Roof-based photovoltaic region extraction method and device, terminal equipment and storage medium

By using an image prediction network in roof photovoltaic area extraction to extract debris features at different scales and quantify the degree of pixel contribution, the problem of inaccurate extraction of roof photovoltaic areas in the prior art is solved, and higher accuracy and accuracy are achieved.

CN118968142APending Publication Date: 2024-11-15GUANGDONG POWER GRID CO LTD
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202411005442.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-25
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

When extracting roof photovoltaic areas in the prior art, debris of different scales, shapes and materials cannot be accurately identified and eliminated, resulting in low extraction accuracy of photovoltaic development areas and large errors.

Method used

Using an image prediction network-based method, debris features at different scales are extracted through encoder and decoder, and the attention weight allocation module is used to quantify the contribution degree of each pixel position, and a fusion feature map is generated to identify and eliminate debris areas. .

Benefits of technology

It improves the identification and extraction accuracy of photovoltaic developmentable areas, reduces errors, and can accurately identify complex and diverse debris characteristics, creating more accurate target areas for building photovoltaics.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118968142A_ABST
    Figure CN118968142A_ABST
Patent Text Reader

Abstract

The invention discloses a photovoltaic region extraction method and device based on a roof, terminal equipment and a storage medium. An image prediction network can identify sundry features under different scales in a roof image; an introduced attention weight distribution module can also calculate the attention weight of each pixel position based on the fused feature map so as to quantify the contribution degree of the sundry features of each pixel position in the fused feature map, so that the accuracy of subsequent sundry recognition can be improved; and after the sundry region in the roof image is removed based on the accurately identified sundry region image, a more accurate photovoltaic construction region can be obtained. According to the method, foreign material feature representation under different scales can be learned, the contribution degree of each pixel position during foreign material recognition can be quantified, and recognition and extraction of complex and diversified foreign material features in the roof image are realized, so that the accuracy of photovoltaic developable effective region extraction is improved, and errors are reduced.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of rooftop photovoltaic area calculation, and in particular to a roof-based photovoltaic area extraction method, device, terminal equipment and storage medium. Background Art

[0002] Based on the current trend of high proportion of rooftop photovoltaic access to the grid, eliminating the undevelopable areas on the roof to obtain the photovoltaic developable areas in order to obtain the goal of accurate photovoltaic capacity has practical application needs and technical research value. Since most roofs have various debris and the roof structure shapes are changeable, it is often necessary to identify roofs of different structures and identify and eliminate various types of debris to obtain the effective area that can be used to build rooftop photovoltaics.

[0003] However, traditional methods or simple machine learning algorithms are often based on fixed feature extraction methods when extracting the photovoltaic construction area on the roof and removing roof debris, such as SIFT (Scale Invariant Feature Transformation Algorithm) and SURF (Speeded Up Robust Feature Extraction Algorithm), which are mainly used to extract key points and their descriptors in the image. These descriptors are usually based on local gradient information or wavelet transform response. However, debris on the roof, such as air conditioner outdoor units, ventilation ducts, wires, etc., may have different scales, that is, different shapes, sizes and materials. The existing technology cannot fully describe the characteristics of these debris, and thus cannot fully capture the complex and diverse debris features in the roof image, and cannot accurately identify and remove various types of debris, resulting in low accuracy in the extraction of effective areas that can be developed for photovoltaics, and there are large errors. Summary of the invention

[0004] The embodiments of the present invention provide a roof-based photovoltaic area extraction method, apparatus, terminal device and storage medium, which can not only learn the feature representation of debris at different scales, but also quantify the contribution of each pixel position in identifying debris, thereby achieving accurate identification and extraction of debris features, thereby improving the accuracy of identification and extraction of effective photovoltaic developable areas, and can effectively solve the problem in the prior art that the accuracy of extraction of effective photovoltaic developable areas is not high due to the inability to accurately identify and eliminate various types of debris.

[0005] An embodiment of the present invention provides a roof-based photovoltaic area extraction method, comprising:

[0006] Acquire a rooftop image containing a debris area;

[0007] Inputting the roof image into a preset image prediction network so that the image prediction network removes debris areas in the roof image, and then extracting target areas that can be used for photovoltaic construction from the roof image;

[0008] The image prediction network includes: an encoder, a decoder, and an attention weight allocation module located between the encoder and the decoder;

[0009] The encoder is used to extract debris features at different scales in the roof image and generate a plurality of first feature maps for representing the debris features;

[0010] The decoder is used to process each first feature map through an upsampling operation to generate a plurality of second feature maps for characterizing the features of the debris;

[0011] The attention weight allocation module is used to fuse each first feature map and each second feature map to generate a fused feature map; according to the feature value of each pixel position in the fused feature map, generate an attention weight for each pixel position; wherein the attention weight is used to characterize the contribution of the debris feature of a pixel position in the fused feature map;

[0012] The decoder is also used to fuse each second feature map based on the attention weight of each pixel position to generate a debris area image; according to the size and position of the debris area image, the debris area corresponding to the debris area image is removed from the roof image, and then the area in the roof image that has not been removed is used as a target area that can be used for photovoltaic construction.

[0013] Preferably, the encoder consists of an initial cyclic residual convolution layer, a plurality of cyclic residual convolution layers and a plurality of pooling layers;

[0014] The encoder is used to extract debris features at different scales in the roof image and generate a plurality of first feature maps for characterizing the debris features, including:

[0015] The initial cyclic residual convolution layer in the encoder is used to obtain the roof image, extract the debris features at different scales in the roof image, and generate the initial feature map of the corresponding level;

[0016] Each cyclic residual convolution layer in the encoder is used to generate an initial feature map of a corresponding level according to the debris features at different scales in the extracted roof image and the initial feature map output by the previous cyclic residual convolution layer;

[0017] Each pooling layer in the encoder is used to downsample the initial feature map generated by the corresponding cyclic residual convolution layer to generate a first feature map for characterizing the features of the debris.

