pixel of the (x, y) pixel point in the image of the first photovoltaic panel with dust accumulation.

5. A real-time dust monitoring method for intelligent operation and maintenance of photovoltaic power plants according to claim 4, characterized in that: The step of calculating the image grayscale ratio by comparing the grayscale values ​​of the second photovoltaic panel's dust accumulation image and the diffuse reflection standard panel image includes the following steps: By converting the grayscale value of the second photovoltaic panel's dust accumulation image, the grayscale value of the second photovoltaic panel's dust accumulation image is obtained: ； in, Let (x, y) be the grayscale value of the image of the second photovoltaic panel with dust accumulation. The red channel value of the second photovoltaic panel's dust accumulation image at coordinates (x, y) is... The green channel value of the second photovoltaic panel's dust accumulation image at coordinates (x, y) is given. The blue channel value of the second photovoltaic panel's dust accumulation image at coordinates (x, y); Calculate the average grayscale value of the image of dust accumulation on the second photovoltaic panel: ； in, Let W be the average grayscale value of the second photovoltaic panel's dust accumulation image, W be the width of the second photovoltaic panel's dust accumulation image, and H be the height of the second photovoltaic panel's dust accumulation image. The image of dust accumulation on the second photovoltaic panel in coordinates The grayscale value below; The diffuse reflection standard plate image is converted to grayscale, and the average grayscale value is calculated to obtain the average grayscale value of the diffuse reflection standard plate. ; The image grayscale ratio is calculated by combining the average grayscale value of the second photovoltaic panel and the average grayscale value of the diffuse reflection standard panel: ； in, Grayscale ratio, The average grayscale value of the image showing dust accumulation on the second photovoltaic panel. This represents the average grayscale value of the diffuse reflection standard plate.

6. The method for real-time dust monitoring in intelligent operation and maintenance of photovoltaic power plants according to claim 5, characterized in that: The step of inputting the second photovoltaic panel's dust accumulation image, image grayscale ratio, and dust accumulation density into the improved DenseNet169 network model for training to obtain the trained DenseNet169 network model includes the following specific steps: The image of the second photovoltaic panel with accumulated dust and the normalized grayscale ratio were used as the input to the main and auxiliary channels of the improved DenseNet169 network model, and the dust accumulation density was used as the output layer of the improved DenseNet169 network model. Random flipping, rotation, different lighting simulation, affine transformation, and Gaussian blur were applied to the image of the second photovoltaic panel with accumulated dust to expand the dataset size. The improved DenseNet169 network model includes: a dual-channel input interface, a backbone network, auxiliary branches, and a gated residual fusion module; Dual-channel input interface: The main channel receives a 224×224 pixel image of the second photovoltaic panel's dust accumulation; the auxiliary channel receives the grayscale ratio. The backbone network consists of an initial convolutional layer, a max pooling layer, four dense blocks, and three transition layers. The initial convolutional layer uses a 7*7 convolutional kernel with a stride of 2; the max pooling layer uses a 3*3 convolutional layer with a stride of 2. Four dense blocks: The first dense block contains a structure of 6 fully connected convolutional layers. Each layer is processed by the sequence "batch normalization-ReLU activation function-1*1 convolution-batch normalization-ReLU activation function-3*3 convolution". The 3*3 convolution in the third layer is replaced by an ACGBlock module. The output channels are expanded to 256 through a dense connection mechanism. The first dense block is followed by the first transition layer. The first transition layer compresses the channels to 128 through 1*1 convolution and uses 2*2 average pooling to achieve spatial downsampling, outputting a 28*28*128 feature map. The ACGBlock module replaces the 1x3 and 3x1 convolutions in the asymmetric convolution module with grouped convolutions; secondly, it replaces the 3x3 convolutions in the original asymmetric convolution module with conditional convolutions of the same size. The second dense block contains 12 fully connected convolutional layers. Each layer also uses the sequence "batch normalization-ReLU activation function-1*1 convolution-batch normalization-ReLU activation function-3*3 convolution", but the 3*3 convolutions in layers 5 and 9 are replaced with ACGBlock modules. The second dense block is followed by a second transition layer. The second transition layer compresses the channels to 256 through 1*1 convolution and uses 2*2 average pooling to achieve spatial downsampling, outputting a 14*14*256 feature map. The third dense block contains 32 fully connected convolutional layers. Each layer also uses the sequence of "batch normalization-ReLU activation function-1*1 convolution-batch normalization-ReLU activation function-3*3 convolution". In layers 8, 16 and 24, the 3*3 convolutions are replaced with ACGBlock modules. The third dense block is followed by the third transition layer. The third transition layer compresses the channels to 256 through 1*1 convolution and uses 2*2 average pooling to achieve spatial downsampling, outputting a 7*7*512 feature map. The fourth dense block contains 32 fully connected convolutional layers. Each layer also uses the sequence of "batch normalization-ReLU activation function-1*1 convolution-batch normalization-ReLU activation function-3*3 convolution", but the 3*3 convolutions in the 11th and 12th layers are replaced with ACGBlock modules. An EMS module is embedded after the fourth dense block. The EMS module includes average pooling and standard deviation pooling. Finally, a 1024-dimensional visual feature vector is output through the Sigmoid function and spatial weighted aggregation. Auxiliary branch: The normalized grayscale ratio is converted into a 16-dimensional feature vector through a two-layer fully connected network. The ReLU activation function is used to enhance the non-linear expressive power to obtain an auxiliary feature vector. Gated residual fusion module: A dynamic gating mechanism is used to establish cross-modal association. The auxiliary feature vector is input into the Sigmoid function to generate 1024-dimensional gating weights. The visual feature vector is multiplied element by element with the gating weights. After transformation by a 128-dimensional fully connected layer, a 128-dimensional fused feature vector is output. Regression output layer: The fused feature vectors are non-linearly mapped through a 128-unit fully connected layer; finally, the predicted ash density is output by a linearly activated neuron. Based on the predicted ash density and the input ash density, the improved DenseNet169 network model is trained using the cross-entropy loss function and optimized using the Ranger optimizer to obtain the trained DenseNet169 network model.

