Intelligent cleaning decision-making method and system for photovoltaic power station based on machine learning

Abstract

The invention provides an intelligent cleaning decision-making method and system for a photovoltaic power station based on machine learning, and relates to the technical field of photovoltaic power stations, and the method comprises the steps: collecting images through a visible light camera and an infrared thermal imaging camera carried by an unmanned plane, segmenting a pollution region based on color, texture and temperature features through a deep learning model after image registration and preprocessing, and obtaining a cleaning decision-making result; and calculating a pollution area proportion and establishing an evaluation index in combination with the temperature difference value so as to analyze the power generation efficiency loss and generate a pollution distribution map and an evaluation report. The pollution area can be accurately identified, the influence of the pollution degree on the power generation efficiency is quantitatively evaluated, and a decision basis is provided for power station cleaning maintenance.

Description

Technical Field

[0001] The present invention relates to the technology of photovoltaic power stations, and particularly to an intelligent cleaning decision-making method and system for photovoltaic power stations based on machine learning. Background Art

[0002] With the global energy transition and the rapid development of renewable energy, as an important part of clean energy, photovoltaic power generation has been widely used globally. Photovoltaic power stations usually consist of a large number of photovoltaic modules, which are exposed to the outdoor environment and are affected by various pollutants such as dust, dirt, sand, bird droppings, leaves, and industrial emissions for a long time. The pollution on the surface of photovoltaic modules will directly lead to a decrease in the absorption and transmittance of sunlight, thereby reducing the power generation efficiency and economic benefits of photovoltaic power stations.

[0003] Traditional maintenance and detection of photovoltaic power stations mainly rely on manual inspections, and workers need to regularly observe and evaluate a large area of photovoltaic modules with the naked eye. With the development of technology, ground monitoring devices such as fixed cameras and temperature measurement devices have been introduced into the detection system, but they still face problems such as limited coverage and poor real-time performance. In recent years, the rapid development of drone technology has provided a new means for the monitoring and evaluation of photovoltaic power stations. Drones equipped with various sensors can quickly obtain the surface information of photovoltaic modules from a high-altitude perspective, and combined with computer vision and deep learning algorithms, it provides a more efficient and accurate solution for the assessment of the pollution degree of photovoltaic power stations.

[0004] The existing pollution degree assessment technologies for photovoltaic power stations have limitations in many aspects, which affect the accuracy and practicality of the assessment results.

[0005] Firstly, the existing assessment methods often only rely on a single type of data source, mainly visible light images or power generation data of the power station. This reliance on a single data source cannot comprehensively reflect the nature and degree of pollution, especially the different impacts of different types of pollutants on power generation efficiency. For example, some transparent pollutants are difficult to be identified under visible light, but they will have a significant impact on the thermal characteristics of the module, resulting in inaccurate assessment results.

[0006] Secondly, traditional assessment methods lack an accurate modeling of the relationship between the pollution degree and the power generation efficiency loss. Existing technologies usually use simple linear models or empirical formulas to estimate the efficiency loss caused by pollution, while ignoring the complex impacts of different pollution types, different pollution distributions, and environmental factors on power generation efficiency, and cannot provide an accurate basis for priority ranking for cleaning and maintenance decisions.

[0007] Third, the prior art lacks a systematic mechanism for visualizing pollution distribution and generating assessment reports. Most assessment systems can only provide preliminary pollution detection results and cannot generate intuitive pollution distribution maps and detailed analysis reports on efficiency losses. This makes it difficult for maintenance personnel to quickly identify problem areas and formulate targeted cleaning plans, resulting in low maintenance efficiency and waste of resources. Summary of the Invention

[0008] Embodiments of the present invention provide an intelligent cleaning decision-making method and system for a photovoltaic power station based on machine learning, which can solve the problems in the prior art.

[0009] In the first aspect of the embodiments of the present invention,

[0010] An intelligent cleaning decision-making method for a photovoltaic power station based on machine learning is provided, including:

[0011] Collect image information of the photovoltaic power station by using an unmanned aerial vehicle (UAV). The UAV is equipped with a visible light camera and an infrared thermal imaging camera. The visible light camera is used to collect a first visible light image of the surface of the photovoltaic module, and the infrared thermal imaging camera is used to collect a first temperature distribution image of the surface of the photovoltaic module. The UAV flies according to a preset flight path, and the preset flight path is generated based on the layout information of the photovoltaic power station, so as to obtain a first visible light image set and a first temperature distribution image set of the photovoltaic power station;

[0012] Perform image registration and preprocessing on the first visible light image set and the first temperature distribution image set to obtain a second visible light image set and a second temperature distribution image set. Analyze the second visible light image set and the second temperature distribution image set through a deep learning model, segment the pollution area based on the color feature, texture feature and temperature feature of the surface of the photovoltaic module, and obtain a pollution area segmentation result. Calculate the pollution area ratio of each photovoltaic module according to the pollution area segmentation result, and combine the temperature difference between the pollution area and the non-pollution area in the second temperature distribution image set to establish a pollution degree assessment index;

[0013] Evaluate the power generation efficiency loss of the photovoltaic power station based on the pollution degree assessment index and the pollution area ratio. By comparing the actual power generation data of the pollution area and the non-pollution area in the pollution area segmentation result, establish a mapping relationship between the pollution degree and the power generation efficiency loss. Generate a pollution distribution map and a power generation efficiency loss assessment report of the photovoltaic power station according to the mapping relationship. The pollution distribution map identifies areas with different pollution degrees, and the power generation efficiency loss assessment report includes the pollution degree assessment index of each area, the temperature difference and the predicted value of the power generation efficiency loss.

[0014] Perform image registration and preprocessing on the first visible light image set and the first temperature distribution image set to obtain a second visible light image set and a second temperature distribution image set, including:

[0015] Establish an initial mapping relationship between the first visible light image set and the first temperature distribution image set; extract feature points for the first visible light image set and the first temperature distribution image set, perform multi-scale decomposition on the first visible light image set and the first temperature distribution image set using a difference-of-Gaussians pyramid, detect local extreme points in the images after multi-scale decomposition, and generate feature descriptor vectors based on the local extreme points;

[0016] Use the nearest neighbor ratio matching strategy to perform initial matching on the feature descriptor vectors, establish the corresponding relationship of feature points, construct an affine transformation matrix according to the corresponding relationship of feature points, and optimize the parameters of the affine transformation matrix using the iterative weighted least squares method; perform geometric correction and radiometric correction on the first visible light image set and the first temperature distribution image set based on the optimized affine transformation matrix, where the geometric correction includes eliminating the distortion of the first visible light image set and the first temperature distribution image set, and the radiometric correction includes eliminating the uneven illumination;

[0017] Perform enhancement processing on the images after geometric correction and radiometric correction using the adaptive histogram equalization method, and perform denoising processing in combination with the bilateral filtering algorithm. The filtering kernel function of the bilateral filtering algorithm takes into account both the spatial distance weight and the gray similarity weight; perform registration accuracy evaluation and image quality evaluation on the enhanced and denoised images, and dynamically adjust the registration parameters according to the results of the registration accuracy evaluation and the image quality evaluation until the preset quality threshold is met to obtain a second visible light image set and a second temperature distribution image set.

[0018] Analyze the second visible light image set and the second temperature distribution image set through a deep learning model, and segment the contaminated area based on the color features, texture features, and temperature features on the surface of the photovoltaic module to obtain the segmentation result of the contaminated area, including:

[0019] Construct a multi-scale feature extraction network. The multi-scale feature extraction network includes a visible light image feature extraction branch and a thermal imaging image feature extraction branch. The visible light image feature extraction branch uses a pyramid structure to perform multi-scale decomposition on the second visible light image set to obtain a visible light multi-scale feature map, and the thermal imaging image feature extraction branch uses a pyramid structure to perform multi-scale decomposition on the second temperature distribution image set to obtain a temperature multi-scale feature map;

[0020] Based on the visible light multi-scale feature maps, convert the second set of visible light images from the RGB space to the CIELAB color space, calculate the color difference values in the CIELAB color space, and obtain a color feature map based on the weighted calculation of the brightness difference values and the chromaticity difference values; construct a gray-level co-occurrence matrix based on the visible light multi-scale feature maps and perform multi-directional multi-scale filtering, and extract the texture feature matrix of the second set of visible light images to obtain a texture feature map, where the texture feature matrix includes direction features and scale features;

[0021] Based on the temperature multi-scale feature maps, calculate the temperature gradient matrix of the second set of temperature distribution images to obtain a gradient feature map, where the temperature gradient matrix includes horizontal direction gradients and vertical direction gradients; construct a thermal anomaly region detection model based on the gradient feature map, and the thermal anomaly region detection model determines the anomaly region by calculating the difference degree between the temperature pixel values and the neighborhood temperature distribution to obtain a temperature anomaly feature map; construct a pollution region segmentation model; process the color feature map, the texture feature map, and the temperature anomaly feature map based on the pollution region segmentation model to obtain a pollution region segmentation result.

[0022] Constructing a pollution region segmentation model includes:

[0023] Construct a dual attention enhancement module. The dual attention enhancement module receives the color feature map, the texture feature map, and the temperature anomaly feature map as inputs, and performs feature enhancement through a regional spatial attention sub-module and a multi-channel feature attention sub-module. The regional spatial attention sub-module includes a deformable convolutional layer and a spatial weight calculation layer. The deformable convolutional layer adaptively adjusts the convolution sampling positions using learnable offsets, and the spatial weight calculation layer generates a two-dimensional spatial weight matrix based on the local region response; the multi-channel feature attention sub-module includes a global pooling layer and a channel mapping layer. The global pooling layer extracts channel-level feature statistical information, and the channel mapping layer generates a channel weight vector through non-linear transformation; multiply the two-dimensional spatial weight matrix element-wise with the input feature map in the spatial dimension, and multiply the channel weight vector with the feature map weighted in the spatial dimension channel-wise in the channel dimension to obtain an attention enhanced feature map;

[0024] Construct an adaptive feature fusion network. The adaptive feature fusion network receives the attention enhanced feature map as an input, and includes a feature correlation calculation unit and a dynamic weight assignment unit. The feature correlation calculation unit calculates the complementary degree between different features in the attention enhanced feature map based on cosine similarity to obtain a feature correlation matrix. The dynamic weight assignment unit generates an adaptive weight coefficient according to the feature correlation matrix, and fuses different features by weighted summation to obtain a fused feature map;

[0025] Construct a multi-level loss calculation module. The multi-level loss calculation module receives the fused feature map as input, and includes a region segmentation loss calculation unit, a boundary refinement loss calculation unit, and a topological structure preservation loss calculation unit. The region segmentation loss calculation unit calculates the category prediction error using a weighted cross-entropy function to obtain a segmentation loss value. The boundary refinement loss calculation unit calculates the boundary position deviation using a distance transformation map to obtain a boundary loss value. The topological structure preservation loss calculation unit calculates the regional connectivity deviation using a structural similarity function to obtain a topological loss value. Generate a comprehensive loss value according to the segmentation loss value, the boundary loss value, and the topological loss value. Use an iterative feedback optimization mechanism to train the adaptive feature fusion network. In each iteration process, calculate the parameter gradient based on the comprehensive loss value and update the weight parameters in the adaptive feature fusion network until the comprehensive loss value converges to obtain a pollution area segmentation model.

[0026] Calculate the pollution area ratio of each photovoltaic module according to the pollution area segmentation result, and combine the temperature difference between the pollution area and the non-pollution area in the second temperature distribution image set to establish a pollution degree evaluation index, including:

[0027] Perform connected component labeling on the pollution area of each photovoltaic module according to the pollution area segmentation result, calculate the pixel area of each connected component to obtain the pixel area of the pollution area, and obtain the total pixel area of each photovoltaic module. Calculate the initial pollution area ratio based on the ratio of the pixel area of the pollution area to the total pixel area. Calculate a perspective transformation matrix according to the imaging perspective of the image acquisition device, and correct the initial pollution area ratio based on the perspective transformation matrix to obtain a corrected pollution area ratio.

