Photovoltaic panel dirt state identification method and system based on dust density and visual diagnosis
By combining a CNN-BiLSTM-Attention hybrid neural network and an improved YOLOv8n model with dust accumulation density and visual diagnostics, the accuracy and efficiency issues of photovoltaic panel dirt status identification are solved, enabling high-precision dirt type identification and providing cleaning strategies.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANJING INST OF TECH
- Filing Date
- 2026-01-19
- Publication Date
- 2026-05-15
AI Technical Summary
Existing methods for identifying the state of dirt on photovoltaic panels in photovoltaic power plants suffer from problems such as excessive human intervention, susceptibility to subjective factors, large computational load, and low accuracy, making it difficult to accurately judge subtle changes in dust accumulation and provide targeted cleaning strategies.
A CNN-BiLSTM-Attention hybrid neural network prediction model combined with an improved YOLOv8n model was adopted. A two-stage logic recognition of the dirt status of photovoltaic panels was constructed through ash density and visual diagnosis. The mapping relationship was established by experimentally analyzing the impact of ash on output power. The improved YOLOv8n model was then used to identify the type of dirt.
It enables accurate identification of the dirt status of photovoltaic panels, improves detection accuracy, provides quantitative and qualitative basis, reduces resource waste, and improves the targeting and safety of cleaning.
Smart Images
Figure CN122049463A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of photovoltaic power generation technology, and in particular to a method and system for identifying the state of dirt on photovoltaic panels based on dust accumulation density and visual diagnosis. Background Technology
[0002] With the increasing global demand for renewable energy, photovoltaic (PV) power plants have developed rapidly worldwide. However, PV power plants operate outdoors for extended periods, and fine particles in the air easily accumulate on the PV panels, forming dirt. Without targeted cleaning, this not only reduces the light transmittance and output power of the PV panels but also alters the heat transfer mechanism, affecting heat dissipation performance. Therefore, identifying the dirt condition of PV panels has become a necessary and challenging task in the daily cleaning and maintenance of PV power plants.
[0003] Currently, the main methods for identifying the state of dirt on photovoltaic panels are as follows:
[0004] Direct observation method: This method is the most direct, but it requires a lot of human intervention. At the same time, it is difficult to carry out in photovoltaic power stations on complex terrain (such as mountains and seas), and human observation is easily affected by subjective factors, making it difficult to accurately judge subtle changes in dust density.
[0005] Image processing method: This method can detect the presence of dirt by processing and analyzing the surface image of the photovoltaic panel, but it is easily affected by light and shadow and has high requirements for model training.
[0006] Big data analysis and prediction: This method predicts the degree of dust accumulation by analyzing historical photovoltaic operating data. However, it involves a large amount of computation and suffers from poor analysis accuracy and low efficiency. Summary of the Invention
[0007] To address the shortcomings of existing technologies, this invention provides a method and system for identifying the state of dirt on photovoltaic panels based on dust accumulation density and visual diagnostics.
[0008] The present invention achieves the above-mentioned technical objectives through the following technical means.
[0009] A method for identifying the fouling status of photovoltaic panels based on dust density and visual diagnostics includes:
[0010] Phase 1: Collect and preprocess actual output power and environmental data of photovoltaic panels; construct a CNN-BiLSTM-Attention hybrid neural network prediction model and conduct experiments on the impact of dust accumulation on the output power of photovoltaic panels; detect the dust accumulation density of photovoltaic panels based on the prediction model and the impact experiments.
[0011] The second stage: When the dust accumulation density exceeds the trigger threshold, visible light images of the photovoltaic panel surface are collected and processed to form a dataset. The type of dirt is identified by improving the YOLOv8n model. The improved YOLOv8n model introduces the NAM lightweight attention mechanism, the RFBNet feature extraction module, and the WloU loss function.
[0012] Furthermore, the environmental data includes temperature, humidity, irradiance, and wind speed.
[0013] Furthermore, the specific process of dust accumulation density detection is as follows: First, the theoretical output power is obtained through the CNN-BiLSTM-Attention hybrid neural network prediction model, and the output power loss is obtained by calculating the difference between the theoretical output power and the actual output power; then, the mapping relationship between dust accumulation density and output power loss is obtained through experiments on the impact of dust accumulation on the output power of photovoltaic panels; finally, the dust accumulation density is obtained by combining the output power loss and the mapping relationship.
