Centralized and distributed photovoltaic identification and prediction method

Abstract

The invention relates to a centralized and distributed photovoltaic identification prediction method, which comprises the following steps: acquiring a high-definition remote sensing image with the resolution of 0.6-0.7 m, and constructing to obtain an image data set; preprocessing the high-definition remote sensing image in the image data set to obtain a processed image and corresponding label data; training and optimizing a U-Net model by using each processed image and the corresponding label data to obtain a photovoltaic panel identification model; and inputting a to-be-detected remote sensing image into the photovoltaic panel identification model, outputting to obtain a photovoltaic panel identification result, and performing calculation processing on the photovoltaic panel identification result to obtain a photovoltaic panel prediction result. Compared with the prior art, the method has the advantages that a U-Net deep learning model is developed on the basis of existing codes, a high-definition remote sensing image is cut in ArcGIS Pro software on the basis of an open-source high-definition remote sensing data set, the U-Net model is developed through training, a photovoltaic panel recognition model is used for new remote sensing image data, and distributed photovoltaic and centralized photovoltaic recognition prediction can be efficiently and accurately performed.

Description

Technical Field

[0001] The invention relates to the technical field of remote sensing data processing, and in particular to a centralized and distributed photovoltaic identification and prediction method. Background Art

[0002] As one of the clean and pollution-free renewable energy sources, the efficient development and utilization of solar energy is an important measure to achieve low-carbon and environmental protection. Photovoltaic technology, that is, the technology of converting solar energy into electrical energy using solar cells, has become one of the key technologies for global energy transformation due to its potential in reducing greenhouse gas emissions and improving energy security. Photovoltaic technology mainly includes centralized and distributed forms. Centralized photovoltaic power stations are usually built in areas with sufficient sunshine, while distributed photovoltaics are closer to the user end, such as rooftop photovoltaic systems.

[0003] As a distributed power generation method, rooftop photovoltaic technology has become an important way to promote energy transformation and achieve renewable energy development goals due to its characteristics of being close to the user end and flexible deployment. Rooftop photovoltaics can not only improve energy efficiency, but also reduce dependence on traditional fossil energy and reduce greenhouse gas emissions, which is of great significance to addressing climate change.

[0004] With the rapid increase in photovoltaic installed capacity, how to accurately and efficiently plan and manage photovoltaic resources has become a key challenge facing the development of the industry. Traditional photovoltaic resource planning and management rely on manual statistics and ground measurement methods, which are not only time-consuming and labor-intensive, but also difficult to achieve large-scale, high-frequency monitoring. In addition, with the increasingly wide geographical distribution of photovoltaic projects, traditional ground measurement methods have become unsustainable in terms of time and economic costs. Therefore, the development of a technology that can quickly and accurately identify and segment photovoltaic solar panels from satellite and aerial images is of great significance for the planning and management of the photovoltaic industry.

[0005] In recent years, with the rapid development of remote sensing technology and artificial intelligence, it has become possible to automatically identify and segment photovoltaic solar panels using satellite and aerial images. This method can not only provide large-scale, high-resolution photovoltaic installed capacity data, but also monitor the changes in photovoltaic resources in real time, providing a scientific basis for the planning, construction, operation and maintenance of photovoltaic projects. In addition, through automated image processing and analysis, the efficiency and accuracy of photovoltaic resource assessment can be greatly improved, and the manpower and time costs can be reduced, thereby promoting the sustainable development of the photovoltaic industry.

[0006] As the importance of this technology becomes increasingly prominent, researchers around the world have begun to explore how to use advanced computer vision and machine learning techniques to improve the accuracy of automatic identification and segmentation of photovoltaic solar panels. In this field, many research teams have made significant progress. For example, Zhang Haoran's research team deeply analyzed the characteristics of photovoltaic panel semantic segmentation from the perspective of computer vision. The suggestions for data acquisition and model design for each observed feature provide valuable references for future improvements in the semantic segmentation of photovoltaic panels. The study found that in the semantic segmentation of photovoltaic panels, image data has problems such as class imbalance, color and texture complexity. Although solutions have been proposed, there is still room for improvement in the accuracy and adaptability of the model for large-scale distributed photovoltaic prediction. L. Kruitwagen's team used machine learning and remote sensing technology to conduct a comprehensive survey of non-residential photovoltaic solar installations around the world, creating a global public database of 68,661 spatially located facility footprints. This study is the first to use machine learning and remote sensing technology to search for specific types of infrastructure assets across the entire earth, providing a new perspective for the global monitoring and management of photovoltaic resources. A global survey of non-residential photovoltaic systems was conducted using remote sensing and machine learning techniques. Although a large database has been established, the research has focused on non-residential photovoltaic systems and has failed to fill the research gap in rooftop distributed photovoltaics. Abraham NoahWu's team developed the Roofpedia model, which adds a new dimension of consideration to assessing the overall sustainability of cities. Through this model, researchers can more accurately understand the current and actual status of roofs, including the coverage of solar panels and green plants. However, in actual operation, calculating the distribution of solar panels on roofs and the coexistence of green plants is a lot of work and complexity.