[0028] Based on the temperature data of each photovoltaic module in the second temperature distribution image set, extract the average temperature value of the pollution area and the average temperature value of the non-pollution area, and calculate the temperature difference between the average temperature value of the pollution area and the average temperature value of the non-pollution area. Obtain the maximum temperature difference and the minimum temperature difference of all photovoltaic modules, and normalize the temperature difference based on the maximum temperature difference and the minimum temperature difference to obtain a normalized temperature difference.

[0029] Calculate the variance of the pollution area correction ratio to obtain the area feature variance, and calculate the variance of the normalized temperature difference to obtain the temperature feature variance; construct an adaptive weight allocation function based on the area feature variance and the temperature feature variance, and the adaptive weight allocation function uses an exponential form to map the area feature variance and the temperature feature variance to obtain the area feature weight and the temperature feature weight; construct a pollution degree evaluation index based on the area feature weight, the temperature feature weight, the pollution area correction ratio, and the normalized temperature difference.

[0030] Evaluate the power generation efficiency loss of the photovoltaic power station based on the pollution degree evaluation index and the pollution area ratio, and establish a mapping relationship between the pollution degree and the power generation efficiency loss by comparing the actual power generation data of the polluted area and the non-polluted area in the pollution area segmentation result, including:

[0031] Extract the polluted area and the non-polluted area of each photovoltaic module based on the pollution area segmentation result, and obtain the real-time power generation data and the corresponding solar irradiance data of the polluted area and the non-polluted area; divide the real-time power generation data of the polluted area by the area of the polluted area to obtain the power generation per unit area, and calculate the ratio of the power generation per unit area to the solar irradiance data to obtain the power generation efficiency of the polluted area; divide the real-time power generation data of the non-polluted area by the area of the non-polluted area to obtain the power generation per unit area, and calculate the ratio of the power generation per unit area to the solar irradiance data to obtain the power generation efficiency of the non-polluted area;

[0032] Obtain the real-time temperature data and wind speed data on the surface of the photovoltaic module, multiply the difference between the real-time temperature data and the temperature reference value by the temperature coefficient to obtain the temperature correction coefficient, and multiply the difference between the wind speed data and the wind speed reference value by the wind speed coefficient to obtain the wind speed correction coefficient; add the product of the power generation efficiency of the polluted area and the temperature correction coefficient and the wind speed correction coefficient to obtain the corrected power generation efficiency of the polluted area, and add the product of the power generation efficiency of the non-polluted area and the temperature correction coefficient and the wind speed correction coefficient to obtain the corrected power generation efficiency of the non-polluted area;

[0033] Subtract the corrected power generation efficiency of the non-polluted area from the corrected power generation efficiency of the polluted area to obtain the power generation efficiency difference, and divide the power generation efficiency difference by the corrected power generation efficiency of the non-polluted area to obtain the power generation efficiency loss ratio; calculate the daily variation curve of solar irradiance intensity based on the solar altitude angle, and construct a time weight function according to the daily variation curve. The time weight function has a larger weight during periods with high solar irradiance intensity; multiply the power generation efficiency loss ratio by the weight of the time weight function at the corresponding moment and accumulate, and divide by the sum of the weights of the time weight function to obtain the daily average power generation efficiency loss; construct a power generation efficiency loss evaluation model based on the daily average power generation efficiency loss, and establish a mapping relationship between the pollution degree and the power generation efficiency loss.

[0034] Construct a power generation efficiency loss evaluation model based on the daily average power generation efficiency loss, and establish a mapping relationship between the pollution degree and the power generation efficiency loss, including:

[0035] Take the pollution degree evaluation index as the independent variable and the daily average power generation efficiency loss as the dependent variable, and construct an initial mapping model using a quadratic polynomial function; construct a sum of squared errors objective function based on the initial mapping model, and use the least squares method to iteratively optimize the coefficients of the quadratic polynomial function to obtain an optimized pollution degree mapping model. The prediction error of the optimized pollution degree mapping model is less than a preset threshold;

[0036] Take the pollution area ratio as the independent variable, and construct an area impact function based on the piecewise linear interpolation method. Use a linear function to describe the corresponding relationship between the pollution area ratio and the power generation efficiency loss in each piecewise interval; take the output result of the optimized pollution degree mapping model and the output result of the area impact function as inputs, and construct an adaptive weight fusion layer. The adaptive weight fusion layer includes a pollution degree weight and an area ratio weight; set the length and sliding step of the sliding time window, and collect the measured power generation efficiency loss data in each sliding time window;

[0037] Construct a weight optimization objective function based on the measured power generation efficiency loss data, and use the gradient descent method to perform online optimization and update of the pollution degree weight and the area ratio weight; integrate the pollution degree weight, the area ratio weight, the optimized pollution degree mapping model, and the area impact function to obtain a power generation efficiency loss evaluation model; according to the power generation efficiency loss evaluation model, input the pollution degree evaluation index and the pollution area ratio into the power generation efficiency loss evaluation model to obtain a predicted value of the power generation efficiency loss; establish a quantitative mapping relationship between the pollution degree and the power generation efficiency loss based on the predicted value of the power generation efficiency loss.

[0038] In the second aspect of the embodiments of the present invention,

[0039] Provided is an intelligent cleaning decision-making system for a photovoltaic power station based on machine learning, including:

[0040] A first unit for deploying a multi-dimensional sensor network to collect multi-source monitoring data of the photovoltaic power station, where the multi-source monitoring data includes dust accumulation amount data on the surface of photovoltaic modules, power generation efficiency data of photovoltaic modules, environmental meteorological data, and historical cleaning record data; among them, the dust accumulation amount data is collected collaboratively by a distributed image sensor and a laser scattering sensor, the power generation efficiency data is collected by a current-voltage sensor array, and the environmental meteorological data includes multi-parameter combined data;

[0041] A second unit for respectively inputting the multi-source monitoring data into a physical model and a data model, constructing a hybrid digital twin system of the photovoltaic power station, and using the hybrid digital twin system of the photovoltaic power station to output a dynamic evaluation result of the photovoltaic power station; according to the dynamic evaluation result, constructing a multi-objective constrained deep reinforcement learning framework to train an intelligent cleaning decision-making model, and obtaining a calculation result of the intelligent cleaning decision-making model; based on the calculation result of the intelligent cleaning decision-making model, constructing an adaptive cleaning threshold triggering mechanism, and when the reduction value of the power generation efficiency of the photovoltaic module detected by the current-voltage sensor array is greater than the critical value of the dynamic power generation efficiency loss, generating a refined cleaning operation plan by using a spatio-temporal optimization algorithm;

[0042] A third unit for controlling multiple cleaning devices to perform cleaning operations according to the cleaning time sequence based on the refined cleaning operation plan by using a distributed task scheduling algorithm, and collecting multi-dimensional operation state data through an edge computing node; inputting the multi-dimensional operation state data and the multi-source monitoring data into a transfer learning module, calculating the policy network parameters of the intelligent cleaning decision-making model by using an online incremental learning method, and updating the calculated policy network parameters to a dual value network structure in the deep reinforcement learning framework.

[0043] In the third aspect of the embodiments of the present invention,

[0044] Provided is an electronic device, including:

[0045] A processor;

[0046] A memory for storing instructions executable by the processor;

[0047] Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0048] In the fourth aspect of the embodiments of the present invention,

[0049] Provided is a computer-readable storage medium, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0050] The beneficial effects of the present application are as follows:

[0051] 1. By using an unmanned aerial vehicle (UAV) to carry a visible light camera and an infrared thermal imaging camera to synchronously collect the surface information of photovoltaic modules, the present invention realizes the comprehensiveness and high efficiency of pollution detection. The design of the preset flight path is generated based on the layout information of the photovoltaic power station, ensuring the complete coverage of data collection, improving the detection efficiency, and avoiding the problems of time-consuming, laborious, and low accuracy of traditional manual detection methods.

[0052] 2. The present invention uses a deep learning model to analyze and process the collected images, comprehensively considers the color features, texture features, and temperature features of the surface of photovoltaic modules for pollution area segmentation, and improves the accuracy of pollution detection. By establishing pollution degree evaluation indicators and combining the pollution area ratio and temperature difference data, the quantitative evaluation of the pollution degree is realized, providing a scientific basis for the maintenance and management of photovoltaic power stations.

[0053] 3. By establishing a mapping relationship between the pollution degree and the loss of power generation efficiency, the present invention generates a pollution distribution map and a power generation efficiency loss assessment report of the photovoltaic power station, enabling the power station management personnel to intuitively understand the pollution situation in different regions and the corresponding impact on power generation efficiency, arranging cleaning and maintenance work targeted, improving the resource utilization efficiency, and maximizing the power generation benefit of the photovoltaic power station. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 It is a schematic flow chart of the intelligent cleaning decision-making method for photovoltaic power stations based on machine learning according to an embodiment of the present invention;

[0055] Figure 2 It is a schematic diagram of the comparison of processing times at different image resolutions according to an embodiment of the present invention;

[0056] Figure 3 It is a schematic diagram of the performance comparison of different segmentation methods on various evaluation indicators according to an embodiment of the present invention;

[0057] Figure 4 It is a schematic diagram of the boundary segmentation performance comparison of different pollution types according to an embodiment of the present invention;

[0058] Figure 5 It is a tabular diagram of the comparison of the prediction accuracy of power generation efficiency loss of different prediction models according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0059] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are only a part rather than all of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0060] The technical solutions of the present invention will be described in detail below with specific embodiments. These specific embodiments may be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0061] Figure 1 The following is a schematic flowchart of a method for intelligent cleaning decision-making of a photovoltaic power station based on machine learning according to an embodiment of the present invention, as Figure 1 shown. The method includes:

[0062] Collect image information of a photovoltaic power station by using an unmanned aerial vehicle (UAV). The UAV is equipped with a visible light camera and an infrared thermal imaging camera. The visible light camera is used to collect a first visible light image of the surface of a photovoltaic module, and the infrared thermal imaging camera is used to collect a first temperature distribution image of the surface of the photovoltaic module. The UAV flies according to a preset flight path, and the preset flight path is generated based on the layout information of the photovoltaic power station, so as to obtain a first visible light image set and a first temperature distribution image set of the photovoltaic power station.

[0063] Perform image registration and preprocessing on the first visible light image set and the first temperature distribution image set to obtain a second visible light image set and a second temperature distribution image set. Analyze the second visible light image set and the second temperature distribution image set through a deep learning model, segment the contaminated area based on the color feature, texture feature, and temperature feature of the surface of the photovoltaic module, and obtain a segmentation result of the contaminated area. Calculate the pollution area ratio of each photovoltaic module according to the segmentation result of the contaminated area, and establish a pollution degree evaluation index in combination with the temperature difference between the contaminated area and the non-contaminated area in the second temperature distribution image set.

[0064] Evaluate the power generation efficiency loss of the photovoltaic power station based on the pollution degree evaluation index and the pollution area ratio. By comparing the actual power generation data of the contaminated area and the non-contaminated area in the segmentation result of the contaminated area, establish a mapping relationship between the pollution degree and the power generation efficiency loss. Generate a pollution distribution map and a power generation efficiency loss evaluation report of the photovoltaic power station according to the mapping relationship. The pollution distribution map identifies areas with different pollution degrees, and the power generation efficiency loss evaluation report includes the pollution degree evaluation index, the temperature difference, and the predicted value of the power generation efficiency loss of each area.