[0014] Furthermore, the mapping relationship expression is as follows: ,in, D represents the output power loss, and D represents the ash density.
[0015] Furthermore, the trigger threshold is set according to the on-site requirements of the photovoltaic power station.
[0016] Furthermore, the improvements to the YOLOv8n model include: the YOLOv8n model includes Input, Backbone, Neck, and Head; a lightweight NAM attention mechanism is introduced; the C2f modules in the Backbone and Neck are replaced with C2f-NAM modules; and an RFBNet feature extraction module is introduced at the end of the Backbone. At the same time, the WloU loss function is used instead of the CloU loss function used by YOLOv8n.
[0017] Furthermore, the performance indicators of the ash density detection method include the difference between the actual ash density and the predicted ash density, and MAPE; the performance indicators of the dirt type identification method include recall and accuracy.
[0018] A photovoltaic panel dirt status identification system based on dust density and visual diagnostics includes:
[0019] The data acquisition and preprocessing module is used to collect meteorological data of the area where the photovoltaic panel is located, output power data of the photovoltaic panel under clean conditions, visible light images of the photovoltaic panel surface, and to preprocess the collected data.
[0020] The model building module is used to build a CNN-BiLSTM-Attention hybrid neural network prediction model for ash density detection and an improved YOLOv8n model for different types of dirt identification.
[0021] The model training and evaluation module is used to train the CNN-BiLSTM-Attention hybrid neural network prediction model and the improved YOLOv8n model using the training dataset as input data, and to set metrics to evaluate the performance of the two trained models.
[0022] The detection application module is used to predict the dust accumulation density through the CNN-BiLSTM-Attention hybrid neural network prediction model, the mapping relationship between dust accumulation density and output power loss, and to determine whether the dust accumulation density exceeds the trigger threshold. If it exceeds the trigger threshold, the module uses the improved YOLOv8n model to determine the specific type of dirt on the photovoltaic panel.
[0023] An electronic device includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the program to implement a method for identifying the state of dirt on photovoltaic panels based on dust density and visual diagnosis.
[0024] A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements a method for identifying the state of dirt on photovoltaic panels based on dust density and visual diagnostics.
[0025] The beneficial effects of this invention are:
[0026] (1) In the first stage of this invention, a CNN-BiLSTM-Attention hybrid neural network prediction model was adopted, which has a significant advantage in prediction accuracy compared with the traditional single neural network model. The degree of dirt on the surface of the photovoltaic panel can be accurately determined by the prediction model and the experiment on the impact of dust accumulation on the output power of the photovoltaic panel. The method is robust and provides a reliable quantitative basis for subsequent intervention decisions.
[0027] (2) The second stage of this invention constructs a two-stage logic of "quantitative + qualitative" to realize the identification of dirt status in all dimensions and effectively save detection resources. Using the dust accumulation density as the visual detection trigger threshold, the improved YOLOv8n model achieves an accuracy of 94.2% to address the problems of small targets being easily missed and the large amount of visual detection data being difficult to quantify. It can accurately identify a variety of dirt, solve the problem of targeted cleaning, and provide qualitative basis. Attached Figure Description
[0028] Figure 1 This is a flowchart of the photovoltaic panel dirt status identification method described in this invention.
[0029] Figure 2This is a diagram of the CNN-BiLSTM-Attention hybrid neural network prediction model framework described in this invention.
[0030] Figure 3 This is a mapping diagram of the ash accumulation density and output power loss described in this invention.
[0031] Figure 4 This is a framework diagram of the improved YOLOv8n model described in this invention.
[0032] Figure 5(a) is a schematic diagram of the accuracy of the YOLOv8n model before the improvement described in this invention.
[0033] Figure 5(b) is a schematic diagram of the recall rate of the YOLOv8n model before the improvement described in this invention.
[0034] Figure 6(a) is a schematic diagram of the accuracy of the improved YOLOv8n model described in this invention.