[0018] Preferably, the decoder is composed of an initial convolutional layer and a plurality of convolutional layers stacked in sequence;

[0019] The decoder is used to process each first feature map through an upsampling operation to generate a plurality of second feature maps for characterizing the features of the debris, including:

[0020] The initial convolution layer in the decoder is used to obtain a first feature map corresponding to the output of the initial cyclic residual convolution layer, and gradually restore the spatial resolution of the first feature map through an upsampling operation to generate a second feature map for characterizing the characteristics of the debris;

[0021] Each convolution layer in the decoder is used to gradually restore the spatial resolution of the first feature map output by the corresponding cyclic residual convolution layer through an upsampling operation to generate a feature map that has been upsampled; then the second feature map output by the previous convolution layer and the feature map that has been upsampled are fused to generate a second feature map for characterizing the characteristics of the debris.

[0022] Preferably, the attention weight allocation module is used to generate the attention weight of each pixel position according to the feature value of each pixel position in the fused feature map, including:

[0023] The attention weight allocation module is used to extract the channel feature value of each pixel position in the fusion feature map;

[0024] The fused feature map is linearly transformed, and then the result value of the linear transformation is integrated with the channel feature value of each pixel position to generate the attention weight corresponding to each pixel position.

[0025] Preferably, the decoder further comprises: a feature pyramid module;

[0026] The attention weight allocation module is further used to perform feature pruning on each second feature map according to the attention weight corresponding to each pixel position;

[0027] The decoder is further used to fuse the second feature maps based on the attention weights of the pixel positions to generate a debris area image, including:

[0028] The feature pyramid module in the decoder is used to perform multi-scale feature fusion on each second feature map that has been feature pruned based on the attention weight of each pixel position through a lateral connection operation to generate a debris area image for representing the debris area contained in the roof image;

[0029] The lateral connection operation is used to represent an operation of adjusting the number of channels of a feature map through multiple convolutional layers.

[0030] Preferably, the attention weight allocation module is further used to perform feature pruning on each second feature map according to the attention weight corresponding to each pixel position, including:

[0031] The attention weight allocation module is further used to perform feature pruning on each second feature map according to the following formula:

[0032]

[0033] in, is the feature value corresponding to a pixel position in the second feature map after feature pruning, is the channel feature value corresponding to a pixel position, is the attention weight corresponding to a pixel position.

[0034] Preferably, the training process of the preset image prediction network includes:

[0035] The sample data of roof images with marked areas that can be used for photovoltaic construction are used as training samples;

[0036] Iteratively train the neural network model to be trained according to the training samples until the neural network model converges to generate a preset image prediction network;

[0037] In each training, the training samples are input into the neural network model so that the neural network model removes the debris areas in the training samples and then outputs the prediction results of the areas in the training samples that can be used for photovoltaic construction;

[0038] The prediction results of the areas that can be used for photovoltaic construction corresponding to the training samples are compared with the actual area annotation results of the training samples, and the network parameters of the neural network model are adjusted according to the comparison results.

[0039] Based on the above method embodiments, the present invention provides corresponding device embodiments.

[0040] An embodiment of the present invention provides a roof-based photovoltaic area extraction device, comprising: a roof image acquisition module and a photovoltaic area extraction module;

[0041] The roof image acquisition module is used to acquire a roof image containing a debris area;

[0042] The photovoltaic area extraction module is used to input the roof image into a preset image prediction network so that the image prediction network removes the debris area in the roof image and then extracts the target area that can be used for photovoltaic construction from the roof image;

[0043] The image prediction network includes: an encoder, a decoder, and an attention weight allocation module located between the encoder and the decoder;

[0044] The encoder is used to extract debris features at different scales in the roof image and generate a plurality of first feature maps for representing the debris features;

[0045] The decoder is used to process each first feature map through an upsampling operation to generate a plurality of second feature maps for characterizing the features of the debris;

[0046] The attention weight allocation module is used to fuse each first feature map and each second feature map to generate a fused feature map; according to the feature value of each pixel position in the fused feature map, generate an attention weight for each pixel position; wherein the attention weight is used to characterize the contribution of the debris feature of a pixel position in the fused feature map;

[0047] The decoder is also used to fuse each second feature map based on the attention weight of each pixel position to generate a debris area image; according to the size and position of the debris area image, the debris area corresponding to the debris area image is removed from the roof image, and then the area in the roof image that has not been removed is used as a target area that can be used for photovoltaic construction.

[0048] Based on the above method embodiments, the present invention provides corresponding terminal device embodiments.

[0049] Another embodiment of the present invention provides a terminal device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, and when the processor executes the computer program, a roof-based photovoltaic area extraction method described in the above-mentioned embodiment of the invention is implemented.

[0050] Based on the above method embodiments, the present invention provides corresponding storage medium item embodiments.

[0051] Another embodiment of the present invention provides a storage medium, wherein the computer-readable storage medium includes a stored computer program, wherein when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute a roof-based photovoltaic area extraction method described in the above-mentioned embodiment of the invention.

[0052] The following beneficial effects are achieved by implementing the present invention:

[0053] The embodiment of the present invention provides a method, apparatus, terminal device and storage medium for extracting photovoltaic areas based on roofs. When extracting photovoltaic construction areas on roofs, the present invention can identify and remove debris areas in roof images through a preset image prediction network, and then after removing the identified debris areas, generate a target area that can be used for photovoltaic construction. Specifically, the image prediction network of the present invention can identify debris features at different scales in roof images based on the encoder therein, and then can identify and extract debris of various sizes or shapes at different scales in roof images. After the corresponding generated first feature map is input into the decoder for upsampling operation, a second feature map containing richer debris detail information is further obtained. The attention weight allocation module introduced in the present invention can generate a fused feature map containing richer feature information by fusing the first feature map (generated by the encoder) at different scales and the upsampled second feature map (generated by the decoder), and can also calculate the attention weight of each pixel position in the map based on the fused feature map to obtain the contribution of the debris feature of each pixel position in the fused feature map, and can quantify the contribution of each pixel in identifying debris, so as to improve the accuracy of debris recognition, so that when the subsequent fusion is based on each attention weight and each second feature map, the complex and diverse debris features in the roof image can be accurately captured, and the debris area image can be generated. After the debris area in the roof image is eliminated based on the accurately identified debris area image, a more accurate target area that can be used for photovoltaic construction can be obtained. Compared with the prior art, the present invention can not only learn the debris feature representation at different scales, but also quantify the contribution of each pixel position in identifying debris (i.e., attention weight), realize the recognition and extraction of complex and diverse debris features in the roof image, thereby improving the accuracy of the extraction of effective areas that can be developed for photovoltaics and reducing errors. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 It is a flow chart of a roof-based photovoltaic area extraction method provided in one embodiment of the present invention.