[0065] In an alternative embodiment, image registration and preprocessing are performed on the first visible light image set and the first temperature distribution image set to obtain a second visible light image set and a second temperature distribution image set, including:

[0066] Establish an initial mapping relationship between the first visible light image set and the first temperature distribution image set; extract feature points for the first visible light image set and the first temperature distribution image set, perform multi-scale decomposition on the first visible light image set and the first temperature distribution image set using a difference of Gaussian pyramid, detect local extreme points in the multi-scale decomposed images, and generate feature descriptor vectors based on the local extreme points;

[0067] Perform initial matching on the feature descriptor vectors using a nearest neighbor ratio matching strategy, establish feature point correspondence relationships, construct an affine transformation matrix according to the feature point correspondence relationships, and optimize the parameters of the affine transformation matrix using an iterative weighted least squares method; perform geometric correction and radiometric correction on the first visible light image set and the first temperature distribution image set based on the optimized affine transformation matrix, where the geometric correction includes eliminating the distortion of the first visible light image set and the first temperature distribution image set, and the radiometric correction includes eliminating uneven illumination;

[0068] Perform enhancement processing on the geometrically and radiometrically corrected images using an adaptive histogram equalization method, and perform denoising processing in combination with a bilateral filtering algorithm. The filtering kernel function of the bilateral filtering algorithm takes into account both spatial distance weights and gray similarity weights; perform registration accuracy evaluation and image quality evaluation on the enhanced and denoised images, and dynamically adjust the registration parameters according to the results of the registration accuracy evaluation and the image quality evaluation until a preset quality threshold is met, to obtain a second visible light image set and a second temperature distribution image set.

[0069] Establish an initial mapping relationship between the first visible light image set and the first temperature distribution image set. The first visible light image set may include multiple frames of images obtained using an ordinary optical camera, such as RGB images with a resolution of 1920×1080 pixels; the first temperature distribution image set may include multiple frames of images obtained using an infrared thermal imager, such as temperature distribution images with a resolution of 640×480 pixels. The initial mapping relationship is established based on the relative positions and shooting parameters of the two imaging devices for subsequent precise registration.

[0070] Feature points are extracted from the first visible light image set and the first temperature distribution image set. The Gaussian difference pyramid method is used to perform multi-scale decomposition on the two image sets. Specifically, first, Gaussian blur is applied to the original image to obtain images with different degrees of blur, and then adjacent Gaussian blurred images are subtracted to obtain Gaussian difference images. For example, a Gaussian difference pyramid with 5 layers and 4 groups is constructed, and the ratio of adjacent Gaussian kernel parameters for each layer is 1.6. Local extreme points are detected in the images after multi-scale decomposition by comparing the gray value of each pixel with the gray values of its 26 neighborhood points (3×3×3 cube neighborhood). If the gray value of a certain point is greater than or less than the gray values of all its neighborhood points, then this point is regarded as a local extreme point.

[0071] For the detected extreme points, feature descriptor vectors are generated. With each extreme point as the center, a 16×16 pixel area around it is extracted and divided into 4×4 sub-blocks. The gradient histogram in 8 directions is calculated in each sub-block, thereby generating a 128-dimensional feature descriptor vector. To enhance the robustness of the feature descriptor, the vector is normalized, and the gradient magnitude threshold is set to 0.2 to reduce the sensitivity to illumination changes.

[0072] The nearest neighbor ratio matching strategy is used to perform initial matching on the feature descriptor vectors. For each feature point in the first visible light image set, find the two feature points with the smallest Euclidean distance in the first temperature distribution image set. If the ratio of the minimum distance to the second minimum distance is less than 0.8, then the matching is considered successful. Through this method, the corresponding relationship of feature points is established, and accurate matching point pairs are screened out.

[0073] According to the corresponding relationship of feature points, an affine transformation matrix is constructed. The affine transformation matrix contains six parameters, which are used to describe the rotation, scaling, shearing, and translation relationships between images. The iterative weighted least squares method is used to optimize the parameters of the affine transformation matrix. In each iteration, the weight is calculated based on the error of the current matching point pair. The greater the error of the matching point pair, the smaller the weight. For example, the weight of the matching point with an error exceeding 5 pixels is set to 0, the weight of the matching point with an error within 1 pixel is set to 1, and the weights of other matching points are set according to a linear ratio. Through 5 to 10 iterations of optimization, the optimal affine transformation matrix is obtained.

[0074] Based on the optimized affine transformation matrix, geometric correction and radiometric correction are performed on the first visible light image set and the first temperature distribution image set. Geometric correction includes eliminating the distortion of the two image sets, mainly dealing with the deformation caused by perspective distortion and lens distortion. For common radial distortion, a polynomial model is used for correction, and the coefficient values are determined according to the camera calibration results. For example, the radial distortion coefficients are k1 = -0.215 and k2 = 0.088. For tangential distortion, the corresponding correction formula is used for correction, and the coefficients are usually p1 = 0.003 and p2 = -0.001.

[0075] Radiometric correction mainly eliminates the uneven illumination. For visible light images, the background estimation and division correction method is adopted. First, the background brightness distribution of the image is estimated, then the original image is divided by the background brightness, and then multiplied by the average brightness value. For temperature distribution images, the look-up table correction method is used to establish the mapping relationship between the reference temperature and the image gray value according to the standard blackbody radiation source. For example, a correction point is recorded every 5 °C in the range of 20 °C to 100 °C to construct a complete correction curve.

[0076] After correction, the adaptive histogram equalization method is used to enhance the image. The image is divided into 8×8 small blocks, and histogram equalization is performed on each small block separately. Bilinear interpolation is used for smooth transition at the block boundaries, and the contrast change threshold is limited to 3.0 to avoid excessive amplification of noise.

[0077] The bilateral filtering algorithm is combined for denoising. The filtering kernel function of the bilateral filtering algorithm takes into account both the spatial distance weight and the gray similarity weight. The spatial distance weight uses a Gaussian function with a standard deviation set to 3.0 pixels; the gray similarity weight also uses a Gaussian function with a standard deviation set to 15.0 gray levels. The filtering window size is set to 11×11 pixels, which can effectively suppress noise while retaining edge information.

[0078] The registration accuracy evaluation and image quality evaluation are performed on the images after enhancement and denoising. The registration accuracy evaluation uses the root mean square error (RMSE) index to calculate the average error distance between corresponding points, and it is required that the RMSE is less than 1.5 pixels. The image quality evaluation uses the structural similarity index (SSIM), and it is required that the SSIM value of the similar region is greater than 0.75. The registration parameters are dynamically adjusted according to the evaluation results, such as adjusting the feature matching threshold, the affine transformation optimization weight, or the enhancement processing parameters, until the preset quality threshold is met.

[0079] Through the above processing, a high-quality second set of visible light images and a second set of temperature distribution images are finally obtained. These images have accurate registration correspondence, improved geometric and radiometric characteristics, enhanced details, and reduced noise, and can be used for subsequent temperature and visible light information fusion analysis.

[0080] Figure 2 Schematic diagram of the processing time comparison under different image resolutions in the embodiment of the present invention:

[0081] The figure shows the comparison of the processing times of three different technical solutions at different image resolutions. The chart is presented in the form of a line chart, where the horizontal axis represents the image resolution (from 640×480 to 2560×1920), and the vertical axis represents the processing time (seconds). The three solutions are represented by different line types: the solid line with triangular markers represents "this technical solution", the dashed line with circular markers represents "feature point traversal matching", and the dotted line with square markers represents "traditional registration and correction".

[0082] From the data trend, as the resolution increases, the processing times of all three solutions show exponential growth. At low resolution (640×480), the processing times of all three solutions are close to 1 second, with little difference. However, as the resolution increases, the difference becomes more significant: at the high resolution of 2560×1920, the processing time of "this technical solution" is approximately 6 seconds, "feature point traversal matching" requires approximately 12 seconds, and "traditional registration and correction" requires approximately 16 seconds. This clearly shows that "this technical solution" has obvious performance advantages in high-resolution image processing, especially when the resolution is above 1600×1200, its processing efficiency is significantly better than the other two solutions. Overall, the figure effectively demonstrates the differences in image processing efficiency among different technical solutions, especially highlighting the superiority of "this technical solution" in processing high-resolution images.

[0083] In an optional implementation manner, the second visible light image set and the second temperature distribution image set are analyzed through a deep learning model, and the pollution area is segmented based on the color characteristics, texture characteristics, and temperature characteristics on the surface of the photovoltaic module to obtain a pollution area segmentation result, including:

[0084] Construct a multi-scale feature extraction network, which includes a visible light image feature extraction branch and a thermal imaging image feature extraction branch. The visible light image feature extraction branch uses a pyramid structure to perform multi-scale decomposition on the second visible light image set to obtain visible light multi-scale feature maps, and the thermal imaging image feature extraction branch uses a pyramid structure to perform multi-scale decomposition on the second temperature distribution image set to obtain temperature multi-scale feature maps;

[0085] Based on the visible light multi-scale feature maps, the second visible light image set is converted from the RGB space to the CIELAB color space, and the color difference value in the CIELAB color space is calculated. The color difference value is obtained by weighted calculation of the brightness difference value and the chromaticity difference value to obtain a color feature map; a gray-level co-occurrence matrix is constructed based on the visible light multi-scale feature maps and multi-directional multi-scale filtering is performed, and the texture feature matrix of the second visible light image set is extracted to obtain a texture feature map, and the texture feature matrix includes direction features and scale features;

[0086] Based on the temperature multi-scale feature map, calculate the temperature gradient matrix of the second temperature distribution image set to obtain a gradient feature map, where the temperature gradient matrix includes horizontal direction gradient and vertical direction gradient; construct a thermal anomaly region detection model based on the gradient feature map, and the thermal anomaly region detection model determines the anomaly region by calculating the difference degree between the temperature pixel value and the neighborhood temperature distribution to obtain a temperature anomaly feature map; construct a pollution region segmentation model; based on the pollution region segmentation model, process the color feature map, the texture feature map and the temperature anomaly feature map to obtain a pollution region segmentation result.

[0087] In the data acquisition stage, use a high-resolution visible light camera and an infrared thermal imaging camera to collect the visible light image set and the temperature distribution image set on the surface of the photovoltaic module respectively. The resolution of the visible light camera is 4000×3000 pixels, the acquisition distance is 2 - 5 meters, and the shooting angle is perpendicular to the surface of the photovoltaic module. The resolution of the infrared thermal imaging camera is 640×480 pixels, the temperature measurement range is -20°C to 150°C, the temperature sensitivity is 0.05°C, and the acquisition distance and angle are the same as those of the visible light camera.

[0088] Construct a multi-scale feature extraction network, which includes a visible light image feature extraction branch and an infrared thermal imaging image feature extraction branch. In the visible light image feature extraction branch, a pyramid structure is used to perform multi-scale decomposition on the visible light image. Specifically, the original image size is 4000×3000 pixels, and through continuous downsampling operations, images of 2000×1500, 1000×750, 500×375, and 250×188 pixels are generated in sequence to form a 5-layer pyramid structure. Each layer of the image is subjected to feature extraction using a 3×3 convolutional kernel, the number of convolutional layers is 4 layers, and the number of convolutional kernels in each layer is 64, 128, 256, and 512 respectively. In this way, a visible light multi-scale feature map is obtained, which contains image feature representations from coarse-grained to fine-grained.

[0089] In the infrared thermal imaging image feature extraction branch, a pyramid structure is also used to perform multi-scale decomposition on the temperature distribution image. The original infrared thermal imaging image size is 640×480 pixels, and through the same downsampling strategy, an image hierarchy of 320×240, 160×120, 80×60, and 40×30 pixels is formed. Considering the characteristics of the infrared thermal imaging image, a 5×5 convolutional kernel is used for convolutional feature extraction to capture a larger range of temperature distribution patterns, the number of convolutional layers is 3 layers, and the number of convolutional kernels in each layer is 32, 64, and 128 respectively. In this way, a temperature multi-scale feature map containing multi-scale temperature distribution features is obtained.

