Photovoltaic panel surface pollution degree image evaluation system

Abstract

The invention discloses a photovoltaic panel surface pollution degree image evaluation system. The system comprises an image acquisition module, an image processing module, a multi-task evaluation module and a pollution degree determination module. Multi-spectral image data of the surface of a photovoltaic panel are collected, after radiation calibration and atmospheric correction processing, spectral reflectivity features and spatial texture features are extracted and fused, a multi-task deep learning model is input, a pollution type classification result and a power generation efficiency loss weight of each pixel region are output, and a pollution type classification result of each pixel region is obtained. And finally, generating a pollution type spatial distribution diagram, calculating the overall power generation efficiency loss percentage, and determining the pollution degree grade. According to the method, the problem that the mixed pollution type cannot be distinguished and the differentiation influence cannot be evaluated in the prior art is solved, accurate quantitative evaluation of the pollution degree is realized, and a reliable basis is provided for fine operation and maintenance of a photovoltaic power station.

Description

TECHNICAL FIELD

[0001] The present application relates to an image evaluation system, in particular to a photovoltaic panel surface contamination degree image evaluation system. BACKGROUND

[0002] As a key support of clean energy, the power generation efficiency of photovoltaic power generation directly determines the energy utilization efficiency. Photovoltaic panels are exposed to the natural environment for a long period of time, and the surface is easy to accumulate dust, bird droppings, pollen and other pollutants. These contaminants can block light and reduce power generation efficiency. Studies have shown that severe contamination can reduce the power generation efficiency of photovoltaic systems by more than 20%, so accurate evaluation of the contamination degree of the surface of the photovoltaic panel is of great significance to improve the operation and maintenance efficiency.

[0003] Current image-based photovoltaic panel contamination evaluation methods are mainly divided into two categories: traditional image processing technology and single deep learning model. The traditional method collects photovoltaic panel images, extracts features such as gray scale and texture, and compares them with standard clean images to quantify the contamination degree. Although this method is simple to implement, it is significantly affected by lighting conditions and is difficult to distinguish between different types of contamination (such as dust, bird droppings, water stains, etc.). To overcome this limitation, existing technology introduces a convolutional neural network, which improves the accuracy of overall contamination coverage evaluation, but only evaluates the overall contamination coverage and cannot identify the differentiated impact of different types of contamination on power generation efficiency, still with significant defects.

[0004] This limitation is manifested in the following aspects: first, existing systems focus on quantifying the overall contamination coverage, ignoring the non-uniform impact of different types of contamination such as dust, bird droppings, and water stains on power generation efficiency, and the actual impact of different types of contamination varies significantly; second, the evaluation method relies too much on single visible light image data and lacks the fusion of multispectral data, resulting in a single feature extraction dimension and insufficient information representation; third, the correlation between the evaluation results and the actual power generation efficiency loss is weak, making it difficult to support the accurate formulation of operation and maintenance decisions. SUMMARY

[0005] The present application overcomes the shortcomings of the prior art and provides a photovoltaic panel surface contamination degree image evaluation system.

[0006] To achieve the above purpose, the technical solution adopted by the present application is as follows: a photovoltaic panel surface contamination degree image evaluation system, comprising: an image acquisition module, an image processing module, a multi-task evaluation module, and a contamination degree determination module.

[0007] The image acquisition module is used to acquire multispectral image data of the surface of the photovoltaic panel.

[0008] The image processing module is used to receive the multispectral image data and perform radiation calibration and atmospheric correction processing to obtain a multispectral image set.

[0009] The multi-task evaluation module is used to extract spectral reflectance features and spatial texture features from the multispectral set, and perform feature-level fusion to generate a fused feature vector; then the fused feature vector is input into the multi-task deep learning model to output the pollution type classification result and power generation efficiency loss weight for each pixel region.

[0010] The contamination level determination module generates a spatial distribution map of contamination types based on the contamination type classification results. Simultaneously, based on the power generation efficiency loss weights of all pixel regions, it uses a weighted summation algorithm to calculate the overall power generation efficiency loss percentage of the photovoltaic panel. Through a preset loss percentage-contamination level mapping relationship, it determines the contamination level of the photovoltaic panel surface.