[0007] In summary, when faced with the planning and management of photovoltaic installed capacity, the existing technology still has the following shortcomings:

[0008] 1. Manual statistical methods are inefficient: Traditional photovoltaic resource planning relies on manual statistics and ground measurement methods, which are time-consuming and labor-intensive. It is difficult to monitor photovoltaic installed capacity on a large scale and at a high frequency, especially in large-scale applications around the world, and it is impossible to achieve fast and effective coverage.

[0009] 2. Distributed photovoltaic systems lack efficient monitoring tools: Existing research mainly focuses on centralized photovoltaic systems, while distributed photovoltaic systems, especially rooftop photovoltaic systems, lack efficient monitoring and management tools due to their wide distribution and large number.

[0010] 3. Poor adaptability of semantic segmentation technology: In existing research on semantic segmentation of photovoltaic panels, challenges such as category imbalance, color heterogeneity and texture homogeneity are faced, which makes it difficult for the model to maintain high accuracy in large-scale remote sensing images. In addition, some models have high requirements for image resolution, which increases the complexity and cost of data processing.

[0011] 4. Lack of globally applicable photovoltaic prediction models: Although some studies have achieved results in local areas, there is still a lack of universal models that can adapt to large-scale photovoltaic predictions on a global scale, especially the monitoring needs of distributed rooftop photovoltaic systems have not been effectively met. Summary of the invention

[0012] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide a centralized and distributed photovoltaic identification and prediction method, which can efficiently and accurately perform rooftop distributed photovoltaic and ground centralized photovoltaic identification and prediction.

[0013] The object of the present invention can be achieved by the following technical solution: A centralized and distributed photovoltaic identification and prediction method comprises the following steps:

[0014] S1. Obtain high-definition remote sensing images with a resolution of 0.6m-0.7m and construct an image data set;

[0015] S2, preprocessing the high-definition remote sensing images in the image data set to obtain processed images and corresponding label data;

[0016] S3, using the processed images and corresponding label data, training and optimizing the U-Net model to obtain a photovoltaic panel recognition model;

[0017] S4. Input the remote sensing image to be detected into the photovoltaic panel recognition model, output the photovoltaic panel recognition result, perform calculation processing on the photovoltaic panel recognition result, and obtain the photovoltaic panel prediction result.

[0018] Furthermore, the step S1 specifically utilizes satellite or aerial images to obtain high-definition remote sensing images with a resolution of 0.6m-0.7m.

[0019] Furthermore, the process of constructing the image data set in step S1 includes:

[0020] Locate the target area and save the corresponding image perspective;

[0021] Open source tools and third-party software are used to obtain high-definition images of the target area and stitch them into a complete image dataset, where the image dataset includes multiple images and corresponding geographic coordinate information.

[0022] Furthermore, step S2 specifically uses the deep learning tool in ArcGIS Pro software to draw labels, uses polygons to mark the photovoltaic panel area in the image, and rasterizes and exports the marked image to TIFF format, and then normalizes and crops the image and label data.

[0023] Furthermore, the training and optimization of the U-Net model in step S3 specifically adjusts the U-Net model parameters to adapt to the characteristics of the photovoltaic panel, and optimizes the model training process for the problems of category imbalance, color heterogeneity and texture homogeneity in the image data.

[0024] Furthermore, the step S4 comprises the following steps:

[0025] S41. Import the photovoltaic panel identification model into ArcGIS Pro software;

[0026] S42, applying the photovoltaic panel recognition model in ArcGIS Pro software, inputting the remote sensing image to be detected into the photovoltaic panel recognition model, and obtaining the photovoltaic panel recognition result;

[0027] S43, calculating the photovoltaic panel area according to the photovoltaic panel identification result;

[0028] S44. Calculate the installed capacity of photovoltaic panels based on the area of ​​photovoltaic panels.

[0029] Furthermore, the step S41 specifically converts the photovoltaic panel identification model into .emd or **.dlpk** format and then imports it into ArcGIS Pro software.