[0090] Based on the obtained visible light multi-scale feature maps, the system further extracts color features and texture features. During the color feature extraction process, the visible light image in the RGB space is converted to the CIELAB color space (Lab*). The L channel represents luminance, with a value range from 0 to 100; the a channel represents the range from green to red, with a value range from -128 to 127; the b channel represents the range from blue to yellow, also with a value range from -128 to 127. By comparing the differences between the area to be detected and the reference area (usually a clean area on the photovoltaic module) in the Lab space, the color difference value is calculated. The weight of the luminance difference value is set to 0.6, and the weight of the chromaticity difference value is set to 0.4. The color feature map is comprehensively obtained, effectively reflecting the color differences between the polluted area and the normal area.

[0091] During the texture feature extraction process, a gray-level co-occurrence matrix is constructed based on the visible light multi-scale feature maps. The system selects four directions of 0°, 45°, 90°, and 135°, calculates the gray-level co-occurrence matrix in each direction, and sets the distance parameter to 1, 2, and 3 pixels, thereby obtaining 12 gray-level co-occurrence matrices. Four statistical features of contrast, energy, homogeneity, and correlation are extracted from each matrix to form a 48-dimensional texture feature vector. These features are mapped to a texture feature map through a fully connected layer, with the same size as the input image, and the value of each pixel represents the texture complexity at that position.

[0092] Based on the temperature multi-scale feature maps, the system calculates the temperature gradient matrix to obtain the gradient feature map. The specific implementation method is to calculate the temperature change rate of each pixel point in the horizontal and vertical directions respectively. The Sobel operator is used for calculation, and the kernel size is 3×3. The sum of the absolute values of the horizontal gradient and the vertical gradient is used as the total gradient value of the pixel point. The gradient feature map intuitively reflects the areas with drastic temperature changes, which usually correspond to the pollution boundaries.

[0093] Based on the gradient feature map, a thermal anomaly region detection model is constructed. The model uses the sliding window method, and the window size is 7×7 pixels. For each pixel at the center of the window, calculate the difference between its temperature value and the temperature values of other pixels within the window. If the difference exceeds the preset threshold (set to 2°C in the experiment), then mark this pixel as an abnormal point. At the same time, fit the temperature value distribution within the window to a Gaussian distribution, and calculate the number of standard deviations by which the temperature value of the central pixel deviates from the mean. If it exceeds 2.5 standard deviations, also mark it as an abnormal point. In this way, the temperature anomaly feature map is obtained, where the high-value regions correspond to the possible pollution areas.

[0094] Build a pollution area segmentation model using the U-Net architecture, taking the color feature map, texture feature map, and temperature anomaly feature map as the inputs for three channels. The encoder part consists of 4 downsampling blocks, each block containing two 3×3 convolutional layers and one 2×2 max pooling layer. The decoder part consists of 4 upsampling blocks, each block containing one 2×2 transposed convolutional layer and two 3×3 convolutional layers. The feature maps are passed from the encoder to the corresponding layers of the decoder through skip connections. The last layer uses a 1×1 convolution and a Sigmoid activation function to generate the probability map of the pollution area. The probability threshold is set to 0.5, and the pixels with probabilities greater than the threshold are classified as the pollution area to obtain the final pollution area segmentation result.

[0095] In practical applications, this method was verified on a test set containing 500 photovoltaic module images, including various situations such as different degrees of dust pollution, bird droppings pollution, and leaf occlusion. The segmentation accuracy (IoU) reached 89.7%, which was 8.3 percentage points higher than the method using only visible light images and 12.1 percentage points higher than the method using only thermal imaging images.

[0096] Figure 3 Schematic diagram of the performance comparison of different segmentation methods in the embodiments of the present invention on various evaluation metrics:

[0097] This table compares the performance of seven different technical solutions on multiple evaluation metrics. Specifically, the "present technical solution" shows the best performance in all indicators: the IoU reaches 87.6%, the F1 score is 0.934, the Dice coefficient is 0.927, the precision is 93.2%, the recall is 93.6%, the boundary F value is 0.897, and the processing time is 142 ms. In contrast, although the "single visible light feature" and the "single temperature feature" have shorter processing times (118 ms and 95 ms respectively), their performance indicators are generally lower than those of the present technical solution. Traditional methods such as "traditional U-Net" and "DeepLabV3+" are at a medium level in terms of performance, but have longer processing times (167 ms and 198 ms respectively). The performance indicators of "handcrafted features + SVM" and "traditional threshold segmentation" are relatively poor, especially the latter's IoU is only 61.2%, although its processing time is the shortest (48 ms). From the overall data, the present technical solution achieves a good balance between performance and efficiency, showing obvious comprehensive advantages, especially in terms of accuracy-related indicators, far exceeding other solutions. Although it is not the fastest in terms of processing time, the processing speed of 142 ms is still within an acceptable range, especially considering its significantly improved performance indicators.

[0098] For the improvement of the prior art, traditional photovoltaic module pollution detection methods mainly rely on a single image source (usually visible light images) and use threshold segmentation or simple machine learning methods for detection, making it difficult to adapt to complex environmental conditions and different types of pollution. The improvements of this application are as follows: First, it integrates two complementary information sources, visible light images and thermal imaging images, capable of simultaneously capturing the visual and thermal characteristics of pollutants; Second, it designs a multi-scale feature extraction network to adapt to pollutants of different scales; Third, it constructs a feature extraction module for the characteristics of photovoltaic modules, including the accurate extraction of color features, texture features, and temperature anomaly features; Fourth, it uses a deep learning model for feature fusion and region segmentation, improving the adaptive ability and generalization performance of the system. Through these technical improvements, the method provided by this application can accurately identify different types of pollution areas under various lighting conditions, providing reliable technical support for the intelligent cleaning decision-making of photovoltaic power stations.

[0099] In an alternative embodiment, a pollution area segmentation model is constructed, including:

[0100] Construct a dual attention enhancement module. The dual attention enhancement module receives the color feature map, the texture feature map, and the temperature anomaly feature map as inputs, and performs feature enhancement through a regional spatial attention sub-module and a multi-channel feature attention sub-module. The regional spatial attention sub-module includes a deformable convolutional layer and a spatial weight calculation layer. The deformable convolutional layer adaptively adjusts the convolutional sampling positions using learnable offsets, and the spatial weight calculation layer generates a two-dimensional spatial weight matrix based on the local region response. The multi-channel feature attention sub-module includes a global pooling layer and a channel mapping layer. The global pooling layer extracts channel-level feature statistical information, and the channel mapping layer generates a channel weight vector through a non-linear transformation. Multiply the two-dimensional spatial weight matrix element-wise with the input feature map in the spatial dimension, and multiply the channel weight vector channel-wise with the feature map weighted in the spatial dimension to obtain an attention-enhanced feature map;

[0101] Construct an adaptive feature fusion network. The adaptive feature fusion network receives the attention-enhanced feature map as an input, and includes a feature correlation calculation unit and a dynamic weight allocation unit. The feature correlation calculation unit calculates the complementary degree between different features in the attention-enhanced feature map based on cosine similarity to obtain a feature correlation matrix. The dynamic weight allocation unit generates an adaptive weight coefficient according to the feature correlation matrix, and fuses different features through weighted summation to obtain a fused feature map;

[0102] Construct a multi-level loss calculation module. The multi-level loss calculation module receives the fused feature map as input and includes a region segmentation loss calculation unit, a boundary refinement loss calculation unit, and a topological structure preservation loss calculation unit. The region segmentation loss calculation unit calculates the category prediction error using a weighted cross-entropy function to obtain a segmentation loss value. The boundary refinement loss calculation unit calculates the boundary position deviation using a distance transformation map to obtain a boundary loss value. The topological structure preservation loss calculation unit calculates the regional connectivity deviation using a structural similarity function to obtain a topological loss value. Generate a comprehensive loss value based on the segmentation loss value, the boundary loss value, and the topological loss value. Use an iterative feedback optimization mechanism to train the adaptive feature fusion network. In each iteration process, calculate the parameter gradient based on the comprehensive loss value and update the weight parameters in the adaptive feature fusion network until the comprehensive loss value converges to obtain a pollution area segmentation model.

[0103] The dual attention enhancement module receives the color feature map, the texture feature map, and the temperature anomaly feature map as input and performs feature enhancement through the regional spatial attention sub-module and the multi-channel feature attention sub-module.

[0104] The regional spatial attention sub-module first processes the input features through a deformable convolutional layer. The deformable convolutional layer uses a 3×3 convolutional kernel, and the number of output channels is twice the number of input channels. Half of them are used to generate the feature map, and the other half are used to predict the offsets of the sampling positions. For example, when the size of the input feature map is 64×64×128, the deformable convolutional layer generates a 64×64×128 feature map and a 64×64×128 offset map. Each position in the offset map contains 18 values, corresponding to the x and y direction offsets of 9 sampling points. By applying these offsets to the sampling grid of the standard convolution, the deformable convolution can adaptively adjust the receptive field shape according to the input features, so as to better capture the irregularly shaped pollution area.

[0105] The spatial weight calculation layer generates a two-dimensional spatial weight matrix based on the local region response. First, apply a 1×1 convolution to the output feature map of the deformable convolution to reduce the number of channels to 32, and then extract the significant features of the local region through a max-pooling operation. The size of the pooling kernel is 3×3, and the stride is 1. Then, through two consecutive 3×3 convolutional layers (the number of channels is 16 and 1 respectively) and a Sigmoid activation function, generate a two-dimensional spatial weight matrix with a value range of 0 to 1, and the size is 64×64×1. The larger the weight value, the more important the spatial position is.

[0106] The multi-channel feature attention sub-module includes a global pooling layer and a channel mapping layer. The global pooling layer performs average pooling on the feature map in the spatial dimension, converting the 64×64×128 feature map into a 1×1×128 channel descriptor to extract the global statistical information of each channel. The channel mapping layer adopts a two-layer fully connected network structure with 128 / 4 = 32 neurons in the middle layer. A non-linear transformation is introduced through the ReLU activation function, and finally a 128-dimensional channel weight vector is generated through the Sigmoid function, with the value range of each element being from 0 to 1.

[0107] The generation of the attention-enhanced feature map is divided into two steps: First, the two-dimensional spatial weight matrix is multiplied element-wise with the input feature map in the spatial dimension to enhance the feature response at important spatial positions; then, the channel weight vector is multiplied channel-wise with the spatially weighted feature map in the channel dimension to enhance the feature response of important channels. The finally generated attention-enhanced feature map has a size of 64×64×128.

[0108] The adaptive feature fusion network receives the attention-enhanced feature map as input and includes a feature correlation calculation unit and a dynamic weight allocation unit.

[0109] The feature correlation calculation unit calculates the complementary degree between different features based on cosine similarity. First, the attention-enhanced feature map is divided into three groups according to the source: color features, texture features, and temperature anomaly features, and the number of channels for each group of features is 128 / 3≈42. Then, global average pooling is performed on each group of features to obtain a 128 / 3-dimensional feature descriptor. The pairwise cosine similarities between these three feature descriptors are calculated to form a 3×3 feature correlation matrix. For example, the similarity between color features and texture features is 0.65, the similarity between color features and temperature anomaly features is 0.38, and the similarity between texture features and temperature anomaly features is 0.42.

[0110] The dynamic weight allocation unit generates adaptive weight coefficients based on the feature correlation matrix. First, the elements in the correlation matrix are normalized so that the sum of the elements in each row is 1. Then, the supplementary information amount of each group of features is calculated: 1 minus the average similarity of this feature with other features. For example, the supplementary information amount of color features is 1 - 0.515 = 0.485. Then, the supplementary information amounts of the three groups of features are converted into weight coefficients through the Softmax function. For example, the weight of color features is 0.33, the weight of texture features is 0.31, and the weight of temperature anomaly features is 0.36. Finally, different features are fused through weighted summation to obtain a fused feature map with a size of 64×64×128.

[0111] The multi-level loss calculation module receives the fused feature map as input, including a region segmentation loss calculation unit, a boundary refinement loss calculation unit, and a topological structure preservation loss calculation unit.