[0011] In a preferred embodiment of the present invention, the image acquisition module includes a multispectral camera and an environmental parameter sensor. The multispectral camera is installed above or to the side of the photovoltaic panel array, and the environmental parameter sensor is used to simultaneously acquire ambient light intensity, ambient temperature, and humidity parameters. The multispectral image data includes spectral information of the visible light band and at least two non-visible light bands.

[0012] In a preferred embodiment of the present invention, the step of the multi-task evaluation module generating the fused feature vector includes:

[0013] S01. Receive a multispectral image set from the image processing module, calculate the spectral reflectance features for each pixel region, including extracting the reflectance values ​​of each band, forming a spectral reflectance curve, and calculating the mean reflectance, variance reflectance, spectral absorption depth, spectral absorption area, and spectral slope parameters from the curve.

[0014] S02. Spatial texture features are extracted in parallel. Local binary mode and gray-level co-occurrence matrix algorithm are used to calculate contrast, correlation, energy and homogeneity texture index.

[0015] S03. Perform feature-level fusion, concatenating spectral reflectance features and spatial texture feature parameters into a multi-dimensional fused feature vector.

[0016] In a preferred embodiment of the present invention, the fused feature vector is input into a multi-task deep learning model. The first subtask of the multi-task deep learning model outputs the probability distribution of the type of contamination in each pixel region, including the classification probability of dust, bird droppings, water stains, pollen and clean surfaces. The second subtask queries a preset contamination type-shading efficiency loss coefficient mapping table based on the classification probability to obtain the corresponding contamination type classification result and power generation efficiency loss weight.

[0017] In a preferred embodiment of the present invention, the process of generating the spatial distribution map of the dirt types includes:

[0018] S11. Receive the classification results of the dirt type of each pixel region output by the multi-task evaluation module, assign the dirt type with the highest probability to the corresponding pixel according to the classification probability, and generate a spatial distribution map of dirt type, in which different dirt types are visualized by different color codes.

[0019] S12. Calculate the coverage area ratio of each type of pollution by counting the number of pixels belonging to the same type of pollution and dividing it by the total number of pixels in the photovoltaic panel.

[0020] S13. Based on the power generation efficiency loss weights of all pixel regions, a weighted summation algorithm is used to calculate the overall power generation efficiency loss percentage of the photovoltaic panel. The formula is as follows:

[0021] ;

[0022] Where i represents the filth type index.

[0023] In a preferred embodiment of the present invention, the process of presetting the percentage loss-fouling level mapping relationship includes:

[0024] S21. Define the percentage loss - soiling level mapping relationship as a piecewise linear function, where the soiling level is divided into four levels, corresponding to clean, lightly soiled, moderately soiled and heavily soiled.

[0025] S22. Set threshold parameters for three stages. When the overall power generation efficiency loss percentage of the photovoltaic panel is less than the first threshold, the pollution level is determined to be Level 1; when the loss percentage is greater than or equal to the first threshold and less than the second threshold, it is Level 2; when it is greater than or equal to the second threshold and less than the third threshold, it is Level 3; when it is greater than or equal to the third threshold, it is Level 4.

[0026] S23. Based on historical operating data and cleaning cost-benefit analysis, dynamically adjust and optimize the threshold parameters of the three stages through machine learning algorithms to adapt to different photovoltaic power station environments. When determining the pollution level, combine the spatial distribution map of pollution type for auxiliary decision-making.

[0027] In a preferred embodiment of the present invention, the first threshold, the second threshold, and the third threshold are determined based on the actual operating data of the photovoltaic power station and the cost-benefit analysis of cleaning.

[0028] In a preferred embodiment of the present invention, the system further includes a feedback learning module, which receives actual power generation efficiency monitoring data, compares the actual power generation efficiency monitoring data with the percentage of overall power generation efficiency loss of the photovoltaic panel evaluated by the system, and calculates the evaluation error; when the evaluation error exceeds a preset threshold, a multi-task deep learning model update mechanism is triggered, and the multi-task deep learning model is incrementally trained using the latest monitoring data to adjust the parameters of the multi-task deep learning model.