[0055] Figure 2 It is a schematic diagram of the training process of the U-Net neural network provided by one embodiment of the present invention.

[0056] Figure 3 Schematic diagram of the network structure of a U-Net network provided by an embodiment of the present invention.

[0057] Figure 4 It is a structural diagram of a cyclic residual convolution module in a U-Net network provided in one embodiment of the present invention.

[0058] Figure 5It is a schematic diagram of the structure of an AG attention gate provided by an embodiment of the present invention.

[0059] Figure 6 It is a structural diagram of a feature pyramid module provided by an embodiment of the present invention.

[0060] Figure 7 It is a structural schematic diagram of a roof-based photovoltaic area extraction device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0061] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0062] like Figure 1 FIG. 1 is a flow chart of a method for extracting photovoltaic areas based on roofs according to an embodiment of the present invention. The method for extracting photovoltaic areas based on roofs comprises:

[0063] Step S1: Acquire a roof image containing debris areas;

[0064] Step S2: inputting the roof image into a preset image prediction network, so that the image prediction network removes the debris area in the roof image, and then extracts the target area that can be used for photovoltaic construction from the roof image;

[0065] The image prediction network includes: an encoder, a decoder, and an attention weight allocation module located between the encoder and the decoder;

[0066] The encoder is used to extract debris features at different scales in the roof image and generate a plurality of first feature maps for representing the debris features;

[0067] The decoder is used to process each first feature map through an upsampling operation to generate a plurality of second feature maps for characterizing the features of the debris;

[0068] The attention weight allocation module is used to fuse each first feature map and each second feature map to generate a fused feature map; according to the feature value of each pixel position in the fused feature map, generate an attention weight for each pixel position; wherein the attention weight is used to characterize the contribution of the debris feature of a pixel position in the fused feature map;

[0069] The decoder is also used to fuse each second feature map based on the attention weight of each pixel position to generate a debris area image; according to the size and position of the debris area image, the debris area corresponding to the debris area image is removed from the roof image, and then the area in the roof image that has not been removed is used as a target area that can be used for photovoltaic construction.

[0070] For step S1, in a preferred embodiment, the present invention can obtain a roof image of a single-family roof. Schematically, the roof image contains debris areas of various types of debris, and blank available areas that can be used to build photovoltaic power generation sources such as photovoltaic panels. Then, when the roof image is subsequently identified and processed, each debris can be identified, and after obtaining the debris area corresponding to each debris, the debris area can be removed and the blank available area can be obtained, so that the effective construction area of ​​photovoltaic can be determined.

[0071] For step S2, in a preferred embodiment, the training process of the preset image prediction network includes:

[0072] The sample data of roof images with marked areas that can be used for photovoltaic construction are used as training samples;

[0073] Iteratively train the neural network model to be trained according to the training samples until the neural network model converges to generate a preset image prediction network;

[0074] In each training, the training samples are input into the neural network model so that the neural network model removes the debris areas in the training samples and then outputs the prediction results of the areas in the training samples that can be used for photovoltaic construction;

[0075] The prediction results of the areas that can be used for photovoltaic construction corresponding to the training samples are compared with the actual area annotation results of the training samples, and the network parameters of the neural network model are adjusted according to the comparison results.

[0076] In a preferred embodiment, when obtaining the sample data of the roof image, the key features of the roof in the sample data can be analyzed to classify the single-family roofs, thereby obtaining a single-family roof classification sample library with clear roof foreign body information;

[0077] Therefore, through image filtering, grayscale transformation, mathematical morphology optimization, data enhancement and other operations, the roof images in the single-family roof classification sample library are preprocessed to construct training samples that can be used to train the neural network model.

[0078] Specifically, Figure 2 The training process diagram shown in the figure includes the following three parts in the process of constructing training samples:

[0079] (1) Analysis of key roof features:

[0080] The key features of the roof image in the sample data are extracted, and the key features include the two-dimensional shape of the roof, spectral information, and foreign matter information.

[0081] (2) Classification method of single-family roofs:

[0082] According to the spectral information (i.e. color information) of the roof buildings, the roofs can be divided into nine types: gray, white, blue, red, yellow, green, black, brown, and mixed colors; and according to the two-dimensional shape of the buildings, the roofs can be divided into five types: rectangular, circular, trapezoidal, letter-shaped, and combined shapes, among which the letter-shaped can be subdivided into seven types: H, L, T, U, Z, O, and E; thereby obtaining sample data of roofs of different colors and types.

[0083] (3) Establishment of roof sample library:

[0084] The pictures of single-family roofs with clear foreign object information are used as the original sample library to ensure that there are sufficient samples of various shapes and colors to cover various spectral information.

[0085] Furthermore, the image preprocessing process mainly includes the following steps:

[0086] (1) Use bilateral filtering to process the noise of each roof image, taking into account the spatial distance and image brightness information, so as to remove the noise while retaining the key information;

[0087] (2) Using mathematical morphology and wavelet transform methods to optimize the edge information of the roof can make the edge of the roof of the image with incomplete information clearer, so that the shape of the roof can be as complete as possible;

[0088] (3) The histogram grayscale transformation method is used to reduce the number of grayscale bits of the roof image and optimize the spectral characteristics. The number of grayscale bits is reduced as much as possible while being able to distinguish the foreground and background, and the interval from the minimum to the maximum value of the grayscale value of the roof image is mapped to the corresponding interval.

[0089] (4) Using different processing methods such as random horizontal flipping, random vertical flipping, random rotation, random brightness adjustment, random contrast adjustment, random noise addition and random cropping to process each roof sample image to form a final data set, so that training samples that can be used to train the neural network model can be obtained.