[0112] The region segmentation loss calculation unit calculates the class prediction error using a weighted cross-entropy function. First, the fused feature map is converted into a class probability map with a size of 64×64×2 (foreground and background) through a 1×1 convolutional layer and a Softmax activation function. Since the pixels in the contaminated area (foreground) usually account for a relatively small proportion, a class weighting strategy is adopted to balance positive and negative samples, with the foreground weight set to 0.75 and the background weight set to 0.25. The weighted cross-entropy of each pixel is calculated, and the average value of all pixels is taken as the segmentation loss value, such as 0.235.

[0113] The boundary refinement loss calculation unit calculates the boundary position deviation using a distance transformation map. First, a distance transformation map is generated from the ground truth label, representing the distance of each pixel to the nearest boundary. The distance of boundary pixels is 0, and the pixel values are larger for pixels farther from the boundary. A corresponding distance transformation map is also generated from the predicted segmentation result. Then, the mean absolute error between the two distance transformation maps is calculated as the boundary loss value, such as 1.87.

[0114] The topological structure preservation loss calculation unit calculates the regional connectivity deviation using a structural similarity function. First, morphological operations (erosion and dilation) are performed on the ground truth label and the prediction result respectively to extract the topological skeletons of the regions. Then, the structural similarity between the two skeleton images is calculated, with a value range from -1 to 1, where 1 indicates exactly the same. The topological loss value is defined as 1 minus the structural similarity, such as 0.15.

[0115] The comprehensive loss value is generated by weighted summation: 0.6×segmentation loss value + 0.25×boundary loss value + 0.15×topological loss value.

[0116] An iterative feedback optimization mechanism is used to train the adaptive feature fusion network. In each iteration process, the parameter gradients are calculated based on the comprehensive loss value and the network weight parameters are updated. The specific steps are as follows:

[0117] Initialize the model parameters, with the learning rate set to 0.001, the batch size to 32, and the total number of iterations to 200 rounds. In each training batch, calculate the prediction result and the comprehensive loss value under the current model parameters. Use the Adam optimizer to calculate the gradients of the loss function with respect to each parameter and update the model parameters. After every 10 iterations, the learning rate is decayed to 0.95 times the original value. When the change in the comprehensive loss value for 5 consecutive iterations is less than 0.001, the model is considered to have converged and the training ends. The finally obtained model parameters are the contaminated area segmentation model.

[0118] Experiments show that after about 150 rounds of iteration, the comprehensive loss value of the model decreases from the initial 2.45 to 0.31, the average intersection over union (IoU) on the test set reaches 87.3%, and the boundary F1 score reaches 84.6%, verifying the effectiveness of the model in the pollution area segmentation task.

[0119] Figure 4 Schematic diagram for comparing the boundary segmentation performance of different pollution types in the embodiments of the present invention:

[0120] This figure clearly shows the comparison of the boundary F values of four different technical solutions (this technical solution, DeepLabV3+, standard U-Net, and single-modal segmentation) under eight different pollution types. The horizontal axis in the figure lists eight pollution types such as dust pollution, bird droppings pollution, water stain pollution, leaf occlusion, sugar stain coverage, hot spot anomaly, factory emissions, and sand dust coverage, and the vertical axis represents the boundary F value (ranging from 0.70 to 0.95). The data shows that this technical solution (represented by blank columns) performs the best among all pollution types, and the boundary F value generally remains between 0.87 and 0.92. Among them, it reaches the highest of about 0.92 in the leaf occlusion type and the lowest but still reaches about 0.86 in factory emissions. DeepLabV3+ (represented by dotted filling) ranks second, and the boundary F value fluctuates between 0.80 and 0.85. The performance of the standard U-Net (represented by horizontal filling) is slightly lower, and the boundary F value is between 0.78 and 0.84. The single-modal segmentation (represented by grid filling) performs the worst, and the boundary F value is mostly between 0.75 and 0.80. These data fully illustrate that this technical solution has significant performance advantages in various pollution detections, and this advantage remains stable in different pollution types, demonstrating good versatility and reliability.

[0121] In an alternative embodiment, the pollution area ratio of each photovoltaic module is calculated according to the pollution area segmentation result, and combined with the temperature difference between the polluted area and the non-polluted area in the second temperature distribution image set, a pollution degree evaluation index is established, including:

[0122] The polluted areas of each photovoltaic module are marked with connected components according to the pollution area segmentation result, the pixel area of each connected component is calculated to obtain the pixel area of the polluted area, and the total pixel area of each photovoltaic module is obtained; the initial pollution area ratio is calculated based on the ratio of the pixel area of the polluted area to the total pixel area; the perspective transformation matrix is calculated according to the imaging perspective of the image acquisition device, and the initial pollution area ratio is corrected based on the perspective transformation matrix to obtain the corrected pollution area ratio;

[0123] Based on the temperature data of each photovoltaic module in the second temperature distribution image set, extract the average temperature value of the contaminated area and the average temperature value of the non-contaminated area, and calculate the temperature difference between the average temperature value of the contaminated area and the average temperature value of the non-contaminated area; obtain the maximum temperature difference and the minimum temperature difference of all photovoltaic modules, and perform normalization processing on the temperature difference based on the maximum temperature difference and the minimum temperature difference to obtain the normalized temperature difference;

[0124] Calculate the variance of the pollution area correction ratio to obtain the area feature variance, and calculate the variance of the normalized temperature difference to obtain the temperature feature variance; construct an adaptive weight allocation function based on the area feature variance and the temperature feature variance, and the adaptive weight allocation function uses an exponential form to map the area feature variance and the temperature feature variance to obtain the area feature weight and the temperature feature weight; construct a pollution degree evaluation index based on the area feature weight, the temperature feature weight, the pollution area correction ratio, and the normalized temperature difference.

[0125] Perform connected component labeling on the pollution area segmentation result. Connected component labeling is to use image processing algorithms to label the pixel regions connected to each other in the image, so that each connected region has a unique identifier. In this embodiment, for the pollution area of each photovoltaic module, the eight-neighborhood connectivity criterion can be used, that is, to judge whether there are similar pixels connected to it in the eight directions around the pixel. For example, for the pollution area image of module A, after connected component labeling, 3 different pollution connected components may be identified, which are labeled as pollution area 1, pollution area 2, and pollution area 3 respectively.

[0126] Calculate the pixel area of each connected component. For each labeled connected component, count the number of pixels it contains. For example, pollution area 1 contains 1200 pixels, pollution area 2 contains 800 pixels, and pollution area 3 contains 500 pixels, then the total pixel area of the pollution area of module A is 2500 pixels. At the same time, obtain the total pixel area of the photovoltaic module. Assume that the total pixel area of module A is 10000 pixels.

[0127] Calculate the initial proportion of the pollution area. The initial proportion of the pollution area is equal to the pixel area of the pollution area divided by the total pixel area of the photovoltaic module. For module A, its initial proportion of the pollution area is 2500 / 10000 = 0.25, that is, 25%.

[0128] Since the imaging perspective of the image acquisition device may cause perspective distortion, it is necessary to correct the initial proportion of the contaminated area. First, according to parameters such as the installation position and inclination angle of the image acquisition device, as well as the layout of the photovoltaic modules in the real world, a perspective transformation matrix is established. This matrix can map the coordinates in the image plane to the coordinates in the real world. Suppose the perspective transformation matrix obtained through calibration generates a correction coefficient of 1.2 for the initial proportion of the contaminated area of Module A, then the corrected proportion of the contaminated area is 0.25×1.2 = 0.3, that is, 30%.

[0129] Analyze the impact of contamination on the module temperature based on the second temperature distribution image set. Extract the temperature data of the contaminated area and the non - contaminated area in each photovoltaic module. For example, the average temperature of the contaminated area of Module A is 48°C, and the average temperature of the non - contaminated area is 42°C, then the temperature difference is 6°C. Similarly, calculate the temperature differences of other photovoltaic modules. Suppose the maximum temperature difference among all modules is 10°C and the minimum temperature difference is 2°C, then normalize the temperature difference of Module A: (6 - 2) / (10 - 2)=0.5, and the normalized temperature difference of Module A is 0.5.

[0130] To comprehensively consider the influence of area characteristics and temperature characteristics on the degree of contamination, it is necessary to calculate the characteristic variance and construct an adaptive weight assignment function. First, calculate the variance of the corrected proportion of the contaminated area of all photovoltaic modules. Suppose the calculated area - characteristic variance is 0.04. Then, calculate the variance of the normalized temperature differences of all photovoltaic modules. Suppose the calculated temperature - characteristic variance is 0.06.

[0131] Based on the area - characteristic variance and the temperature - characteristic variance, construct an adaptive weight assignment function. This function adopts an exponential form and assigns weights according to the magnitude of the characteristic variance. The characteristic with a larger variance has a higher discrimination degree and should be given a larger weight. Suppose the area - characteristic weight is calculated as 0.4 and the temperature - characteristic weight is calculated as 0.6.

[0132] Construct an evaluation index for the degree of contamination. This index comprehensively considers the corrected proportion of the contaminated area and the normalized temperature difference, and is weighted according to the adaptive weights. For Module A, its evaluation index for the degree of contamination is calculated as 0.4×0.3 + 0.6×0.5 = 0.42. The larger the evaluation index, the more serious the contamination degree of the module, and it needs to be cleaned and maintained preferentially.

[0133] A threshold for the evaluation index of the degree of contamination can be set. For example, the evaluation index is divided into three levels: slight contamination (0 - 0.3), moderate contamination (0.3 - 0.6), and severe contamination (0.6 - 1.0). According to the evaluation results, preferentially clean and maintain the severely contaminated modules to improve the overall power generation efficiency of the photovoltaic power station.

[0134] This embodiment can also visually display the evaluation results. For example, the pollution degree evaluation index is mapped to colors to generate a pollution degree heat map, intuitively showing the pollution status of each photovoltaic module. In addition, by combining historical data, the time variation trend of the pollution degree can be analyzed to predict the future pollution development situation, providing decision-making support for the preventive maintenance of the photovoltaic power station.

[0135] Through the above method, the pollution degree of photovoltaic modules can be accurately evaluated, effectively guiding the cleaning and maintenance work of the photovoltaic power station, avoiding waste of resources caused by blind cleaning, and at the same time preventing the decrease in power generation efficiency caused by severe pollution, improving the economic benefits and operation reliability of the photovoltaic power station.

[0136] The pollution degree evaluation method not only considers the pollution area factor, but also combines the temperature distribution data. Through the adaptive weight allocation mechanism, a more comprehensive and accurate pollution degree evaluation is achieved, providing strong support for the refined operation and maintenance of the photovoltaic power station. In addition, this method is applicable to various types of photovoltaic modules and pollution evaluations under different environmental conditions, having broad application value.

[0137] In an alternative embodiment, based on the pollution degree evaluation index and the pollution area ratio, the power generation efficiency loss of the photovoltaic power station is evaluated. By comparing the actual power generation data of the polluted area and the non-polluted area in the pollution area segmentation result, a mapping relationship between the pollution degree and the power generation efficiency loss is established, including:

[0138] Based on the pollution area segmentation result, the polluted area and the non-polluted area of each photovoltaic module are extracted, and the real-time power generation data and the corresponding solar irradiance data of the polluted area and the non-polluted area are obtained; the real-time power generation data of the polluted area is divided by the area of the polluted area to obtain the power generation per unit area, and the ratio of the power generation per unit area to the solar irradiance data is calculated to obtain the power generation efficiency of the polluted area; the real-time power generation data of the non-polluted area is divided by the area of the non-polluted area to obtain the power generation per unit area, and the ratio of the power generation per unit area to the solar irradiance data is calculated to obtain the power generation efficiency of the non-polluted area;

[0139] The real-time temperature data and wind speed data on the surface of the photovoltaic module are obtained. The difference between the real-time temperature data and the temperature reference value is multiplied by the temperature coefficient to obtain the temperature correction coefficient, and the difference between the wind speed data and the wind speed reference value is multiplied by the wind speed coefficient to obtain the wind speed correction coefficient; the power generation efficiency of the polluted area is added to the product of the temperature correction coefficient and the wind speed correction coefficient to obtain the corrected power generation efficiency of the polluted area, and the power generation efficiency of the non-polluted area is added to the product of the temperature correction coefficient and the wind speed correction coefficient to obtain the corrected power generation efficiency of the non-polluted area;

[0140] Subtract the corrected power generation efficiency of the non-polluted area from the corrected power generation efficiency of the polluted area to obtain the power generation efficiency difference, and divide the power generation efficiency difference by the corrected power generation efficiency of the non-polluted area to obtain the power generation efficiency loss ratio; calculate the daily variation curve of solar irradiance intensity based on the solar altitude angle, and construct a time weight function according to the daily variation curve. The time weight function has a larger weight in the period with high solar irradiance intensity; multiply the power generation efficiency loss ratio by the weight of the time weight function at the corresponding moment and accumulate, and divide by the sum of the weights of the time weight function to obtain the daily average power generation efficiency loss; construct a power generation efficiency loss evaluation model based on the daily average power generation efficiency loss, and establish a mapping relationship between the pollution degree and the power generation efficiency loss.