[0029] In a preferred embodiment of the present invention, the spatial distribution map of contamination types, the percentage of overall power generation efficiency loss of photovoltaic panels, and the level of contamination on the surface of photovoltaic panels are integrated into a comprehensive assessment report, which is displayed through a graphical user interface or transmitted to a remote monitoring center through a network interface. At the same time, cleaning suggestions are generated based on the level of contamination. When the level of contamination reaches or exceeds a preset warning level, a cleaning work order is automatically generated and sent to the operation and maintenance personnel.

[0030] This invention addresses the shortcomings of the prior art and has the following beneficial effects:

[0031] (1) This invention uses a multispectral image acquisition module to acquire spectral information beyond the visible light range, and combines a feature extraction module to fuse spectral reflectance features and spatial texture features. The first subtask of the multi-task deep learning model is used to achieve pixel-level classification of dirt types, which enables accurate identification and spatial positioning of mixed dirt such as dust, bird droppings, water stains, and pollen. The generated dirt type spatial distribution map can intuitively show the distribution of various types of dirt, providing a visual basis for targeted cleaning.

[0032] (2) Through the second subtask of the multi-task deep learning model, combined with the pre-set pollution type-shading efficiency loss coefficient mapping table, the power generation efficiency loss weight of different pollution areas is calculated, and the overall power generation efficiency loss percentage is obtained by using a weighted summation algorithm. Different types of pollution have different light-blocking effects due to their different physical characteristics such as composition, thickness, and light transmittance. This invention establishes a quantitative relationship model between pollution type and power generation loss by experimentally determining the standard shading efficiency loss coefficient of various types of pollution, and realizes accurate power generation efficiency loss assessment based on pollution type and spatial distribution.

[0033] (3) This invention constructs a complete photovoltaic panel pollution assessment technology system by organically combining multispectral data acquisition, feature fusion, and multi-task evaluation. Data flow between modules is smooth, forming a closed-loop processing flow from raw data acquisition to final decision support. The system not only achieves accurate identification of pollution types but also establishes a precise mapping between pollution distribution and power generation efficiency loss, providing technical support for predictive maintenance of photovoltaic power plants. Simultaneously, the system's feedback learning mechanism can continuously optimize the assessment model based on actual operating data, ensuring accuracy and adaptability for long-term use. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 This is a system workflow diagram of a preferred embodiment of the present invention. Detailed Implementation

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein. Therefore, the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0038] It should be noted that in assessing surface contamination on photovoltaic panels, different types of contamination have significantly different impacts on power generation efficiency. The coverage area or visual thickness of the contamination alone cannot accurately reflect the actual degree of impact. For example, although locally concentrated bird droppings cover a small area, they completely block light transmission, creating severe hot spots on the photovoltaic panel and causing a complete loss of power generation efficiency in that area. In contrast, while uniformly distributed fine dust has a wider visual coverage, its partial light transmission means its impact on overall power generation efficiency is relatively limited. Furthermore, special contaminants such as water stains alter surface optical properties, and their impact cannot be quantified by simple area accumulation. Therefore, traditional assessment methods based on a single visual feature are insufficient to establish an accurate correlation between contamination distribution and power generation loss. A comprehensive consideration of the optical characteristics of contamination types, distribution patterns, and their differentiated impacts on the photoelectric conversion process is necessary to achieve a scientific assessment of the degree of contamination.

[0039] like Figure 1 As shown, the photovoltaic panel surface contamination level image assessment system includes: an image acquisition module, an image processing module, a multi-task assessment module, and a contamination level determination module;

[0040] The image acquisition module is used to acquire multispectral image data of the photovoltaic panel surface.

[0041] To further explain, the specific execution steps of the image acquisition module include:

[0042] Multispectral image data of the photovoltaic panel surface is acquired by a multispectral camera, which includes multiple optical sensor arrays, each corresponding to the acquisition of spectra in different bands, including at least the 450~510nm blue light band, the 520~590nm green light band, the 600~680nm red light band, the 700~780nm near-infrared band, and the 800~900nm short-wave infrared band.