[0090] The embodiment of the present invention removes image noise through bilateral filtering and other methods, which can reduce interference information, improve the signal-to-noise ratio of the image, and make the model training more focused on key features. In addition, using different preprocessing methods to process image samples can greatly increase the diversity of sample data and improve the generalization ability of the model, so that the model can perform well under different conditions. By classifying roofs according to spectral information and two-dimensional shapes, a more sophisticated and targeted sample library can be constructed, making model training more efficient.

[0091] In a preferred embodiment, the image prediction network of the present invention is a U-Net neural network, and an improved U-Net neural network, that is, the present invention can add a cyclic residual convolution module, an AG attention gate and a feature pyramid module to the original network structure to obtain an improved U-Net neural network. Then, based on the improved U-Net neural network, the deep and shallow features can be comprehensively utilized to complete the roof area extraction and foreign object removal work to obtain a buildable area.

[0092] Specifically, based on the U-Net network, the convolution blocks in the encoding path and the decoding path are replaced by the Recurrent Residual Convolutional Units (RRCU), which can enhance information transfer and feature learning through the recursive mechanism, so as to enhance the accuracy of roof area recognition and effectively solve the overfitting problem under small sample output. Furthermore, the attention weight allocation module is introduced in the jump connection part of the U-Net network, that is, the attention gate (AG) is introduced as the attention mechanism to dynamically adjust the degree of attention to different inputs, and the attention coefficient is calculated for each pixel position, so that the subsequent feature pruning can be performed to achieve dynamic adjustment of weights, so that the model can pay attention to the roof area of ​​more concern and enhance the foreign body removal function.

[0093] The feature pyramid module is also introduced in the improved U-Net neural network. The specific working principle is as follows:

[0094] In the U-Net network decoder, the feature pyramid module can extract feature maps from various sizes (i.e., different scales), comprehensively utilize shallow and deep features, construct multi-scale image representation, and detect image features at different scales. Therefore, it can retrieve and match features of different sizes, thereby analyzing and identifying images at different scales, enhancing the ability to identify foreign objects of different scales, and effectively achieving edge detail retention and contour segmentation accuracy of buildings and foreign objects, so as to improve the robustness and accuracy of the algorithm.

[0095] Specifically, the encoder in the improved U-Net neural network consists of an initial cyclic residual convolution layer, multiple cyclic residual convolution layers, and multiple pooling layers; while the decoder consists of an initial convolution layer and several convolution layers stacked in sequence, and a feature pyramid module is also connected at the end of the decoder. Schematically, the convolution layer in the decoder is also a cyclic residual convolution layer.

[0096] The training process of the improved U-Net network can also be specifically as follows:

[0097] The sample data sets of each type of rooftop image are input into the improved U-Net neural network for training respectively, and different parameters are obtained. During training, the training set and the validation set are divided into a ratio of 9:1. The backbone network (Backbone Network) of the improved network selects the VGG-13 network, the batch size (Batch Size) is 1, the number of training times (epoch) is 20, and the other parameters are the default values. The activation function selects the ELU activation function that can effectively combine the ReLu activation function and the sigmoid activation function to avoid the gradient descent when the input is positive and the insensitive response when the input is negative. The input image has a unified resolution of 512×512. The experimental environment is: Intel Core i7-12700K CPU, 32GB RAM hardware environment, the operating system is Windows10, and the operating environment is the Facebook company's open source Pytorch deep learning framework based on Python3.8.

[0098] At the same time, an ablation experiment was designed. According to the same input and parameter settings, the image was input into the baseline U-Net model, and only R2U-Net with cyclic residual convolution module and AttU-Net with attention mechanism were used. After training, four models were obtained to test the accuracy indicators of each network.

[0099] In order to evaluate the extraction quality of multiple models, this embodiment uses the following three indicators for evaluation:

[0100] (1) Accuracy: Accuracy is one of the most commonly used evaluation indicators, which is used to measure the proportion of samples correctly classified by the model:

[0101]

[0102] The four indicators TP, TN, FP, and FN reflect the roof area correctly predicted as the roof, the background area correctly predicted as the background, the background area incorrectly predicted as the roof, and the roof area incorrectly predicted as the background, respectively.

[0103] (2) F1-Score: It is a weighted index that combines precision and recall. Precision is used to measure the accuracy of the model in predicting the certificate category, and recall is used to measure the proportion of correct predictions among all certificate category samples:

[0104]

[0105] in represent precision and recall respectively.

[0106] (3) Intersection over Union (IOU): used to measure the degree of overlap between the predicted image result and the true label:

[0107]

[0108] Where C is the segmentation result of the network, and G is the segmentation result of the label annotation.

[0109] Furthermore, after training the two cascade networks and ablation networks for roof area extraction and foreign object removal for 20 epochs, the loss functions of each network have stabilized and the model training has been basically completed. By comparing the roof area extraction networks, the average accuracy indicators for various types of roof extraction are shown in the table:

[0110] network Acc F1-Score IOU U-Net 0.823 0.750 0.813 R2U-Net 0.902 0.840 0.821 AttU-Net 0.923 0.883 0.833 Methods 0.930 0.889 0.841

[0111] Through the analysis of accuracy indicators, it can be found that the improved U-Net neural network in the embodiment of the present invention has certain advantages over the baseline network and the ablation network in terms of F1-Score, Acc, IOU and other indicators, which can prove the effectiveness of the method in this paper.

[0112] like Figure 3 From the network structure diagram of the U-Net network shown in the figure, it can be understood that there is a cyclic residual module (i.e., a cyclic residual convolution layer) at the end of each layer of the encoding path and the decoding path, and there is an attention weight allocation module based on the attention mechanism at the junction of the encoder and the decoder and at the junction of upsampling, and there is a feature pyramid module at the end of each layer of the decoding path.

[0113] In the network structure design of the above-mentioned U-Net network, the encoding path and the decoding path are closely connected through the cyclic residual convolution layer and the upsampling layer, and are interspersed with the attention weight allocation module and the feature pyramid module, which not only reduces the model parameters and calculation amount, but also improves the feature extraction and recognition efficiency of the model.