[0141] Extract the polluted area and non-polluted area of each photovoltaic module based on the pollution area segmentation result. Taking a standard photovoltaic module of a certain photovoltaic power station as an example, the total area of this module is 1.96 square meters, the polluted area obtained by the image processing method is 0.58 square meters, and the non-polluted area is 1.38 square meters.

[0142] Obtain the real-time power generation data and the corresponding solar irradiance data of the above-mentioned polluted area and non-polluted area. Taking the data at 10 am on a certain day as an example, the real-time power generation of the polluted area is 72 watts, and the solar irradiance at that time is 850 watts per square meter; the real-time power generation of the non-polluted area is 226 watts, and the solar irradiance is also 850 watts per square meter.

[0143] Calculate the power generation efficiency of the polluted area and the non-polluted area. Divide the real-time power generation data of 72 watts in the polluted area by the area of the polluted area of 0.58 square meters to obtain the power generation per unit area of 124.14 watts per square meter. Calculate the ratio of the power generation per unit area to the solar irradiance data of 850 watts per square meter to obtain the power generation efficiency of the polluted area of 0.146. Similarly, the power generation per unit area of the non-polluted area is 226 watts ÷ 1.38 square meters = 163.77 watts per square meter, and the power generation efficiency of the non-polluted area is 163.77 watts per square meter ÷ 850 watts per square meter = 0.193.

[0144] In order to improve the evaluation accuracy, it is necessary to obtain the real-time temperature data and wind speed data on the surface of the photovoltaic module and perform correction. Taking the same moment as above, the temperature on the surface of the module is measured to be 42 °C, the ambient wind speed is 2.5 m / s, the temperature reference value is 25 °C, and the wind speed reference value is 1.0 m / s. Set the temperature coefficient to -0.0038 / °C and the wind speed coefficient to 0.005 / (m / s).

[0145] Multiply the difference between the real-time temperature data and the temperature reference value by the temperature coefficient to obtain the temperature correction coefficient, i.e., (42 - 25)×(-0.0038) = -0.0646. Multiply the difference between the wind speed data and the wind speed reference value by the wind speed coefficient to obtain the wind speed correction coefficient, i.e., (2.5 - 1.0)×0.005 = 0.0075.

[0146] Add the power generation efficiency in the polluted area to the product of the temperature correction coefficient and the wind speed correction coefficient to obtain the corrected power generation efficiency in the polluted area, i.e., 0.146 + (-0.0646 + 0.0075) = 0.0889. Add the power generation efficiency in the non-polluted area to the product of the temperature correction coefficient and the wind speed correction coefficient to obtain the corrected power generation efficiency in the non-polluted area, i.e., 0.193 + (-0.0646 + 0.0075) = 0.1359.

[0147] Subtract the corrected power generation efficiency in the polluted area from the corrected power generation efficiency in the non-polluted area to obtain the difference in power generation efficiency, i.e., 0.1359 - 0.0889 = 0.047. Divide this difference in power generation efficiency by the corrected power generation efficiency in the non-polluted area to obtain the power generation efficiency loss ratio, i.e., 0.047÷0.1359 = 0.3458, indicating that the pollution causes a power generation efficiency loss of approximately 34.58%.

[0148] Since the solar irradiance intensity varies at different times of the day, it is necessary to calculate the daily variation curve of the solar irradiance intensity based on the solar altitude angle and construct a time weight function according to this curve. This weight function has a larger weight in the period with high solar irradiance intensity. For example, on a typical day, the weight from 6:00 to 8:00 is 0.05, the weight from 8:00 to 10:00 is 0.15, the weight from 10:00 to 14:00 is 0.4, the weight from 14:00 to 16:00 is 0.3, and the weight from 16:00 to 18:00 is 0.1.

[0149] Collect data at multiple time points within a day, calculate the power generation efficiency loss ratio for each period, then multiply the power generation efficiency loss ratio for each period by the corresponding time weight and accumulate them, and finally divide by the sum of the weights of the time weight function to obtain the daily average power generation efficiency loss.

[0150] For example, in the above typical day, the power generation efficiency loss ratio during the period from 6:00 to 8:00 is 0.21, from 8:00 to 10:00 is 0.28, from 10:00 to 14:00 is 0.35 (i.e., 0.3458 calculated previously), from 14:00 to 16:00 is 0.32, and from 16:00 to 18:00 is 0.23. The calculated daily average power generation efficiency loss is: (0.21×0.05 + 0.28×0.15 + 0.35×0.4 + 0.32×0.3 + 0.23×0.1)÷(0.05 + 0.15 + 0.4 + 0.3 + 0.1) = 0.313, that is, the daily average power generation efficiency loss of this photovoltaic module caused by pollution is 31.3%.

[0151] After obtaining the data of photovoltaic modules with multiple pollution levels over multiple days, a mapping relationship between the pollution level and the power generation efficiency loss can be established. For example, it is determined through a large amount of experimental data that when the pollution level assessment index (a value between 0 and 1, where 1 represents the most severe) reaches 0.3 and the pollution area ratio is 20%, the power generation efficiency loss is approximately 10%; when the pollution level assessment index reaches 0.5 and the pollution area ratio is 30%, the power generation efficiency loss is approximately 18%; when the pollution level assessment index reaches 0.7 and the pollution area ratio is 40%, the power generation efficiency loss is approximately 27%; when the pollution level assessment index reaches 0.9 and the pollution area ratio is 60%, the power generation efficiency loss exceeds 35%.

[0152] By constructing a power generation efficiency loss assessment model, an accurate mapping between the pollution level and the power generation efficiency loss can be achieved. For example, the following rules can be adopted: The power generation efficiency loss caused by mild pollution (pollution level index < 0.4 and pollution area ratio < 30%) generally does not exceed 15%; the power generation efficiency loss caused by moderate pollution (pollution level index between 0.4 and 0.7 or pollution area ratio between 30% and 50%) is between 15% and 30%; the power generation efficiency loss caused by severe pollution (pollution level index > 0.7 or pollution area ratio > 50%) usually exceeds 30%.

[0153] This mapping relationship can be used to guide the operation and maintenance management of photovoltaic power stations. When it is predicted that the power generation efficiency loss caused by pollution exceeds a certain threshold (such as 20%), cleaning and maintenance can be arranged, so as to balance the relationship between the cleaning cost and the improvement of power generation efficiency and achieve the economic and efficient operation of photovoltaic power stations.

[0154] Figure 5 The following is a comparison table graph of the power generation efficiency loss prediction accuracies of different prediction models in the embodiments of the present invention:

[0155] This table details the performance test results of this technical solution under different environmental conditions. The table includes 12 different test environments, covering standard test conditions (25°C, no wind) to various extreme weather conditions. From the data, under standard test conditions, both the temperature and wind speed correction factors are 1.000, and the corrected efficiency is 18.70%. Under high temperature conditions (42 - 45°C), the temperature correction factor ranges from 0.897 to 0.911, and the wind speed correction factor ranges from 1.012 to 1.085, with the accuracy improvement of 3.68 - 4.06%. The system performs more prominently under low temperature conditions (5 - 8°C), where the temperature correction factor reaches 1.077 - 1.088, and the wind speed correction factor is 1.020 - 1.102, with a significant accuracy improvement of 6.53 - 7.74%. Under pollution conditions, the corrected efficiency ranges from 15.76 - 18.26% under mild pollution (temperature 10 - 40°C), and the efficiency drops to 12.74% under severe pollution. Especially under sandstorm weather conditions (32°C, wind speed 5.0 m / s), although the temperature correction factor is relatively low (0.962), the wind speed correction factor reaches the highest value of 1.122, achieving an accuracy improvement of 8.21%, which is the highest improvement among all test conditions. Overall, the average corrected efficiency of this technical solution reaches 17.74%, compared with 16.71% of the fixed coefficient correction method, with an average accuracy improvement of 5.07%, demonstrating good adaptability and performance advantages.

[0156] The adaptive correction technical solution proposed in this application innovatively improves the limitations of the existing technology that uses fixed correction coefficients. Traditional methods usually use a single fixed correction coefficient to handle data deviations under different environmental conditions. This method cannot adapt to the complex and changeable actual environment, resulting in unsatisfactory correction effects, especially under extreme weather conditions.

[0157] The technical solution proposed in this application innovatively introduces a dual adaptive correction mechanism for temperature and wind speed. This mechanism can dynamically adjust the correction coefficient according to the real-time changes of environmental temperature and wind speed, realizing the adaptive optimization of correction parameters. Specifically, this solution maintains the benchmark correction effect under standard environmental conditions, while automatically adjusting the correction strategy under extreme environmental conditions (such as high temperature, low temperature, strong wind, etc.). Especially under low temperature conditions, through the collaborative correction of temperature and wind speed, the correction effect is significantly improved; under polluted weather conditions, the system can adaptively adjust the correction strategy according to the degree of pollution, ensuring the stability of the correction effect.

[0158] Compared with the prior art, the improvements of this solution are mainly reflected in the following aspects: First, it breaks through the limitations of traditional fixed - coefficient calibration and realizes the dynamic adaptability of calibration parameters; Second, through the dual - calibration mechanism of temperature and wind speed, it improves the system's adaptability to complex environments; Third, it shows obvious performance advantages under extreme weather conditions, especially the calibration effect is significantly improved under low - temperature and polluted weather conditions.

[0159] Through the above - mentioned technical improvements, this solution not only improves the overall calibration efficiency but also significantly enhances the accuracy. Especially under extreme environmental conditions, it shows stronger environmental adaptability and stability, providing more reliable technical support for practical applications.

[0160] In an alternative embodiment, an evaluation model of power generation efficiency loss is constructed based on the daily average power generation efficiency loss, and a mapping relationship between the pollution degree and the power generation efficiency loss is established, including:

[0161] Taking the pollution degree evaluation index as the independent variable and the daily average power generation efficiency loss as the dependent variable, an initial mapping model is constructed using a quadratic polynomial function; Based on the initial mapping model, an objective function of the sum of squared errors is constructed, and the coefficients of the quadratic polynomial function are iteratively optimized using the least - squares method to obtain an optimized pollution degree mapping model, and the prediction error of the optimized pollution degree mapping model is less than a preset threshold;

[0162] Taking the pollution area ratio as the independent variable, an area impact function is constructed based on the piece - wise linear interpolation method, and the corresponding relationship between the pollution area ratio and the power generation efficiency loss is described using a linear function in each piece - wise interval; Taking the output results of the optimized pollution degree mapping model and the output results of the area impact function as inputs, an adaptive weight fusion layer is constructed, and the adaptive weight fusion layer includes a pollution degree weight and an area ratio weight; The length and sliding step of the sliding time window are set, and the measured power generation efficiency loss data is collected within each sliding time window;

[0163] Based on the measured power generation efficiency loss data, a weight optimization objective function is constructed, and the pollution degree weight and the area ratio weight are optimized and updated online using the gradient - descent method; The pollution degree weight, the area ratio weight, the optimized pollution degree mapping model, and the area impact function are integrated to obtain an evaluation model of power generation efficiency loss; According to the evaluation model of power generation efficiency loss, the pollution degree evaluation index and the pollution area ratio are input into the evaluation model of power generation efficiency loss to obtain a predicted value of power generation efficiency loss; Based on the predicted value of power generation efficiency loss, a quantitative mapping relationship between the pollution degree and the power generation efficiency loss is established.