[0030] Furthermore, the step S42 is specifically to perform image reasoning in ArcGIS Pro software by using the Detect Objects Using Deep Learning tool to generate object detection results in a specific area, and to adjust model parameters and optimize model accuracy by using the Train Deep Learning Model tool;

[0031] Once processing is complete, ArcGIS Pro generates feature classes showing the locations of the PV panels detected by the model or the classification results.

[0032] Furthermore, the step S43 specifically calculates the photovoltaic panel area through the Calculate Geometry tool in ArcGIS software.

[0033] Furthermore, the formula for calculating the installed capacity of the photovoltaic panels in step S44 is: installed capacity = photovoltaic panel area × installed power per square meter.

[0034] Compared with the prior art, the present invention has the following advantages:

[0035] The present invention first obtains high-definition remote sensing images with a resolution of 0.6m-0.7m for constructing an image data set; then preprocesses the high-definition remote sensing images in the image data set to obtain processed images and corresponding label data; then uses each processed image and corresponding label data to train and optimize the U-Net model to obtain a photovoltaic panel recognition model; finally, the remote sensing image to be detected is input into the photovoltaic panel recognition model, and the photovoltaic panel recognition result is output, and the photovoltaic panel recognition result is calculated and processed to obtain the photovoltaic panel prediction result. In this way, combined with the high-definition remote sensing data set, the U-Net model can be trained and optimized to achieve a recognition accuracy rate of 91.6% under low resolution conditions, thereby improving the accuracy of recognition prediction for distributed photovoltaic and ground centralized photovoltaic, and the image resolution requirement is low, and photovoltaic prediction can be completed more efficiently, greatly reducing data costs and time consumption.

[0036] The present invention adopts the U-Net model to construct a photovoltaic panel recognition model. The U-Net model has excellent image segmentation capabilities and is suitable for processing complex remote sensing images. It can efficiently and accurately identify distributed and centralized photovoltaic systems from remote sensing images and perform large-scale photovoltaic predictions.

[0037] The present invention uses ArcGIS Pro software to preprocess open source remote sensing images, and imports and applies photovoltaic panel recognition models in ArcGIS Pro software, which can effectively reduce manpower and time costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a schematic diagram of the method flow of the present invention;

[0039] Figure 2 Schematic diagram of the U-Net model training process in the embodiment;

[0040] Figure 3 It is a schematic diagram of the application process of the embodiment;

[0041] Figure 4a is a remote sensing image input in the embodiment;

[0042] Figure 4b It is the photovoltaic panel identification result output in the embodiment. DETAILED DESCRIPTION

[0043] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0044] Example

[0045] like Figure 1As shown, a centralized and distributed photovoltaic identification and prediction method comprises the following steps:

[0046] S1. Obtain high-definition remote sensing images with a resolution of 0.6m-0.7m and construct an image data set;

[0047] S2, preprocessing the high-definition remote sensing images in the image data set to obtain processed images and corresponding label data;

[0048] S3, using the processed images and corresponding label data, training and optimizing the U-Net model to obtain a photovoltaic panel recognition model;

[0049] S4. Input the remote sensing image to be detected into the photovoltaic panel recognition model, output the photovoltaic panel recognition result, perform calculation processing on the photovoltaic panel recognition result, and obtain the photovoltaic panel prediction result.

[0050] This embodiment applies the above solution, and the main contents include:

[0051] 1. Acquisition of high-definition remote sensing images: Use satellite or aerial images to obtain high-definition remote sensing images with a resolution of 0.6m-0.7m.

[0052] This embodiment uses platforms such as Google Earth, World Imagery, and Mapbox to locate the target area and save the appropriate image perspective; then uses open source tools and third-party software to obtain high-definition images of a large area and stitch them into a complete image data set.

[0053] The image is saved in TIFF format. To ensure that the image can be used in GIS software, geographic coordinate information needs to be added to generate a GeoTiff format file.

[0054] 2. Data Preprocessing

[0055] The acquired high-definition remote sensing images were imported into ArcGIS Pro software, and labels were drawn using deep learning tools, and the photovoltaic panel areas were annotated using polygons. The processed images were rasterized and exported to TIFF format. The images and label data were further processed using Python scripts to make them suitable for training the U-Net model.

[0056] The main reason for choosing the U-Net model in this solution is its excellent image segmentation capabilities and suitability for processing complex remote sensing images. The U-Net model has the following key advantages, which make it very suitable for the photovoltaic panel identification and prediction tasks in this solution:

[0057] 1. High-precision segmentation capability: U-Net’s U-shaped architecture combines encoding and decoding paths, enabling it to simultaneously capture the global context and local detail information of an image. This is critical for accurately identifying photovoltaic panels from high-definition remote sensing images, especially when the photovoltaic panels in the image may be unevenly distributed or have complex color and texture features.