[0114] Through the synergistic effect of the above-mentioned multiple modules, the improved U-Net neural network of the present invention can accurately identify and eliminate debris areas in complex and changeable roof environments, and extract effective areas that can be used for photovoltaic construction. The U-Net neural network model of the present invention also has good generalization performance on different types of roof images, and is suitable for diversified needs in actual application scenarios.

[0115] Specifically, when the improved U-Net neural network is used as the image prediction network of the present invention, the image prediction network removes the debris area in the roof image, and then extracts the target area that can be used for photovoltaic construction from the roof image. The specific process includes:

[0116] First, the current roof image to be predicted is input into the encoder, and the encoder is used to extract the debris features at different scales in the roof image and generate a plurality of first feature maps for characterizing the debris features, specifically including:

[0117] The initial cyclic residual convolution layer in the encoder is used to obtain the roof image, extract the debris features at different scales in the roof image, and generate the initial feature map of the corresponding level;

[0118] Each cyclic residual convolution layer in the encoder is used to generate an initial feature map of a corresponding level according to the debris features at different scales in the extracted roof image and the initial feature map output by the previous cyclic residual convolution layer;

[0119] Each pooling layer in the encoder is used to downsample the initial feature map generated by the corresponding cyclic residual convolution layer to generate a first feature map for characterizing the features of the debris.

[0120] Specifically, Figure 4 The structural diagram of the circular residual convolution module shown in the figure, the RRCU of the circular residual convolution module (RRCU) includes an initial circular residual convolution layer and several circular residual convolution layers, which exist at the end of the encoding path and the decoding path; each convolution layer is composed of a 3X3 convolution kernel, which can be based on the element-by-element feature summation, while retaining the advantages of U-Net, integrating high-dimensional abstract features and low-dimensional detail feature information.

[0121] In the improved U-Net neural network, the encoder part is responsible for extracting the clutter features at different scales in the roof image. The current roof image to be predicted is taken as input. It first passes through the initial circular residual convolution layer (the starting layer of RRCU). The initial circular residual convolution layer uses a 3x3 convolution kernel to perform a convolution operation on the roof image to obtain preliminary clutter features and generate an initial feature map of the corresponding level. Each circular residual convolution layer in the encoder is further processed based on the output of the previous layer (i.e., the feature map output by the previous RRCU). Each circular residual convolution layer extracts features through multiple 3x3 convolution layers and uses residual connections to add the input and output, which can retain more feature information while reducing the problem of gradient disappearance or explosion. Therefore, the encoder can extract clutter features at different scales in the roof image layer by layer.

[0122] Furthermore, each pooling layer (such as the maximum pooling layer) in the encoder downsamples the initial feature map generated by the corresponding cyclic residual convolution layer. Since the downsampling operation can reduce the size of the feature map and increase the receptive field, higher-level features can be extracted, and the global context information of the image can be gradually constructed. Finally, after being processed by multiple cyclic residual convolution layers and pooling layers, the encoder finally generates several first feature maps for characterizing the clutter features, and these feature maps contain clutter information of different scales and levels in the roof image. The encoder of the present invention can extract clutter features at different scales in the roof image through multiple cyclic residual convolution layers. This is particularly important for the roof image segmentation task in complex clutter scenes, which can enable the neural network model to more accurately identify and segment clutter areas of different sizes in the future.

[0123] Schematically, the residual connection in the cyclic residual convolution module helps to retain more feature information and reduce information loss in the deep network. At the same time, by fusing high-dimensional abstract features and low-dimensional detail feature information, the encoder can generate a richer and more comprehensive feature representation. The improved U-Net neural network of the present invention has improved performance compared to the traditional U-Net network by introducing structures such as the cyclic residual convolution module, that is, it can more accurately identify the debris area in the roof image and extract the target area that can be used for photovoltaic construction.

[0124] Next, the decoder is used to process each first feature map through an upsampling operation to generate a plurality of second feature maps for characterizing the features of the debris, specifically including:

[0125] The initial convolution layer in the decoder is used to obtain a first feature map corresponding to the output of the initial cyclic residual convolution layer, and gradually restore the spatial resolution of the first feature map through an upsampling operation to generate a second feature map for characterizing the characteristics of the debris;

[0126] Each convolution layer in the decoder is used to gradually restore the spatial resolution of the first feature map output by the corresponding cyclic residual convolution layer through an upsampling operation to generate a feature map that has been upsampled; then the second feature map output by the previous convolution layer and the feature map that has been upsampled are fused to generate a second feature map for characterizing the characteristics of the debris.

[0127] Specifically, each convolution layer in the decoder corresponds to a cyclic residual convolution layer (or the layer before the pooling layer) in the encoder. For each layer in the decoder, the spatial resolution of the first feature map output by the corresponding cyclic residual convolution layer is first gradually restored through upsampling operations to generate an upsampled feature map. For non-initial layers, the second feature map output by the previous convolution layer can be fused with the upsampled feature map of the current layer, aiming to combine low-level detail information with high-level semantic information, thereby deriving a second feature map for characterizing various debris features. Schematically, the fused feature map is processed again through the convolution layer to generate a second feature map for characterizing debris features in the current layer.

[0128] Each layer in the decoder repeats the above process, layer by layer, gradually restoring the spatial resolution of the feature map and fusing multi-level feature information. Finally, the output layer of the decoder will generate multiple second feature maps with the same spatial resolution as the input roof image. The second feature map contains rich clutter feature information and can be used for subsequent clutter area removal and target area extraction.

[0129] Therefore, the decoder of this embodiment effectively combines the feature maps at different scales generated by the encoder through layer-by-layer upsampling and feature fusion, so that the feature map finally generated contains both rich detail information and advanced semantic information. Moreover, through the upsampling operation, the decoder can gradually restore the spatial resolution of the feature map, so that the feature map finally generated is consistent with the input image in space, which is convenient for subsequent pixel-level operations (such as debris area removal). Since the decoder fuses multi-level feature information, it can make it possible to more accurately identify the debris area in the roof image and generate a more refined debris feature map.