[0164] Collect the pollution data of photovoltaic modules and the corresponding power generation efficiency loss data. In a specific embodiment, 20 photovoltaic modules of a certain photovoltaic power station are selected as the research objects, and the pollution degree evaluation index values are obtained through image processing technology, ranging from 0.15 to 0.78; at the same time, record the daily average power generation efficiency loss of these modules, ranging from 3.2% to 16.5%. In addition, the pollution area ratio is calculated through image analysis, ranging from 0.08 to 0.63.

[0165] Taking the pollution degree evaluation index as the independent variable and the daily average power generation efficiency loss as the dependent variable, a quadratic polynomial function is used to construct an initial mapping model. Specifically, the form of this quadratic polynomial function is: the daily average power generation efficiency loss is equal to A times the square of the pollution degree evaluation index, plus B times the pollution degree evaluation index, plus C. Where A, B, and C are coefficients to be determined. Initially, A is set to 2.5, B is set to 3.8, and C is set to 1.2.

[0166] Based on this initial mapping model, an objective function of the sum of squared errors is constructed. For each collected sample point, calculate the difference between the model prediction value and the actual daily average power generation efficiency loss, square the differences of all sample points and sum them up to obtain the total sum of squared errors. The least squares method is used to iteratively optimize the coefficients A, B, and C. Specifically, the gradient descent algorithm is used, the iteration step size is set to 0.01, and the maximum number of iterations is 1000 times.

[0167] During the iteration process, when the change in the sum of squared errors between two adjacent iterations is less than 0.0001 or the maximum number of iterations is reached, the iteration stops. Finally, the optimized coefficient A is 2.83, B is 5.12, and C is 0.95. At this time, the average prediction error of this optimized pollution degree mapping model on the test set is 2.3%, which is less than the preset 3% threshold.

[0168] Taking the pollution area ratio as the independent variable, an area influence function is constructed. Using the piecewise linear interpolation method, the pollution area ratio interval [0, 1] is divided into three sub-intervals: [0, 0.3], [0.3, 0.7], and [0.7, 1]. Within each sub-interval, a linear function is used to describe the relationship between the pollution area ratio and the efficiency loss.

[0169] In the first interval [0, 0.3], the slope of the linear function is 0.18 and the intercept is 0.02; in the second interval [0.3, 0.7], the slope is 0.25 and the intercept is -0.01; in the third interval [0.7, 1], the slope is 0.32 and the intercept is -0.06. For example, when the pollution area ratio is 0.5, the output of the area influence function is 0.115.

[0170] Taking the output results of the optimized pollution level mapping model and the output results of the area impact function as inputs, an adaptive weight fusion layer is constructed. This fusion layer includes a pollution level weight and an area ratio weight, both initially set to 0.5.

[0171] To achieve dynamic adjustment of the weights, a sliding time window is set, with a window length of 7 days and a sliding step of 1 day. Within each sliding time window, measured power generation efficiency loss data is collected. For example, the average power generation efficiency loss of 7 groups of measured data collected within a certain time window is 8.7%.

[0172] Based on the measured power generation efficiency loss data, a weight optimization objective function is constructed. This objective function is: the sum of the squares of the difference between the power generation efficiency loss predicted by the model and the measured power generation efficiency loss. The gradient descent method is used to perform online optimization and update of the pollution level weight and the area ratio weight, with a learning rate set to 0.05 and the number of iterations set to 100 times.

[0173] In a specific embodiment, after optimization, the pollution level weight is adjusted from the initial 0.5 to 0.64, and the area ratio weight is adjusted from 0.5 to 0.36. This indicates that in this scenario, the impact of the pollution level on the power generation efficiency loss is slightly greater than the impact of the pollution area ratio.

[0174] Integrating the above optimized pollution level weight, area ratio weight, optimized pollution level mapping model, and area impact function, the final power generation efficiency loss assessment model is obtained. The integration method of this model is: multiplying the output value of the pollution level mapping model by the pollution level weight, adding the output value of the area impact function multiplied by the area ratio weight, to obtain the final predicted value of the power generation efficiency loss.

[0175] Through this assessment model, a quantitative mapping between the pollution level and the power generation efficiency loss can be achieved. For example, when the pollution level assessment index of a certain photovoltaic module is 0.45 and the pollution area ratio is 0.38, the predicted value of the power generation efficiency loss calculated by substituting into the power generation efficiency loss assessment model is 9.7%.

[0176] To verify the effectiveness of the model, 10 photovoltaic modules with different pollution states are selected for testing. The test results show that the average prediction error of this model is 2.1%, and the maximum prediction error is 3.8%, meeting the actual application requirements.

[0177] The model can also be adaptively updated according to time series data. In practical applications, new sample data is collected every 15 days to re-optimize the model parameters, enabling the model to adapt to the changes in pollution characteristics under different seasons and environmental conditions. For example, in the rainy season and the dry season, the optimal values of the pollution degree weight and the area ratio weight are (0.58, 0.42) and (0.67, 0.33) respectively, reflecting the differences in pollution characteristics under different climate conditions.

[0178] By establishing such a quantitative mapping relationship, the impact of the pollution status of photovoltaic modules on power generation efficiency can be accurately evaluated, providing a scientific basis for the cleaning and maintenance decision-making of photovoltaic power plants and realizing precise and intelligent management.

[0179] In the second aspect of the embodiments of the present invention,

[0180] a machine learning-based intelligent cleaning decision-making system for a photovoltaic power plant is provided, including:

[0181] a first unit for deploying a multi-dimensional sensor network to collect multi-source monitoring data of the photovoltaic power plant, where the multi-source monitoring data includes dust accumulation data on the surface of photovoltaic modules, power generation efficiency data of photovoltaic modules, environmental meteorological data, and historical cleaning record data; among them, the dust accumulation data is collected collaboratively by a distributed image sensor and a laser scattering sensor, the power generation efficiency data is collected by a current-voltage sensor array, and the environmental meteorological data includes multi-parameter combined data;

[0182] a second unit for respectively inputting the multi-source monitoring data into a physical model and a data model to construct a hybrid digital twin system of the photovoltaic power plant, and using the hybrid digital twin system of the photovoltaic power plant to output a dynamic evaluation result of the photovoltaic power plant; according to the dynamic evaluation result, constructing a multi-objective constrained deep reinforcement learning framework to train an intelligent cleaning decision-making model and obtaining the calculation result of the intelligent cleaning decision-making model; based on the calculation result of the intelligent cleaning decision-making model, constructing an adaptive cleaning threshold triggering mechanism, and when the reduction value of the power generation efficiency of the photovoltaic module detected by the current-voltage sensor array is greater than the critical value of the dynamic power generation efficiency loss, generating a refined cleaning operation plan by using a spatio-temporal optimization algorithm;

[0183] a third unit for controlling multiple cleaning devices to perform cleaning operations according to the cleaning time series based on the refined cleaning operation plan, and collecting multi-dimensional operation status data through an edge computing node; inputting the multi-dimensional operation status data and the multi-source monitoring data into a transfer learning module, calculating the policy network parameters of the intelligent cleaning decision-making model by an online incremental learning method, and updating the calculated policy network parameters to the dual value network structure in the deep reinforcement learning framework.

[0184] In a third aspect of the embodiments of the present invention,

[0185] a kind of electronic device is provided, including:

[0186] a processor;

[0187] a memory for storing instructions executable by the processor;

[0188] wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0189] In a fourth aspect of the embodiments of the present invention,

[0190] a computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0191] The present invention may be a method, an apparatus, a system, and / or a computer program product. The computer program product may include a computer-readable storage medium, on which computer-readable program instructions for executing various aspects of the present invention are uploaded.

[0192] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An intelligent cleaning decision-making method for a photovoltaic power station based on machine learning, characterized in that, Including: Deploying a multi-dimensional sensor network to collect multi-source monitoring data of a photovoltaic power station, where the multi-source monitoring data includes dust accumulation amount data on the surface of photovoltaic modules, power generation efficiency data of photovoltaic modules, environmental meteorological data, and historical cleaning record data; among which the dust accumulation amount data is collected collaboratively by a distributed image sensor and a laser scattering sensor, the power generation efficiency data is collected by a current-voltage sensor array, and the environmental meteorological data includes multi-parameter combined data; Inputting the multi-source monitoring data into a physical model and a data model respectively to construct a hybrid digital twin system of the photovoltaic power station, and using the hybrid digital twin system of the photovoltaic power station to output a dynamic evaluation result of the photovoltaic power station; according to the dynamic evaluation result, constructing a multi-objective constrained deep reinforcement learning framework to train an intelligent cleaning decision-making model, and obtaining the calculation result of the intelligent cleaning decision-making model; Based on the calculation result of the intelligent cleaning decision-making model, constructing an adaptive cleaning threshold triggering mechanism. When the reduction value of the power generation efficiency of the photovoltaic module detected by the current-voltage sensor array is greater than the critical value of the dynamic power generation efficiency loss, using a spatio-temporal optimization algorithm to generate a refined cleaning operation plan; Based on the refined cleaning operation plan, using a distributed task scheduling algorithm to control multiple cleaning devices to perform cleaning operations according to the cleaning time sequence, and collecting multi-dimensional operation status data through an edge computing node; inputting the multi-dimensional operation status data and the multi-source monitoring data into a transfer learning module, calculating the policy network parameters of the intelligent cleaning decision-making model through an online incremental learning method, and updating the calculated policy network parameters to the dual value network structure in the deep reinforcement learning framework.

2. The method according to claim 1, characterized in that Inputting the multi-source monitoring data into a physical model and a data model respectively to construct a hybrid digital twin system of the photovoltaic power station, and using the hybrid digital twin system of the photovoltaic power station to output a dynamic evaluation result of the photovoltaic power station, including: Inputting the dust accumulation amount data on the surface of the photovoltaic module and the power generation efficiency data of the photovoltaic module into the physical model to construct a bidirectional mapping relationship. The physical model establishes a power generation characteristic model based on the single-diode equivalent circuit principle with temperature compensation. The power generation characteristic model uses an adaptive parameter estimation algorithm to dynamically compensate the series resistance and parallel resistance, and inputs the compensated power generation efficiency data into a characteristic curve fitting module to obtain the output power curve of the photovoltaic module; Establishing a dust accumulation characteristic model based on the multi-scale dust accumulation attenuation characteristic. The dust accumulation characteristic model uses the optical scattering theory to calculate the dust accumulation transmittance, and fuses the dust accumulation thickness parameter, dust accumulation distribution density parameter, and dust accumulation type parameter in the dust accumulation amount data through an exponential weighting function to output a physical characteristic parameter matrix representing the performance of the photovoltaic module; inputting the environmental meteorological data and the historical cleaning record data into the data model for deep feature extraction. The data model adopts a multi-level hybrid neural network structure, including a convolutional neural network layer for local feature extraction, a bidirectional long short-term memory network layer for capturing time series features, and a hierarchical multi-head attention layer for multi-source feature fusion. The convolutional neural network layer performs spatial feature decomposition on temperature, humidity, wind speed, wind direction, air pressure, precipitation, and light intensity in the environmental meteorological data. The bidirectional long short-term memory network layer performs forward and backward temporal encoding on the decomposed features based on a gating mechanism. The hierarchical multi-head attention layer adaptively aligns and fuses the encoded environmental features and the historical cleaning record data through a cross-modal attention mechanism, and outputs a spatio-temporal feature tensor representing the environmental impact. The physical property parameter matrix and the spatio-temporal feature tensor are input into the digital twin system of the hybrid photovoltaic power station, and a dynamic evaluation result of the photovoltaic power station is output.