[0043] The spatial position and shooting angle of the multispectral camera are adjusted using an auxiliary positioning device. The camera attitude is monitored in real time by a gyroscope and an angle sensor to ensure that the angle between the shooting angle and the normal of the photovoltaic panel surface is controlled within a range of less than 10 degrees, so as to minimize perspective distortion.

[0044] The environmental parameter sensor is triggered synchronously to collect ambient light intensity, ambient temperature and humidity parameters. The ambient light intensity data is used for illumination normalization in subsequent image processing, and the ambient temperature and humidity data are used for atmospheric correction compensation.

[0045] Record the shooting timestamp and photovoltaic panel identification information, including the photovoltaic panel number, installation location and battery type, and store this metadata in association with the multispectral image data;

[0046] Multispectral image data, environmental parameters, timestamps, and identification information are integrated into a structured data packet and transmitted to the image processing module through a data interface to ensure data integrity and traceability.

[0047] The image processing module is used to receive the multispectral image data and perform radiometric calibration and atmospheric correction processing to obtain a multispectral image set.

[0048] To further explain, the specific execution steps of the image processing module include:

[0049] It receives multispectral image data from the image acquisition module, performs radiometric calibration on the image of each band, and converts the digital quantization value of the image into surface reflectance data according to the pre-stored radiometric calibration coefficients, where the radiometric calibration coefficients are obtained through laboratory calibration and include gain and offset parameters.

[0050] Atmospheric correction processing is performed using an atmospheric correction algorithm based on a radiative transfer model. The atmospheric scattering and absorption effects are compensated by synchronously collected ambient light intensity, temperature and humidity parameters. The output is a standardized multispectral image, where the value of each pixel represents the true reflectance, and the reflectance range is normalized to between 0 and 1.

[0051] Image registration is performed, and a feature point matching algorithm is used for multispectral images of different bands to ensure that the pixel coordinates of the same geographical location are consistent in images of different bands, and the registration error is controlled at the sub-pixel level.

[0052] Noise filtering and image enhancement processing are performed on the registered multispectral images. Median filtering is used to remove salt-and-pepper noise, and histogram equalization is used to improve image contrast, generating a high-quality multispectral image set.

[0053] The processed multispectral image set is stored in the cache and a data interface is provided for the multi-task evaluation module to ensure smooth data transmission.

[0054] The multi-task evaluation module is used to extract spectral reflectance features and spatial texture features from the multispectral set, and perform feature-level fusion to generate a fused feature vector; then the fused feature vector is input into the multi-task deep learning model to output the pollution type classification result and power generation efficiency loss weight for each pixel region.

[0055] To further clarify, the specific execution steps for feature extraction and fusion in the multi-task evaluation module include:

[0056] The image processing module receives a set of multispectral images, calculates the spectral reflectance features for each pixel region, including extracting the reflectance values ​​of each band, forming a spectral reflectance curve, and calculating the mean reflectance, variance reflectance, spectral absorption depth, spectral absorption area, and spectral slope parameters from the curve.

[0057] Spatial texture features are extracted in parallel. Local binary mode and gray-level co-occurrence matrix algorithms are used to calculate contrast, correlation, energy and homogeneity texture indices from visible light and near-infrared images.

[0058] Feature-level fusion is performed, concatenating spectral reflectance feature parameters and spatial texture feature parameters into a multi-dimensional fused feature vector, and then normalizing the vector to eliminate dimensional differences.

[0059] The fused feature vector is input into a multi-task deep learning model. The first subtask of the model outputs the probability distribution of the type of pollution in each pixel region, including the classification probability of dust, bird droppings, water stains, pollen and clean surfaces. The second subtask queries a pre-set pollution type-shading efficiency loss coefficient mapping table based on the classification probability to obtain the power generation efficiency loss weight of the corresponding pollution type.

[0060] The output is post-processed, including applying probability thresholds to filter low-confidence classifications and smoothing loss weight data to reduce the impact of noise.

[0061] Furthermore, the structure and operation steps of the multi-task deep learning model in the multi-task evaluation module include:

[0062] A multi-task deep learning model is constructed, which adopts a shared low-level feature extraction network and a branch-specific network structure. The shared network is a deep convolutional neural network, including convolutional layers, pooling layers and activation functions, which are used to extract high-level semantic features from the fused feature vector.