[0130] Furthermore, the attention weight allocation module is connected between the encoder and the decoder by jumping, and the attention weight allocation module is used to generate the attention weight of each pixel position according to the feature value of each pixel position in the fusion feature map, specifically including:

[0131] First, each first feature map and each second feature map are fused to generate a fused feature map;

[0132] Extract the channel feature value of each pixel position in the fusion feature map;

[0133] The fused feature map is linearly transformed, and then the result value of the linear transformation is integrated with the channel feature value of each pixel position to generate the attention weight corresponding to each pixel position.

[0134] Schematically, a linear transformation is performed on the fused feature map, which is usually implemented using a convolutional layer. The purpose of the linear transformation is to map the feature map to a new space in order to better calculate the attention coefficient.

[0135] Specifically, there is an attention gate in the attention weight allocation module, such as Figure 5 As shown in the structural diagram of the AG attention gate, the attention gate (AG) consists of a set of parameters, including the linear transformation W x , W g , ψ, and the bias term b ψ 、b g ; These parameters are used to calculate the attention coefficient (i.e. the attention weight mentioned above), where W x Used to perform linear transformation on the input feature map, W g It is used to perform linear transformation on the preset gate vector, ψ is the function used to calculate the final attention coefficient, and the bias term b ψ and b g Used to adjust the calculation range of the above attention coefficient.

[0136] The attention gate calculates the attention weight (i.e., the attention coefficient) based on the attention mechanism. ) can be divided into the following steps:

[0137] The first step is to use the sigmoid activation function to process the eigenvalues ​​of channel c at each pixel position i in the feature map, specifically:

[0138]

[0139] Among them, σ2 represents the calculation function for pruning the characteristic value of the channel at the pixel position based on the sigmoid activation function, x i,c is the feature value of channel c at each pixel position i on the input feature map. It can be understood that the sigmoid activation function is used to compress the calculation result between 0 and 1, indicating the importance of the pixel position.

[0140] In the second step, use the attention gate parameters to substitute the eigenvalue of channel c at each pixel position i, first solve the intermediate variable q, and then solve the attention coefficient

[0141] The calculation formula based on the additive attention mechanism is:

[0142]

[0143]

[0144] in, is the attention coefficient, is an intermediate variable that trims the eigenvalues ​​to intermediate values, is the feature value of channel c at each pixel position on the input feature map, g i is the eigenvalue of the spatial attention mechanism, θ att is the preset gate vector, and σ1 is the input after the two parts are added together. Schematically, the additive attention mechanism can evaluate the importance of each pixel position by calculating the similarity between the feature value and the gate vector (through dot product or weighted sum, etc.).

[0145] Furthermore, the attention weight allocation module is further used to perform feature pruning on each second feature map according to the attention weight corresponding to each pixel position, including:

[0146] The features of each second feature map are pruned according to the following formula:

[0147]

[0148] in, is the feature value corresponding to a pixel position in the second feature map after feature pruning, is the channel feature value corresponding to a pixel position, is the attention weight corresponding to a pixel position.

[0149] Specifically, the channel feature value of each pixel position is multiplied by its corresponding attention weight to adjust the feature representation of the pixel position, so that the task-related areas will be given higher weights, while irrelevant areas will be suppressed.

[0150] Therefore, the embodiment of the present invention uses the attention gate to enable the network model to dynamically adjust the weight of the feature map, so that in subsequent tasks (such as debris recognition and area removal), the model can pay more attention to important image areas. This improves the accuracy and robustness of the model, especially when dealing with complex and shape-changing debris features. The attention weight allocation module combines the feature maps of the encoder and decoder and uses the attention gate to calculate the attention weight, thereby achieving dynamic attention and feature enhancement of important areas in the image (such as debris areas).

[0151] Finally, the decoder is also used to fuse the second feature maps based on the attention weights of the pixel positions to generate a debris area image, specifically including:

[0152] The feature pyramid module in the decoder is used to perform multi-scale feature fusion on each second feature map that has been feature pruned based on the attention weight of each pixel position through a lateral connection operation to generate a debris area image for representing the debris area contained in the roof image; wherein the lateral connection operation is used to represent an operation of adjusting the number of channels of the feature map through multiple convolutional layers.

[0153] The feature pyramid module of the present invention can extract feature maps of different scales based on the generated attention weights and the second feature map that has been feature pruned, and upsample them to a uniform size to achieve subsequent feature fusion. Figure 6 As shown, it can extract two layers of feature maps at each scale respectively and merge them at the same scale, so as to capture details and contextual information at different levels, thereby more accurately locating and identifying debris areas in roof images.

[0154] Specifically, the lateral connection operation refers to adjusting the number of channels of feature maps of different scales through a series of convolutional layers so that they can maintain a consistent dimension during the fusion process, and each convolutional layer can be regarded as a process of feature conversion or feature mapping, which aims to extract and integrate useful information across scales.

[0155] Through the lateral connection operation, the feature pyramid module can effectively combine the low-resolution but semantically rich feature map with the high-resolution but detail-rich feature map to generate a multi-scale feature representation that contains both global information and local details. In the feature pyramid module, the second feature map that has been pruned (importance adjusted according to the attention weight) will be gradually fused in the order from low resolution to high resolution. The fusion of each layer will take into account the feature map of the previous layer (or layers), and adjust the number of channels through the lateral connection operation to ensure the smooth progress of the fusion process.

[0156] Finally, the feature pyramid module outputs a clutter area image that fuses multi-scale features. The image matches the original roof image in spatial resolution, but contains high-precision information about the clutter area at each pixel location.

[0157] Finally, the decoder can also remove the debris area corresponding to the debris area image from the roof image according to the size and position of the debris area image, and then use the unremoved area in the roof image as the target area that can be used to build photovoltaics. Schematically, in a preferred embodiment, according to the size and position of the mapped debris area, the decoder generates a mask image of the same size as the original roof image. In this mask image, the debris area is marked as a specific value (such as 0), and the non-debris area (i.e., the target area) is marked as another value (such as 1 or remains transparent), and the generated mask image is used to perform a mask operation on the original roof image so as to set the pixel value of the debris area to the background color (such as black) or completely transparent, thereby generating a new image in which the debris area has been effectively removed, and only the target area that can be used to build photovoltaics is retained. Further, the decoder can also perform post-processing steps such as image smoothing, edge detection or color correction to improve the quality and readability of the output image.