[0058] 2. Skip connections preserve spatial information: U-Net’s skip connections enable high-resolution features in the encoding path to be directly passed to the decoding path, avoiding the loss of detail information during upsampling. For tasks that require accurate positioning of the boundaries of photovoltaic panels, this feature greatly improves segmentation accuracy and ensures that the model can maintain stable performance on large-scale remote sensing data.

[0059] 3. Suitable for small data sets and complex tasks: U-Net is also designed with relatively few parameters, which means it can run efficiently with limited computing resources. Since the distribution characteristics of photovoltaic panels in different regions may vary, using U-Net can better adapt to these changes while maintaining a high prediction accuracy.

[0060] 4. Scalability and versatility: U-Net can not only handle the segmentation task of two-dimensional images, but also be extended to three-dimensional data to adapt to more types of remote sensing data and application scenarios. In addition, it also supports data enhancement technology, which helps to improve the robustness of the model when the training data is limited.

[0061] 3. U-Net model training: Use pre-processed images and label data to import the U-Net deep learning model for training. Adjust the model parameters to adapt to the characteristics of photovoltaic panels, especially to optimize the model training process for class imbalance, color heterogeneity, and texture homogeneity in image data. After the training process is completed, evaluate the accuracy and adaptability of the model through the validation data set to ensure that it can maintain a high accuracy rate in different regions and image quality.

[0062] The U-Net model training process in this embodiment is as follows Figure 2 As shown in the figure, it can be seen that the overall training results are good.

[0063] 4. Photovoltaic prediction: Apply the trained U-Net model to the remote sensing image data to be detected to predict distributed rooftop photovoltaics and centralized photovoltaics.

[0064] (I) Convert to ArcGIS supported formats: ArcGIS Pro mainly supports two deep learning model formats: .emd (Esri Model Definition file) and **.dlpk** (Deep Learning Package). Convert the model to one of these two formats. Create a custom inference function in Python or ArcGIS API to support the model.

[0065] (II) Import the model: After converting the model to a suitable format, use the following tools in ArcGIS Pro to import and apply the model:

[0066] Train Deep Learning Model: If you need to further adjust or fine-tune the model, use this tool to load preprocessed image data and train the model.

[0067] Detect Objects Using Deep Learning or Classify Pixels Using DeepLearning: These tools apply the trained model to new remote sensing data to perform object detection or pixel classification to generate distribution predictions for photovoltaic panels.

[0068] (III) Application model: Figure 3 As shown in the figure, the photovoltaic panel recognition model is imported and applied in ArcGIS Pro software; the remote sensing image to be detected is input into the photovoltaic panel recognition model to obtain the photovoltaic panel recognition result; then the photovoltaic panel area is calculated according to the photovoltaic panel recognition result; and then the photovoltaic panel installed capacity is calculated according to the photovoltaic panel area.

[0069] In practical applications, the Detect Objects Using Deep Learning tool can be used for image reasoning to generate object detection results in a specific area. The accuracy of the model can be optimized by adjusting model parameters (such as padding, batch size, etc.). After processing, ArcGIS Pro will generate a feature class that displays the location of the photovoltaic panels or the classification results detected by the model, which facilitates further spatial analysis and planning.

[0070] 5. Results display and application: This embodiment also designs that the prediction results will be displayed through a visualization platform. Users can view the photovoltaic distribution of specific areas on a global map and filter and analyze the data according to actual needs. Based on these data, relevant departments can optimize the planning, construction, operation and maintenance of photovoltaic resources.

[0071] In this embodiment, the input remote sensing image to be detected is as follows: Figure 4a As shown, the output photovoltaic panel identification results are as follows Figure 4b As shown, it can be seen that this solution is effective in identifying photovoltaic panels and can accurately identify rooftop photovoltaics in the area.

[0072] In summary, this solution can effectively solve the photovoltaic monitoring problems in existing technologies and achieve low-cost, high-efficiency large-scale photovoltaic system prediction and management by introducing high-definition remote sensing image processing, deep learning model training and automated prediction processes. The photovoltaic panel recognition model constructed by this solution performs better when processing large-scale data and has better versatility and stability. The model focuses on the prediction of distributed rooftop photovoltaic systems, filling the gap between the non-residential photovoltaic field and the rooftop photovoltaic field, and providing comprehensive monitoring and planning capabilities for global photovoltaic installed capacity. On the technical level, automated processing through deep learning models greatly reduces the workload of manual labeling and monitoring. The model has low requirements for image resolution and can more efficiently complete rooftop photovoltaic predictions around the world, greatly reducing data costs and time consumption. It has the following three characteristics:

[0073] 1. High precision and high accuracy: Compared with the existing photovoltaic panel recognition model, the U-Net model in this solution can still achieve a high accuracy of 91.6% while maintaining a resolution accuracy of 0.6m. Traditional models usually require higher resolutions to achieve the same level of accuracy, so this solution greatly reduces the cost and complexity of image acquisition while ensuring accurate prediction results.