[0035] Figure 6(b) is a schematic diagram of the recall rate of the improved YOLOv8n model described in this invention. Detailed Implementation
[0036] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but the scope of protection of the present invention is not limited thereto.
[0037] A method for identifying the state of dirt on photovoltaic panels based on dust density and visual diagnostics, such as... Figure 1 As shown, it includes:
[0038] In the first phase, for each photovoltaic panel that needs to be identified as clean, the actual output power data of the photovoltaic panel is collected in real time using a data acquisition module; and environmental data (temperature, humidity, irradiance, wind speed, etc.) of the area where the target photovoltaic panel is located is collected in real time through a meteorological monitoring station.
[0039] After preprocessing the collected data, it is input into the CNN-BiLSTM-Attention hybrid neural network prediction model to obtain the theoretical output power under clean photovoltaic conditions on the predicted day. Figure 2As shown, this model utilizes CNN to extract spatial features from multidimensional data, BiLSTM to mine time series features, and an Attention mechanism to focus on key influencing factors. Root mean square error (RMSE) and mean absolute percentage error (MAPE) are used to measure the prediction accuracy of various models (CNN, LSTM, CNN-LSTM, CNN-BiLSTM-Attention). The CNN-BiLSTM-Attention model performs best: for both RMSE and MAPE, this model has the lowest values across various weather types (sunny, cloudy, and rainy), as shown in Table 1. The results show that it maintains stable low errors regardless of whether it is sunny with sufficient sunlight or fluctuating cloudy and rainy conditions, indicating that the model has good robustness.
[0040] Table 1 Performance index values of each neural network prediction model
[0041]
[0042] First, the difference between the theoretical output power predicted by the model and the actual output power is calculated to analyze the real-time output power loss. Then, the mapping relationship between ash density and output power loss is obtained through an experiment on the impact of ash accumulation on the output power of photovoltaic panels, thus obtaining the ash density of the photovoltaic panels. Specifically, the experiment on the impact of ash accumulation on the output power of photovoltaic panels involves preparing six identical photovoltaic panels. For three of these panels, during the midday period (11:00-13:30) when radiation is strongest and less volatile, 5.28g of fly ash is evenly applied to each panel every fifteen minutes, increasing the ash density of each panel sequentially to 5, 10, 15, 20, 25, 30, 35, 40, 45, and 50 g / m³. 2 The other three photovoltaic panels remained clean at all times. The average output power loss under different dust accumulation densities was analyzed and compared, and a linear regression method was used to derive the mapping relationship between dust accumulation density and output power loss, such as... Figure 3 As shown, the expression is: ,in, D represents the output power loss, and D represents the ash density.
[0043] The performance of the above-mentioned ash density detection method was experimentally verified under sunny and cloudy weather conditions: fly ash was covered on photovoltaic panels under sunny and cloudy weather conditions, respectively. The theoretical output power of the photovoltaic panels was obtained through the prediction model. The theoretical output power and the actual output power were compared and analyzed. The ash density was predicted by combining the mapping relationship between ash density and output power loss. The predicted ash density was compared with the actual fly ash density. The effectiveness of the ash density detection method was verified by the difference and MAPE. See Table 2 for details.
[0044] Table 2 Performance index values of ash density testing method
[0045]
[0046] In the second stage, some contaminants (such as bird droppings and sticky dust) will remain on the photovoltaic panels if not addressed promptly, making subsequent cleaning difficult and potentially leading to hot spots and safety risks. Therefore, it is best to identify the type of contaminant in advance and clean it accordingly to minimize revenue loss. However, relying solely on dust density cannot determine the specific cause of the contaminant and prevents targeted human intervention. Therefore, this embodiment uses the calculated "dust density" as a trigger threshold. The threshold is set according to the requirements of the photovoltaic power station site, with a range of 8-15 g / m³. 2 When the detected dust density exceeds the trigger threshold, the system activates the visual detection subsystem, launching a drone to precisely acquire images of the abnormal area. The drone captures visible light images of the photovoltaic panel surface, obtaining image data encompassing different lighting conditions, different types of dirt, and different photovoltaic module types. The image data includes 400 images of bird droppings, 300 images of leaves, 400 images of sea salt, and 400 images of floating dust. Operations such as horizontal flipping, random cropping, and scaling are used to expand the dataset to 6000 images.