[0158] Through the above process, after removing the debris area, the remaining part is the target area that can be used for photovoltaic construction, and the target area should be continuous, unobstructed and suitable for installing photovoltaic equipment in the roof image.

[0159] Therefore, the present invention improves the original U-Net neural network, adds a cyclic residual network module, an AG attention gate and a feature pyramid module, thereby completing the roof area extraction and foreign body removal work based on the improved network, obtaining the buildable area, and outputting it as a reference for photovoltaic development potential prediction. The present invention can effectively avoid the overfitting problem under small samples, and has good recognition ability in roof area extraction and foreign body removal, and can more efficiently and accurately extract the rooftop buildable photovoltaic area, which can efficiently and accurately evaluate the urban photovoltaic development potential, improve energy utilization, and also improve the scientific nature of new distribution network planning.

[0160] like Figure 7 As shown, based on the above-mentioned various embodiments of the photovoltaic area extraction method based on the roof, the present invention provides a corresponding device embodiment;

[0161] An embodiment of the present invention provides a roof-based photovoltaic area extraction device, comprising: a roof image acquisition module and a photovoltaic area extraction module;

[0162] The roof image acquisition module is used to acquire a roof image containing a debris area;

[0163] The photovoltaic area extraction module is used to input the roof image into a preset image prediction network so that the image prediction network removes the debris area in the roof image, and then extracts the target area that can be used for photovoltaic construction from the roof image;

[0164] The image prediction network includes: an encoder, a decoder, and an attention weight allocation module located between the encoder and the decoder;

[0165] The encoder is used to extract debris features at different scales in the roof image and generate a plurality of first feature maps for representing the debris features;

[0166] The decoder is used to process each first feature map through an upsampling operation to generate a plurality of second feature maps for characterizing the features of the debris;

[0167] The attention weight allocation module is used to fuse each first feature map and each second feature map to generate a fused feature map; according to the feature value of each pixel position in the fused feature map, generate an attention weight for each pixel position; wherein the attention weight is used to characterize the contribution of the debris feature of a pixel position in the fused feature map;

[0168] The decoder is also used to fuse each second feature map based on the attention weight of each pixel position to generate a debris area image; according to the size and position of the debris area image, the debris area corresponding to the debris area image is removed from the roof image, and then the area in the roof image that has not been removed is used as a target area that can be used for photovoltaic construction.

[0169] It should be noted that the device embodiments described above are merely schematic, wherein the modules described as separate components may or may not be physically separated, and the components displayed as modules may or may not be physical modules, and may be located in one place or distributed on multiple network modules. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. In addition, in the accompanying drawings of the device embodiments provided by the present invention, the connection relationship between the modules indicates that there is a communication connection between them, which may be specifically implemented as one or more communication buses or signal lines. Those of ordinary skill in the art can understand and implement it without paying creative labor.

[0170] Those skilled in the art can clearly understand that, for the sake of convenience and simplicity, the specific working process of the device described above can refer to the corresponding process in the aforementioned method embodiment, and will not be repeated here.

[0171] Based on the above-mentioned various embodiments of the roof-based photovoltaic area extraction method, the present invention provides corresponding terminal equipment embodiments.

[0172] An embodiment of the present invention provides a terminal device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein when the processor executes the computer program, a roof-based photovoltaic area extraction method described in any method embodiment of the present invention is implemented.

[0173] The terminal device may be a computing terminal device such as a desktop computer, a notebook, a palm computer, a cloud server, etc. The terminal device may include, but is not limited to, a processor and a memory.

[0174] The processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc. The processor is the control center of the terminal device, and uses various interfaces and lines to connect various parts of the entire terminal device.

[0175] The memory can be used to store the computer program, and the processor realizes various functions of the terminal device by running or executing the computer program stored in the memory and calling the data stored in the memory. The memory can mainly include a program storage area and a data storage area, wherein the program storage area can store an operating system, at least one application required for a function, etc.; the data storage area can store data created according to the use of the mobile phone, etc. In addition, the memory can include a high-speed random access memory, and can also include a non-volatile memory, such as a hard disk, a memory, a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), at least one disk storage device, a flash memory device or other volatile solid-state storage device.

[0176] Based on the above-mentioned various embodiments of the roof-based photovoltaic area extraction method, the present invention provides corresponding storage medium item embodiments.

[0177] An embodiment of the present invention provides a storage medium, which includes a stored computer program, wherein when the computer program is running, the device where the computer-readable storage medium is located is controlled to execute a roof-based photovoltaic area extraction method described in any method embodiment of the present invention.

[0178] The storage medium is a computer-readable storage medium, and the computer program is stored in the computer-readable storage medium. When the computer program is executed by the processor, the steps of each method embodiment described above can be implemented. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form, etc. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, disk, optical disk, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electric carrier signal, telecommunication signal and software distribution medium, etc. It should be noted that the content contained in the computer-readable medium can be appropriately increased or decreased according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electric carrier signals and telecommunication signals.

[0179] The above is a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements and modifications without departing from the principle of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A photovoltaic area extraction method based on rooftops, characterized in that: include: Obtain a rooftop image containing debris areas; Inputting the roof image into a preset image prediction network so that the image prediction network removes debris areas in the roof image and then extracts target areas that can be used for photovoltaic construction from the roof image; The image prediction network includes an encoder, a decoder, and an attention weight allocation module located between the encoder and the decoder; The encoder is used to extract debris features at different scales in the roof image and generate a plurality of first feature maps for representing the debris features; The decoder is configured to process each first feature map through an upsampling operation to generate a plurality of second feature maps for representing the characteristics of the debris; The attention weight allocation module is used to fuse each first feature map and each second feature map to generate a fused feature map; generate an attention weight for each pixel position based on the feature value of each pixel position in the fused feature map; wherein the attention weight is used to represent the contribution of the clutter feature at a pixel position to the fused feature map; The decoder is further configured to fuse the second feature maps based on the attention weights of the pixel positions to generate a debris area image; remove the debris area corresponding to the debris area image from the roof image based on the size and position of the debris area image, and then use the unremoved area in the roof image as a target area that can be used for photovoltaic construction.