[0074] 2. Large-scale prediction capability: The model of this solution has been tested in many cities around the world and can effectively predict the distribution of photovoltaic panels in different geographical regions (such as Europe, America, and Asia). The model is suitable for the monitoring and management of large-scale distributed photovoltaic systems, meeting the needs of model development for photovoltaic resource assessment on a global scale and monitoring of large-scale remote sensing areas.

[0075] 3. Low-cost solution: This solution uses open-source remote sensing image data for model training and prediction, reducing the high cost of purchasing commercial data. In addition, the use of deep learning algorithms for automated image processing and analysis significantly reduces manpower and time costs, greatly improving the efficiency of photovoltaic resource assessment.

Claims

1. A centralized and distributed photovoltaic identification and prediction method, characterized in that: The following steps are involved: S1. Obtain high-definition remote sensing images with a resolution of 0.6m-0.7m and construct an image dataset; S2, preprocessing the high-definition remote sensing images in the image data set to obtain processed images and corresponding label data; S3, using the processed images and corresponding label data, training and optimizing the U-Net model to obtain a photovoltaic panel recognition model; S4. Input the remote sensing image to be detected into the photovoltaic panel recognition model, output the photovoltaic panel recognition result, perform calculation processing on the photovoltaic panel recognition result, and obtain the photovoltaic panel prediction result.

2. A centralized and distributed photovoltaic identification and prediction method according to claim 1, characterized in that: The step S1 specifically uses satellite or aerial images to obtain high-definition remote sensing images with a resolution of 0.6m-0.7m.

3. A centralized and distributed photovoltaic identification and prediction method according to claim 1, characterized in that: The process of constructing the image data set in step S1 includes: Locate the target area and save the corresponding image perspective; Open source tools and third-party software are used to obtain high-definition images of the target area and stitch them into a complete image dataset, where the image dataset includes multiple images and corresponding geographic coordinate information.

4. A centralized and distributed photovoltaic identification and prediction method according to claim 1, characterized in that: The step S2 specifically uses the deep learning tool in ArcGIS Pro software to draw labels, uses polygons to mark the photovoltaic panel area in the image, and rasterizes and exports the marked image to TIFF format, and then normalizes and crops the image and label data.

5. A centralized and distributed photovoltaic identification and prediction method according to claim 1, characterized in that: The training and optimization of the U-Net model in step S3 specifically adjusts the U-Net model parameters to adapt to the characteristics of the photovoltaic panel, and optimizes the model training process for the problems of category imbalance, color heterogeneity and texture homogeneity in the image data.

6. A centralized and distributed photovoltaic identification and prediction method according to claim 4, characterized in that: The step S4 comprises the following steps: S41. Import the photovoltaic panel identification model into ArcGIS Pro software; S42, applying the photovoltaic panel recognition model in ArcGIS Pro software, inputting the remote sensing image to be detected into the photovoltaic panel recognition model, and obtaining the photovoltaic panel recognition result; S43, calculating the photovoltaic panel area according to the photovoltaic panel identification result; S44. Calculate the installed capacity of photovoltaic panels based on the area of ​​photovoltaic panels.

7. A centralized and distributed photovoltaic identification and prediction method according to claim 6, characterized in that: The step S41 specifically converts the photovoltaic panel identification model into .emd or **.dlpk** format and then imports it into ArcGIS Pro software.

8. A centralized and distributed photovoltaic identification and prediction method according to claim 6, characterized in that: The step S42 specifically includes performing image reasoning in ArcGIS Pro software by using the Detect Objects Using Deep Learning tool to generate object detection results in a specific area, and adjusting model parameters and optimizing model accuracy by using the Train Deep Learning Model tool; Once processing is complete, ArcGIS Pro generates feature classes showing the locations of the PV panels detected by the model or the classification results.

9. A centralized and distributed photovoltaic identification and prediction method according to claim 6, characterized in that: The step S43 specifically calculates the photovoltaic panel area through the Calculate Geometry tool in ArcGIS software.

10. A centralized and distributed photovoltaic identification and prediction method according to claim 6, characterized in that: The formula for calculating the installed capacity of the photovoltaic panels in step S44 is: installed capacity = photovoltaic panel area × installed power per square meter.

Images