[0047] Data augmentation was used to randomly arrange and crop parts of images with different types of dirt, fusing them into multiple new images to form a dataset. This not only improved the diversity and comprehensiveness of the dataset but also enhanced the generalization ability of subsequent models. The fused dataset was then labeled with LabelMe bounding boxes: dirt on the photovoltaic panel surface included bird droppings, leaves, sea salt, and dust. Using tools provided by LabelMe, regions were drawn on the images, and labels ("BD", "leaf", "salt", "dust") were assigned to the dirt (bird droppings, leaves, sea salt, dust). After labeling, the annotation information was saved in TXT format, and the dataset was divided into training, validation, and test sets in a 7:2:1 ratio.
[0048] An improved YOLOv8n model was constructed to identify dirt categories, such as... Figure 4As shown, the YOLOv8n model consists of an input network, a backbone network, a neck network, and a head network. The backbone is mainly responsible for extracting image features, using a CSPDarknet structure, with the core component being the C2f module (used to extract rich gradient flow information). The neck adopts a PANet structure, containing top-down and bottom-up paths for fusing feature maps of different scales. The head uses a decoupled structure, processing classification and regression tasks separately, and finally outputting the target's category and bounding box location. This embodiment introduces a lightweight NAM attention mechanism, replacing the C2f module in the original backbone and neck with a C2f-NAM module, and introducing an RFBNet feature extraction module at the end of the backbone, enhancing the model's ability to extract features of different scales of dust accumulation on photovoltaic panels. Simultaneously, the WloU loss function is used instead of the default CloU loss function used by YOLOv8n, improving the model's training performance and solving problems such as complex background environments in photovoltaic power plants, low detection accuracy, and susceptibility to missed detections and false detections.
[0049] The improved YOLOv8n model was trained using 70% of the data, and validated using 20% of the data, with parameters adjusted during training. Additionally, the accuracy and recall of the unimproved and improved YOLOv8n models were compared using the 10% test set. As shown in Figures 5(a) and 5(b), the unimproved YOLOv8n model had an average accuracy of 88.3% and an average recall of 90%. As shown in Figures 6(a) and 6(b), the improved YOLOv8n model had an average accuracy of 94.2% and an average recall of 91%, representing improvements of 5.9% and 1%, respectively. This demonstrates that the improved YOLOv8n model can accurately identify the specific types of dirt causing the power drop in the first stage, thus providing a qualitative basis for subsequent on-site personnel to develop differentiated cleaning strategies, such as spraying for floating dust and targeted scrubbing for bird droppings.
[0050] A photovoltaic panel dirt status identification system based on dust density and visual diagnostics includes:
[0051] The data acquisition and preprocessing module is used to collect meteorological data of the area where the photovoltaic panel is located, the output power data of the photovoltaic panel under clean conditions, and the visible light image of the photovoltaic panel surface, and to preprocess the collected data.
[0052] The model building module is used to build a CNN-BiLSTM-Attention hybrid neural network prediction model for ash density detection and an improved YOLOv8n model for different types of dirt recognition.
[0053] The model training and evaluation module is used to train the CNN-BiLSTM-Attention hybrid neural network prediction model and the improved YOLOv8n model using the training dataset as input data, and to set metrics to evaluate the performance of the two trained models.
[0054] The detection application module is used to predict the dust accumulation density through the CNN-BiLSTM-Attention hybrid neural network prediction model and the mapping relationship between dust accumulation density and output power loss. It determines whether the dust accumulation density exceeds the trigger threshold. If it exceeds the trigger threshold, it determines the specific type of dirt on the photovoltaic panel by improving the YOLOv8n model, thereby achieving the purpose of monitoring the dirt status of the photovoltaic panel.
[0055] An electronic device includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the program to implement the above-described method for identifying the state of dirt on photovoltaic panels based on dust density and visual diagnosis.
[0056] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method for identifying the state of dirt on photovoltaic panels based on dust accumulation density and visual diagnostics.