2. The method for extracting photovoltaic areas based on rooftops according to claim 1, wherein: The encoder consists of an initial cyclic residual convolution layer, multiple cyclic residual convolution layers and multiple pooling layers; The encoder is used to extract debris features at different scales in the roof image and generate a plurality of first feature maps for representing the debris features, including: The initial cyclic residual convolution layer in the encoder is used to obtain the roof image and extract the debris features at different scales in the roof image to generate the initial feature map of the corresponding level; Each cyclic residual convolution layer in the encoder is used to generate an initial feature map of the corresponding level based on the debris features at different scales in the extracted roof image and the initial feature map output by the previous cyclic residual convolution layer; Each pooling layer in the encoder is used to downsample the initial feature map generated by the corresponding cyclic residual convolution layer to generate a first feature map for characterizing the clutter feature.

3. The method for extracting photovoltaic areas based on rooftops according to claim 2, wherein: The decoder consists of an initial convolutional layer and several convolutional layers stacked in sequence; The decoder is configured to process each first feature map through an upsampling operation to generate a plurality of second feature maps for characterizing the features of the debris, including: The initial convolution layer in the decoder is used to obtain a first feature map corresponding to the output of the initial cyclic residual convolution layer, and gradually restore the spatial resolution of the first feature map through an upsampling operation to generate a second feature map for characterizing the clutter feature; Each convolution layer in the decoder is used to gradually restore the spatial resolution of the first feature map output by the corresponding cyclic residual convolution layer through an upsampling operation, thereby generating a feature map that has been upsampled; Then, the second feature map output by the previous convolutional layer and the feature map that has been upsampled are fused to generate a second feature map for characterizing the clutter features.

4. The method for extracting photovoltaic areas based on rooftops according to claim 3, wherein: The attention weight allocation module is used to generate the attention weight of each pixel position according to the feature value of each pixel position in the fusion feature map, including: The attention weight allocation module is used to extract the channel feature value of each pixel position in the fusion feature map; Perform a linear transformation on the fused feature map, and then integrate the result value of the linear transformation with the channel feature value of each pixel position to generate the attention weight corresponding to each pixel position.

5. The method for extracting photovoltaic areas based on rooftops according to claim 4, characterized in that: The decoder further includes: a feature pyramid module; The attention weight allocation module is further used to perform feature pruning on each second feature map according to the attention weight corresponding to each pixel position; The decoder is further configured to fuse the second feature maps based on the attention weights of the respective pixel positions to generate a debris area image, including: The feature pyramid module in the decoder is configured to perform multi-scale feature fusion on each of the feature-pruned second feature maps through a lateral connection operation based on the attention weight of each pixel position, thereby generating a debris area image for representing the debris area contained in the roof image; The lateral connection operation is used to represent an operation of adjusting the number of channels of a feature map through multiple convolutional layers.

6. The method for extracting photovoltaic areas based on rooftops according to claim 5, characterized in that: The attention weight allocation module is further configured to perform feature pruning on each second feature map according to the attention weight corresponding to each pixel position, including: The attention weight allocation module is further configured to perform feature pruning on each second feature map according to the following formula: in, is the feature value corresponding to a pixel position in the second feature map after feature pruning, is the channel feature value corresponding to a pixel position, is the attention weight corresponding to a pixel position.

7. The method for extracting photovoltaic areas based on rooftops according to claim 1, wherein: The training process of the preset image prediction network includes: The sample data of rooftop images with marked areas that can be used for photovoltaic construction are used as training samples; Iteratively train the neural network model to be trained according to the training samples until the neural network model converges to generate a preset image prediction network; During each training, the training samples are input into the neural network model so that the neural network model removes the debris areas in the training samples and then outputs the prediction results of the areas in the training samples that can be used for photovoltaic construction; The predicted results of the areas available for photovoltaic construction corresponding to the training samples are compared with the actual area labeling results of the training samples, and the network parameters of the neural network model are adjusted according to the comparison results.

8. A roof-based photovoltaic area extraction device, characterized in that: include: Roof image acquisition module and photovoltaic area extraction module; The roof image acquisition module is used to acquire a roof image containing debris areas; The photovoltaic area extraction module is used to input the roof image into a preset image prediction network so that the image prediction network removes debris areas in the roof image and then extracts target areas that can be used for photovoltaic construction from the roof image; The image prediction network includes an encoder, a decoder, and an attention weight allocation module located between the encoder and the decoder; The encoder is used to extract debris features at different scales in the roof image and generate a plurality of first feature maps for representing the debris features; The decoder is configured to process each first feature map through an upsampling operation to generate a plurality of second feature maps for representing the characteristics of the debris; The attention weight allocation module is used to fuse each first feature map and each second feature map to generate a fused feature map; generate an attention weight for each pixel position based on the feature value of each pixel position in the fused feature map; wherein the attention weight is used to represent the contribution of the clutter feature at a pixel position to the fused feature map; The decoder is further configured to fuse the second feature maps based on the attention weights of the pixel positions to generate a debris area image; remove the debris area corresponding to the debris area image from the roof image based on the size and position of the debris area image, and then use the unremoved area in the roof image as a target area that can be used for photovoltaic construction.

9. A terminal device, characterized in that: The invention comprises a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein when the processor executes the computer program, a roof-based photovoltaic area extraction method according to any one of claims 1 to 7 is implemented.

10. A storage medium, characterized in that: The storage medium includes a stored computer program, wherein when the computer program is running, the device where the storage medium is located is controlled to execute the roof-based photovoltaic area extraction method according to any one of claims 1 to 7.

Citation Information

Cited By

  • Image generation method and device of photovoltaic system, photovoltaic system, electronic equipment, storage medium and computer program product

    CN119963687A

  • Image generation method, device, photovoltaic system, electronic device, storage medium and computer program product for a photovoltaic system

    CN119963687B