3. The method according to claim 2, wherein Inputting the physical property parameter matrix and the spatio-temporal feature tensor into the digital twin system of the hybrid photovoltaic power station, and outputting the dynamic evaluation result of the photovoltaic power station, including: Inputting the physical property parameter matrix and the spatio-temporal feature tensor into the digital twin system of the hybrid photovoltaic power station for multi-modal information fusion. The digital twin system of the hybrid photovoltaic power station includes a Bayesian inference module and a dynamic fusion module. The physical property parameter matrix includes a power prediction curve generated by a power generation characteristic model and a multi-scale attenuation coefficient matrix output by an ash deposition characteristic model. The spatio-temporal feature tensor includes the temporal encoding features of environmental meteorological data and the alignment and fusion features of historical cleaning records. Constructing a hierarchical Bayesian network in the Bayesian inference module. The hierarchical Bayesian network uses a variational inference algorithm to map the power prediction curve to a conditional probability distribution of power generation efficiency, map the multi-scale attenuation coefficient matrix to a conditional probability distribution of ash deposition impact, and map the temporal encoding features and the alignment and fusion features to a conditional probability distribution of environmental factors. An adaptive weight matrix is set in the dynamic fusion module, and the conditional probability distribution of power generation efficiency, the conditional probability distribution of ash deposition impact, and the conditional probability distribution of environmental factors are jointly optimized by the alternating direction method of multipliers, and the weight coefficients in the adaptive weight matrix are iteratively updated. Constructing a multi-layer evaluation index system to evaluate the states of the conditional probability distribution of power generation efficiency, the conditional probability distribution of ash deposition impact, and the conditional probability distribution of environmental factors. The multi-layer evaluation index system includes performance indexes, impact indexes, and environmental indexes. The evaluation deviation is calculated by comparing each layer of indexes with a preset hierarchical threshold vector. Based on the evaluation deviation, a robust feedback mechanism is activated to update the conditional probability distribution parameters in the hierarchical Bayesian network and the weight coefficients in the adaptive weight matrix. Based on the updated conditional probability distribution parameters and the weight coefficients, the conditional probability distribution of power generation efficiency, the conditional probability distribution of ash deposition impact, and the conditional probability distribution of environmental factors are substituted into the multi-layer evaluation index system, and a dynamic evaluation result of the photovoltaic power station is output.

4. The method according to claim 1, wherein According to the dynamic evaluation result, constructing a multi-objective constrained deep reinforcement learning framework to train an intelligent cleaning decision-making model, and obtaining the calculation result of the intelligent cleaning decision-making model, including: Input the dynamic evaluation results into the multi-objective constraint module. The multi-objective constraint module constructs a power generation constraint function, a cleaning cost constraint function, and a device life constraint function. The power generation constraint function constructs a power generation efficiency matrix based on the power prediction curve of the power generation characteristic model. The cleaning cost constraint function constructs a cost feature vector based on the historical cleaning record data. The device life constraint function constructs a life loss tensor based on the ash fouling attenuation characteristic model. The multi-objective constraint module uses an adaptive weight method to perform multi-objective fusion on the power generation efficiency matrix, the cost feature vector, and the life loss tensor to generate a constraint objective function. Input the photovoltaic module status information into the state vector construction module. The state vector construction module constructs the cumulative ash fouling amount and the ash fouling distribution density of the photovoltaic module into an ash fouling state matrix, constructs the component temperature distribution into a temperature state matrix, constructs the local meteorological parameters into an environmental state vector, and constructs the available state of the cleaning resources into a resource state vector. The state vector construction module uses a multi-layer perceptron network to extract features from the ash fouling state matrix, the temperature state matrix, the environmental state vector, and the resource state vector to generate a state representation matrix. Input the cleaning operation parameters into the action vector construction module. The action vector construction module encodes the cleaning time schedule into a timing feature vector, encodes the cleaning path planning into a spatial feature matrix, encodes the cleaning intensity control into a control parameter vector, and encodes the cleaning equipment combination configuration into a configuration feature vector. The action vector construction module uses a recurrent neural network to perform timing modeling on the timing feature vector, uses a graph convolutional network to perform spatial modeling on the spatial feature matrix, uses a fully connected network to perform feature mapping on the control parameter vector and the configuration feature vector, and fuses the modeling results to generate an action probability distribution. Train an intelligent cleaning decision-making model based on the constraint objective function, the state representation matrix, and the action probability, and obtain the calculation result of the intelligent cleaning decision-making model.

5. The method according to claim 4, wherein Training the intelligent cleaning decision-making model based on the constraint objective function, the state representation matrix, and the action probability, and obtaining the calculation result of the intelligent cleaning decision-making model includes: Construct a dual-value network structure, and output a reward signal value and a target value based on the constraint objective function, the state representation matrix, and the action probability. Input the reward signal value and the target value into the loss function module. The loss function module constructs a policy loss function and a value loss function. The policy loss function inputs the optimized action probability distribution into a regularization network to generate constraint parameters. The value loss function inputs the reward signal value and the target value into an error calculation unit to generate a network difference value. Input the constraint parameters and the network difference value into a training optimization unit. The training optimization unit performs iterative training on the intelligent cleaning decision-making model, and inputs the trained intelligent cleaning decision-making model into a Monte Carlo search module. The Monte Carlo search module constructs the state representation matrix as the node data of the search tree, constructs the optimized action probability distribution as the edge data of the search tree, inputs the node data into the probability calculation unit to generate node probability values, inputs the edge data into the count statistics unit to generate the number of visits, inputs the node probability values and the number of visits into the confidence calculation unit to generate search parameters, inputs the search parameters into the traversal control unit, and the traversal control unit traverses the search tree along the depth direction, inputs the traversal path into the constraint objective function to generate a path reward value, and inputs the path reward value into the trained intelligent cleaning decision model to generate search direction data; Input the path reward value into the node update unit to update the node data of the search tree, input the updated node data into the path selection unit, the path selection unit selects a decision path according to the number of visits, input the decision path into the constraint verification unit, the constraint verification unit inputs the verification result into the parameter update unit, and the parameter update unit updates the network parameters of the intelligent cleaning decision model and the optimized action probability distribution, and uses the decision path as the output result of the intelligent cleaning decision model.

6. The method according to claim 1, characterized in that, Based on the calculation result of the intelligent cleaning decision model, construct an adaptive cleaning threshold trigger mechanism. When the reduction value of the power generation efficiency of the photovoltaic module detected by the current-voltage sensor array is greater than the dynamic power generation efficiency loss critical value, use the spatio-temporal optimization algorithm to generate a refined cleaning operation plan, including: Input the calculation result of the intelligent cleaning decision model into the feature extraction network, extract the cleaning decision feature parameters from the calculation result, and adjust the prediction parameters of the power generation efficiency curve prediction model and the prediction parameters of the meteorological time series prediction model based on the cleaning decision feature parameters; input the power generation characteristic model into the adjusted power generation efficiency curve prediction model to generate power generation efficiency prediction data, and input the environmental meteorological data predicted by the deep neural network into the adjusted meteorological time series prediction model to generate meteorological prediction data; Adopt an ensemble learning method to fuse the power generation efficiency prediction data and the meteorological prediction data to generate a dynamic power generation efficiency loss critical value; compare the reduction value of the power generation efficiency of the photovoltaic module detected by the current-voltage sensor array with the dynamic power generation efficiency loss critical value. When the reduction value of the power generation efficiency of the photovoltaic module is greater than the dynamic power generation efficiency loss critical value, input the meteorological prediction data of the meteorological time series prediction model into the time series planning unit, input the dust accumulation data into the space planning unit, and input the historical cleaning record data into the equipment configuration unit; Construct a time window constraint matrix for the meteorological prediction data, perform multi-time scale partitioning on the time window constraint matrix and input it into a time series optimization network to generate an optimal cleaning time arrangement. Construct a spatial clustering feature vector for the dust accumulation data and input it into a path planning network to generate an optimal cleaning path plan. Construct an equipment performance evaluation index for the historical cleaning record data and input it into a resource scheduling network to generate an optimal equipment scheduling plan. Combine the optimal cleaning time arrangement, the optimal cleaning path plan and the optimal equipment scheduling plan to generate a refined cleaning operation plan.

7. The method according to claim 1, characterized in that Based on the refined cleaning operation plan, use a distributed task scheduling algorithm to control multiple cleaning devices to perform cleaning operations according to the cleaning time sequence, and collect multi-dimensional operation status data through the edge computing node, including: Input the optimal cleaning time arrangement, the optimal cleaning path plan and the optimal equipment scheduling plan in the refined cleaning operation plan into a task decomposition module. The task decomposition module decomposes the cleaning operation task into multiple sub-task units, and each sub-task unit includes time window information, path coordinate information and equipment parameter information. Input the sub-task units into a task allocation module. The task allocation module dynamically allocates the sub-task units based on a load balancing algorithm, assigns a corresponding task sequence to each cleaning device, and the sub-task units in the task sequence are sorted according to the time window information. Input the task sequence into an execution control module. The execution control module generates a motion trajectory instruction for the cleaning device according to the path coordinate information, generates an operation control instruction for the cleaning device according to the equipment parameter information, and controls multiple cleaning devices to perform cleaning operations according to the optimal cleaning time arrangement. Collect multi-dimensional operation status data of the cleaning device during the cleaning operation through the edge computing node. Input the multi-dimensional operation status data into a status evaluation module. The status evaluation module calculates the task completion degree index and the equipment health degree index in real time. Dynamically adjust the task sequence based on the task completion degree index and the equipment health degree index, and re-allocate the adjusted task sequence to each cleaning device to continue the cleaning operation.

8. An intelligent cleaning decision-making system for a photovoltaic power station based on machine learning, which is used to implement the method described in any one of the foregoing claims 1-7, characterized in that, Including: A first unit for deploying a multi-dimensional sensor network to collect multi-source monitoring data of a photovoltaic power station. The multi-source monitoring data includes dust accumulation data on the surface of photovoltaic modules, power generation efficiency data of photovoltaic modules, environmental meteorological data and historical cleaning record data. Among them, the dust accumulation data is collected collaboratively by a distributed image sensor and a laser scattering sensor, the power generation efficiency data is collected by a current-voltage sensor array, and the environmental meteorological data includes multi-parameter combination data. A second unit for respectively inputting the multi-source monitoring data into a physical model and a data model, constructing a hybrid photovoltaic power station digital twin system, and using the hybrid photovoltaic power station digital twin system to output a dynamic evaluation result of the photovoltaic power station. According to the dynamic evaluation result, construct a multi-objective constrained deep reinforcement learning framework to train an intelligent cleaning decision-making model, and obtain the calculation result of the intelligent cleaning decision-making model. Based on the calculation results of the intelligent cleaning decision-making model, an adaptive cleaning threshold triggering mechanism is constructed. When the reduction value of the power generation efficiency of the photovoltaic module detected by the current-voltage sensor array is greater than the critical value of the dynamic power generation efficiency loss, a refined cleaning operation plan is generated using the spatio-temporal optimization algorithm; The third unit is configured to control multiple cleaning devices to perform cleaning operations according to the cleaning time sequence by using a distributed task scheduling algorithm based on the refined cleaning operation plan, and collect multi-dimensional operation status data through an edge computing node; input the multi-dimensional operation status data and the multi-source monitoring data into a transfer learning module, calculate the policy network parameters of the intelligent cleaning decision-making model by using an online incremental learning method, and update the calculated policy network parameters to the dual value network structure in the deep reinforcement learning framework.

9. An electronic device, characterized in that, Comprising: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, the method according to any one of claims 1 to 7 is implemented.

Images