[0063] The first subtask, the filth type classification subnetwork, consists of fully connected layers and a softmax activation function. It outputs the probability value of each pixel region belonging to each filth type, and the probability value is optimized by the cross-entropy loss function.

[0064] The second subtask, the power generation efficiency loss assessment subnetwork, receives the output of the first subtask and queries a pre-set pollution type-shading efficiency loss coefficient mapping table. The shading efficiency loss coefficient is determined by experimentally measuring the photoelectric conversion efficiency loss of different pollution types under standard test conditions.

[0065] During model training, an end-to-end learning approach is adopted, while simultaneously optimizing classification loss and regression loss. The adaptive moment estimation algorithm is used to update network parameters, and overfitting is prevented by monitoring the validation set.

[0066] During model inference, GPUs are used to accelerate computation, and the pollution type classification results and power generation efficiency loss weights for each pixel region are output in real time, supporting large-scale image processing.

[0067] The training and optimization steps for multi-task deep learning models include:

[0068] Collect multispectral image sample data with accurate annotations. The annotations include the actual pollution type label for each pixel region and the corresponding measured value of power generation efficiency loss. The measured values ​​are obtained by a portable IV tester under synchronous conditions.

[0069] Preprocess the sample data, including data augmentation operations such as rotation, scaling and color jitter, to increase the diversity of the dataset and improve the model's generalization ability;

[0070] The dataset is randomly divided into training, validation and test sets. The training set is used for model parameter learning, the validation set is used for hyperparameter tuning, and the test set is used for final performance evaluation.

[0071] Define a multi-task loss function that combines the cross-entropy loss of the first subtask and the mean squared error loss of the second subtask. The total loss function is a weighted sum, and the weights are dynamically adjusted according to the importance of the tasks.

[0072] During the training process, the network parameters are iteratively updated using batch gradient descent. After each training round, the accuracy of the validation set is calculated. Early stopping is used to terminate training when performance no longer improves, and the optimal model parameters are saved.

[0073] The contamination level determination module generates a spatial distribution map of contamination types based on the contamination type classification results. Simultaneously, based on the power generation efficiency loss weights of all pixel regions, it uses a weighted summation algorithm to calculate the overall power generation efficiency loss percentage of the photovoltaic panel. Through a preset loss percentage-contamination level mapping relationship, it determines the contamination level of the photovoltaic panel surface.

[0074] To further clarify, the specific steps in the pollution level determination module for generating a spatial distribution map of pollution types and calculating the percentage of power generation efficiency loss include:

[0075] The system receives the classification results of the dirt type for each pixel region from the multi-task evaluation module, assigns the dirt type with the highest probability to the corresponding pixel according to the classification probability, and generates a spatial distribution map of dirt types, in which different dirt types are visualized using different color codes.

[0076] The coverage area ratio of each type of pollution is calculated by counting the number of pixels belonging to the same type of pollution and dividing it by the total number of pixels in the photovoltaic panel.

[0077] Based on the power generation efficiency loss weights of all pixel regions, a weighted summation algorithm is used to calculate the overall power generation efficiency loss percentage of the photovoltaic panel, as shown in the formula:

[0078] ;

[0079] Where i represents the filth type index.

[0080] Verify the calculation results, including checking the integrity and consistency of the data, ensuring that the sum of the area ratios is 1 and the percentage of loss is within a reasonable range;

[0081] The spatial distribution map of pollution types and the percentage of power generation efficiency loss are temporarily stored, and a data interface is provided for the output module.

[0082] The specific definition and application steps of the preset loss percentage-fouling level mapping relationship in the soiling degree determination module include:

[0083] The percentage loss-fouling level mapping relationship is defined as a piecewise linear function, where the fouling level is divided into four levels, corresponding to clean, lightly soiled, moderately soiled, and heavily soiled.

[0084] Set threshold parameters, assuming the first threshold is 2%, the second threshold is 5%, and the third threshold is 10%. When the overall power generation efficiency loss percentage of the photovoltaic panel is less than 2%, the pollution level is determined to be Level 1; when the loss percentage is greater than or equal to 2% and less than 5%, it is Level 2; when it is greater than or equal to 5% and less than 10%, it is Level 3; and when it is greater than or equal to 10%, it is Level 4.