[0057] The embodiments described above are preferred embodiments of the present invention, but the present invention is not limited to the above embodiments. Any obvious improvements, substitutions or modifications that can be made by those skilled in the art without departing from the essence of the present invention shall fall within the protection scope of the present invention.
Claims
1. A method for identifying the state of dirt on photovoltaic panels based on dust accumulation density and visual diagnosis, characterized in that, include: Phase 1: Collect and preprocess actual output power and environmental data of photovoltaic panels; construct a CNN-BiLSTM-Attention hybrid neural network prediction model and conduct experiments on the impact of dust accumulation on the output power of photovoltaic panels; detect the dust accumulation density of photovoltaic panels based on the prediction model and the impact experiments. The second stage: When the dust accumulation density exceeds the trigger threshold, visible light images of the photovoltaic panel surface are collected and processed to form a dataset. The type of dirt is identified by improving the YOLOv8n model. The improved YOLOv8n model introduces the NAM lightweight attention mechanism, the RFBNet feature extraction module, and the WloU loss function.
2. The photovoltaic panel dirt status identification method based on dust accumulation density and visual diagnosis according to claim 1, characterized in that, The environmental data includes temperature, humidity, irradiance, and wind speed.
3. The photovoltaic panel dirt status identification method based on dust density and visual diagnosis according to claim 1, characterized in that, The specific process for detecting the ash density is as follows: First, the theoretical output power is obtained through the CNN-BiLSTM-Attention hybrid neural network prediction model, and the output power loss is obtained by calculating the difference between the theoretical output power and the actual output power. Then, the mapping relationship between ash density and output power loss is obtained through experiments on the impact of ash on the output power of photovoltaic panels. Finally, the ash density is obtained by combining the output power loss and the mapping relationship.
4. The photovoltaic panel dirt status identification method based on dust accumulation density and visual diagnosis according to claim 3, characterized in that, The mapping relationship expression is: ,in, D represents the output power loss, and D represents the ash density.
5. The photovoltaic panel fouling condition identification method based on dust accumulation density and visual diagnosis according to claim 1, characterized in that, The trigger threshold is set according to the on-site requirements of the photovoltaic power station.
6. The photovoltaic panel dirt status identification method based on dust density and visual diagnosis according to claim 1, characterized in that, The improvements to the YOLOv8n model include: the YOLOv8n model includes Input, Backbone, Neck, and Head; the introduction of the NAM lightweight attention mechanism; the replacement of the C2f modules in the Backbone and Neck with C2f-NAM modules; and the introduction of an RFBNet feature extraction module at the end of the Backbone; at the same time, the WloU loss function is used instead of the CloU loss function used by YOLOv8n.
7. The photovoltaic panel dirt status identification method based on dust accumulation density and visual diagnosis according to claim 1, characterized in that, The performance indicators of the ash density detection method include the difference between the actual ash density and the predicted ash density, and MAPE; the performance indicators of the dirt type identification method include recall and accuracy.
8. A photovoltaic panel dirt status identification system based on dust accumulation density and visual diagnosis, characterized in that, include: The data acquisition and preprocessing module is used to collect meteorological data of the area where the photovoltaic panel is located, output power data of the photovoltaic panel under clean conditions, visible light images of the photovoltaic panel surface, and to preprocess the collected data. The model building module is used to build a CNN-BiLSTM-Attention hybrid neural network prediction model for ash density detection and an improved YOLOv8n model for different types of dirt identification. The model training and evaluation module is used to train the CNN-BiLSTM-Attention hybrid neural network prediction model and the improved YOLOv8n model using the training dataset as input data, and to set metrics to evaluate the performance of the two trained models. The detection application module is used to predict the dust accumulation density through the CNN-BiLSTM-Attention hybrid neural network prediction model, the mapping relationship between dust accumulation density and output power loss, and to determine whether the dust accumulation density exceeds the trigger threshold. If it exceeds the trigger threshold, the module uses the improved YOLOv8n model to determine the specific type of dirt on the photovoltaic panel.
9. An electronic device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, When the processor executes the program, it implements the photovoltaic panel dirt status identification method based on ash density and visual diagnosis as described in any one of claims 1 to 7.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the photovoltaic panel dirt status identification method based on ash density and visual diagnosis as described in any one of claims 1 to 7.