7. A real-time dust monitoring method for intelligent operation and maintenance of photovoltaic power plants according to claim 6, characterized in that: The process of collecting the ambient light intensity and clean power generation of the photovoltaic panel under clean conditions, and performing curve fitting to obtain the light intensity-power generation curve under clean conditions includes the following specific steps: Collect the ambient light intensity and corresponding clean power generation when the photovoltaic panels are in a clean state, ensuring that the timestamps of each set of data are consistent and establishing a one-to-one correspondence; preprocess the collected clean power generation data to remove outliers caused by equipment failure, measurement errors, etc., and retain valid data samples. The light intensity-power generation curve under clean conditions was obtained by curve fitting using the least squares method with ambient light intensity as the independent variable and clean power generation as the dependent variable. ； Where P represents clean power generation, GZ represents ambient light intensity, and a, b, and c are the parameters of the light-power generation curve under clean conditions. The temperature coefficient is determined by the material of the photovoltaic panel. For ambient temperature, The reference temperature is 25°C by default.

8. A real-time dust monitoring method for intelligent operation and maintenance of photovoltaic power plants according to claim 7, characterized in that: The process of collecting real-time images of the second photovoltaic panel's accumulated dust and their grayscale ratios, inputting them into a trained DenseNet169 network model, and outputting a predicted dust density value includes the following specific steps: The system collects real-time images of dust accumulation on the second photovoltaic panel and their grayscale ratios, then inputs these images into a trained DenseNet169 network model to output a predicted dust accumulation density value. ； in, To predict ash accumulation density, DenseNet169() is the input function for the DenseNet169 model. This is an image of dust accumulation on the second photovoltaic panel. This represents the image grayscale ratio.

9. A real-time dust monitoring method for intelligent operation and maintenance of photovoltaic power plants according to claim 8, characterized in that: The process of collecting real-time ambient light intensity and power generation of the photovoltaic panel, inputting the ambient light intensity into the light intensity-power generation curve under clean conditions, outputting the clean power generation, and calculating the difference between the clean power generation and the total power generation to obtain the photovoltaic panel's power generation coefficient includes the following specific steps: The ambient light intensity, ambient temperature, and power generation of the photovoltaic panels are collected in real time within a preset time window T. The ambient light intensity is then input into the light intensity-power generation curve under clean conditions to obtain the clean power generation. ； in, Let be the clean power generation capacity at time t. Let be the ambient light intensity at time t, and a, b, and c be the parameters of the light intensity-historical clean power generation curve. The temperature coefficient is determined by the material of the photovoltaic panel. Let be the ambient temperature at time t. For reference temperature, the default value is 25°C; The difference between the clean power generation and the actual power generation is calculated to obtain the photovoltaic power generation coefficient: ； Where C is the photovoltaic power generation coefficient, T is the length of the preset time window, and t is time t. Let be the clean power generation capacity at time t. Let t be the actual power generation at time t.

10. A method for real-time dust monitoring in intelligent operation and maintenance of photovoltaic power plants according to claim 9, characterized in that: The process of calculating the photovoltaic panel dust index by combining the predicted dust density value and the photovoltaic panel power generation coefficient includes the following specific steps: By combining the predicted dust density value and the photovoltaic panel power generation coefficient, the photovoltaic panel dust index is calculated: ； Wherein, DI is the dust index of the photovoltaic panel. To predict ash accumulation density, Here is a reference value for dust density, and C is the photovoltaic power generation coefficient. This is the adjustment coefficient.

Images