[0085] The threshold parameters are dynamically adjusted based on historical operating data and cleaning cost-benefit analysis. The thresholds are optimized through machine learning algorithms to adapt to different photovoltaic power plant environments.

[0086] When determining the level of filth, the spatial distribution map of filth types can be used to assist in decision-making. For example, when local filth such as bird droppings accounts for a high proportion, the level of filth should be appropriately increased.

[0087] Output the contamination level result along with a confidence score, which is calculated based on the consistency of classification probability and loss weight.

[0088] The system also includes a feedback learning module, whose specific execution steps include:

[0089] Receive actual power generation efficiency monitoring data, obtain real-time power generation data from photovoltaic inverters or monitoring systems, and calculate actual power generation efficiency loss;

[0090] The actual power generation efficiency loss is compared with the percentage of overall power generation efficiency loss of the photovoltaic panels as assessed by the system, and the absolute error and relative error are calculated.

[0091] When the error exceeds a preset threshold (e.g., 5%), the model update mechanism is triggered to incrementally train the multi-task deep learning model using the latest data and adjust the network parameters.

[0092] Dynamically update the mapping table of pollution type-shading efficiency loss coefficient and the mapping relationship of loss percentage-pollution level, and recalibrate the coefficients and thresholds based on long-term monitoring data;

[0093] Record feedback data to a historical database to support model version management and performance tracking, and continuously improve the accuracy of system evaluation.

[0094] The system also includes an output module, whose specific execution steps include:

[0095] Receive the pollution level determination module outputting the spatial distribution map of pollution types, the percentage of overall power generation efficiency loss of photovoltaic panels, and the pollution level level;

[0096] The data is integrated to generate a comprehensive assessment report, which includes a text summary, visualizations, and recommended measures. The report is displayed through a graphical user interface or transmitted to a remote monitoring center via a network interface.

[0097] Cleaning suggestions are automatically generated based on the level of contamination. When the level reaches level three or above, a cleaning work order is created and sent to the maintenance personnel. The work order includes the location of the photovoltaic panels, the type of contamination, and the urgency level.

[0098] Store historical assessment data in the database, support querying the trend of pollution level changes by time series, and provide a pollution accumulation rate prediction function based on the ARIMA model;

[0099] It provides a data export interface, supporting report output in PDF and Excel formats, which facilitates subsequent analysis and decision-making.

[0100] Based on the preferred embodiments of the present invention described above, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A photovoltaic panel surface contamination level image assessment system, characterized in that, include: Image acquisition module, image processing module, multi-task evaluation module, and contamination level determination module; The image acquisition module is used to acquire multispectral image data of the photovoltaic panel surface; The image processing module is used to receive the multispectral image data and perform radiometric calibration and atmospheric correction processing to obtain a multispectral image set. The multi-task evaluation module is used to extract spectral reflectance features and spatial texture features from the multispectral set, and perform feature-level fusion to generate a fused feature vector; then the fused feature vector is input into the multi-task deep learning model to output the pollution type classification result and power generation efficiency loss weight for each pixel region. The contamination level determination module generates a spatial distribution map of contamination types based on the contamination type classification results. Simultaneously, based on the power generation efficiency loss weights of all pixel regions, it uses a weighted summation algorithm to calculate the overall power generation efficiency loss percentage of the photovoltaic panel. Through a preset loss percentage-contamination level mapping relationship, it determines the contamination level of the photovoltaic panel surface.

2. The photovoltaic panel surface contamination level image assessment system according to claim 1, characterized in that: The image acquisition module includes a multispectral camera and an environmental parameter sensor. The multispectral camera is installed above or to the side of the photovoltaic panel array, and the environmental parameter sensor is used to simultaneously acquire ambient light intensity, ambient temperature, and humidity parameters. The multispectral image data includes spectral information of the visible light band and at least two non-visible light bands.

3. The photovoltaic panel surface contamination level image assessment system according to claim 1, characterized in that: The steps of generating the fused feature vector by the multi-task evaluation module include: S01. Receive a multispectral image set from the image processing module, calculate the spectral reflectance features for each pixel region, including extracting the reflectance values ​​of each band, forming a spectral reflectance curve, and calculating the mean reflectance, variance reflectance, spectral absorption depth, spectral absorption area, and spectral slope parameters from the curve. S02. Spatial texture features are extracted in parallel. Local binary mode and gray-level co-occurrence matrix algorithm are used to calculate contrast, correlation, energy and homogeneity texture index. S03. Perform feature-level fusion, concatenating spectral reflectance features and spatial texture feature parameters into a multi-dimensional fused feature vector.

4. The photovoltaic panel surface contamination level image assessment system according to claim 3, characterized in that: The fused feature vectors are input into a multi-task deep learning model. The first subtask of the multi-task deep learning model outputs the probability distribution of the type of contamination for each pixel region, including the classification probabilities of dust, bird droppings, water stains, pollen, and clean surfaces. The second subtask queries a pre-set contamination type-shading efficiency loss coefficient mapping table based on the classification probabilities to obtain the corresponding contamination type classification results and power generation efficiency loss weights.

5. The photovoltaic panel surface contamination level image assessment system according to claim 1, characterized in that: The process of generating the spatial distribution map of the pollution types includes: S11. Receive the classification results of the dirt type of each pixel region output by the multi-task evaluation module, assign the dirt type with the highest probability to the corresponding pixel according to the classification probability, and generate a spatial distribution map of dirt type, in which different dirt types are visualized by different color codes. S12. Calculate the coverage area ratio of each type of pollution by counting the number of pixels belonging to the same type of pollution and dividing it by the total number of pixels in the photovoltaic panel. S13. Based on the power generation efficiency loss weights of all pixel regions, calculate the overall power generation efficiency loss percentage of the photovoltaic panel using a weighted summation algorithm. The formula is as follows: ； Where i represents the filth type index.

6. The photovoltaic panel surface contamination level image assessment system according to claim 1, characterized in that: The pre-setting process for the percentage loss - contamination level mapping relationship includes: S21. Define the percentage loss - soiling level mapping relationship as a piecewise linear function, where the soiling level is divided into four levels, corresponding to clean, lightly soiled, moderately soiled and heavily soiled. S22. Set threshold parameters for three stages. When the overall power generation efficiency loss percentage of the photovoltaic panel is less than the first threshold, the pollution level is determined to be Level 1; when the loss percentage is greater than or equal to the first threshold and less than the second threshold, it is Level 2; when it is greater than or equal to the second threshold and less than the third threshold, it is Level 3; when it is greater than or equal to the third threshold, it is Level 4. S23. Based on historical operating data and cleaning cost-benefit analysis, dynamically adjust and optimize the threshold parameters of the three stages through machine learning algorithms to adapt to different photovoltaic power station environments. When determining the pollution level, combine the spatial distribution map of pollution type for auxiliary decision-making.

7. The photovoltaic panel surface contamination level image assessment system according to claim 6, characterized in that: The first threshold, the second threshold, and the third threshold are determined based on the actual operating data of the photovoltaic power station and the cost-benefit analysis of cleaning.

8. The photovoltaic panel surface contamination level image assessment system according to claim 1, characterized in that: The system also includes a feedback learning module, which receives actual power generation efficiency monitoring data, compares the actual power generation efficiency monitoring data with the percentage of overall power generation efficiency loss of photovoltaic panels evaluated by the system, and calculates the evaluation error. When the evaluation error exceeds a preset threshold, a multi-task deep learning model update mechanism is triggered, which uses the latest monitoring data to incrementally train the multi-task deep learning model and adjust the parameters of the multi-task deep learning model.

9. The photovoltaic panel surface contamination level image assessment system according to claim 1, characterized in that: The system integrates the spatial distribution map of contamination types, the percentage of overall power generation efficiency loss of photovoltaic panels, and the level of contamination on the surface of photovoltaic panels into a comprehensive assessment report, which is displayed through a graphical user interface or transmitted to a remote monitoring center via a network interface. At the same time, it also generates cleaning suggestions based on the level of contamination. When the level of contamination reaches or exceeds the preset warning level, it automatically generates a cleaning work order and sends it to the operation and maintenance personnel.

Images