A method for evaluating photovoltaic rooftop resources and carbon reduction potential based on satellite image recognition

Abstract

The present invention discloses a method for evaluating photovoltaic exploitable roof resources and carbon reduction potential based on satellite image recognition, including: downloading a high-resolution building roof satellite image dataset from a public map, obtaining the roof dataset information based on satellite images; cleaning the dataset according to whether there are roof features, dividing the images with roofs into satellite image datasets; extracting roof feature analysis based on deep learning image semantic segmentation methods, accurately identifying the building roof area and classifying it according to the roof structure and material; dividing the building into connected domains, calculating the area of ​​different types of roofs in combination with scales, and evaluating the photovoltaic potential in combination with the installed capacity per unit area rule; and evaluating the carbon reduction potential of the roof according to the photovoltaic potential evaluation results. The present invention integrates innovative solutions of remote sensing technology and machine learning methods, aiming to achieve rapid and accurate evaluation of the photovoltaic potential of large-scale urban building roofs.

Description

Technical Field

[0001] The present invention belongs to the field of photovoltaic power generation system planning, relates to photovoltaic roof resource prediction technology, and specifically relates to a method for evaluating photovoltaic developable roof resources and carbon reduction potential based on satellite image recognition. Background Art

[0002] With the growing global demand for renewable energy, the use of solar energy is of great significance for achieving sustainable development, mitigating climate change, and protecting the environment. Building rooftop photovoltaics have the advantages of being geographically unlimited and not occupying additional land resources, making them the first choice for developing and utilizing solar energy resources in cities. At present, the solar energy resource potential assessment method is mainly based on the potential assessment on a macro scale, which makes it difficult to provide scientific guidance for the construction and planning of building rooftop photovoltaics in cities. Therefore, the key to the development and utilization of photovoltaics in urban buildings is to establish a scientific and feasible method for evaluating the potential of distributed photovoltaics and carbon reduction potential on the roofs of urban buildings. Summary of the invention

[0003] Purpose of the invention: In order to overcome the shortcomings of the prior art, a method for evaluating the exploitable photovoltaic roof resources and carbon reduction potential based on satellite image recognition is provided, which integrates an innovative solution of remote sensing technology and machine learning methods, aiming to achieve rapid and accurate evaluation of the photovoltaic potential of large-scale urban building roofs.

[0004] Technical solution: To achieve the above purpose, the present invention provides a method for evaluating photovoltaic exploitable roof resources and carbon reduction potential based on satellite image recognition, comprising the following steps:

[0005] S1: Dataset acquisition:

[0006] Download high-resolution building roof satellite image datasets from public maps and obtain roof dataset information based on satellite images;

[0007] S2: Preliminary classification of data:

[0008] According to the presence or absence of roof features, the dataset is cleaned and images with roofs are divided to form a satellite image dataset;

[0009] S3: Roof Identification:

[0010] The image semantic segmentation method based on deep learning extracts roof feature analysis, accurately identifies the roof area of ​​the building and classifies it according to the roof structure and material;

[0011] S4: Calculation and evaluation of photovoltaic potential

[0012] Divide the connected domain of the building, calculate the area of ​​different types of roofs based on the scale, and evaluate the photovoltaic potential based on the installed capacity per unit area rule;

[0013] S5: Conduct rooftop carbon reduction potential assessment based on the photovoltaic potential assessment results.

[0014] Furthermore, the specific requirements for data collection in step S1 are as follows:

[0015] For the purpose of building roof recognition, satellite images of building roofs are collected. To ensure the accuracy of the recognition results, the following standards must be followed: the satellite images must be as clear as possible, and when the data set is annotated, the data set volume must account for no less than 10% of the original data, and the number of each category must be balanced; geographic location information must be included to facilitate roof location positioning; satellite images must be of consistent scale and size to facilitate stitching and overall visualization.

[0016] Furthermore, the method for cleaning the data set in step S2 is: classifying the acquired satellite images into images with buildings and images without buildings, and marking them with labels.

[0017] In step S2, image data is collected and labels are prepared. Labels generally refer to the category to which an image belongs. The purpose of the model of the present invention is to separate buildings with roofs from buildings without roofs, so the image classification model is divided into two categories, images with buildings and images without buildings. The acquired satellite images are classified, wherein images containing buildings are often rich in color and the image occupies a large memory size, so the images are pre-classified, and the process of the image occupying a memory size in the middle part is manually labeled.

[0018] Furthermore, in step S2, the satellite images are classified using the trained deep learning model Resnet50.

[0019] Furthermore, the recognition of the roof area of ​​the building using the constructed DeepLab v3+CBAM model in step S3 specifically includes the following steps:

[0020] A1: Feature Extraction

[0021] Use convolutional neural networks to extract features. Through multi-layer convolution and pooling operations, the features of the image are gradually abstracted. In the process of feature extraction, the convolution kernel is used to scan the entire image to obtain a feature map. The first few layers of the network are used to extract local details of the image. As the convolution layer goes deeper, the abstract and complex features in the image are extracted.

[0022] The structure of the convolutional neural network in the present invention includes:

[0023] Convolutional Layer

[0024] The convolution layer is one of the core layers in the convolutional neural network and one of the main components used to extract image features. The convolution layer generates an output feature map by sliding a set of convolution kernels on the input image and performing convolution operations on each position. In the convolutional neural network, the weights of the convolution kernels can automatically learn to extract the features of the input image. The core idea of ​​the convolution layer is local connection and weight sharing. Local connection means that each neuron in the convolution layer is only connected to a local part of the pixels on the input image, which can reduce the number of weights to be learned and reduce the complexity of the model. Weight sharing means that each convolution kernel in the convolution layer shares the same weight on the entire input image, which can further reduce the number of weights to be learned and make the convolution layer translationally invariant, that is, for different positions on the input image, the features extracted by the same convolution kernel are the same. The convolution layer usually includes multiple convolution kernels, each of which can extract different features, and the number of channels of the output feature map is equal to the number of convolution kernels. Convolution layers usually use activation functions to introduce nonlinear characteristics to enhance the expressive power of the model.

[0025] The pooling layer is a commonly used hierarchical structure in convolutional neural networks. It is mainly used to reduce the size and number of feature maps, reduce the number of parameters, and alleviate the overfitting of the model. It reduces the parameters that need to be learned in the network by downsampling the spatial size of the feature map, and can also improve the robustness of the model to position and size. The operation of the pooling layer is to take a statistical value, such as the maximum value, average value, etc., for each sub-region in the feature map, and use the statistical value as the output of the sub-region. Usually, the kernel function size and step size of the pooling layer are smaller than those of the convolution layer, thereby reducing the size of the feature map.

[0026] The main functions of the pooling layer are: reducing the size and number of feature maps, reducing model parameters, and alleviating model overfitting; increasing the robustness of the model to transformations such as translation and rotation; improving the model's feature extraction capabilities, which is conducive to detecting the location of the target in the image; reducing the amount of calculation and accelerating model training and prediction. Common pooling layers include the maximum pooling layer and the average pooling layer.

[0027] Spatial Pyramid Pooling

[0028] Spatial Pyramid Pooling (SPP) is used to solve the problem of different sizes of output feature maps after convolutional neural network processing for input images of different sizes. Traditional fully connected layers are designed based on fixed input dimensions, so manual adjustment is required when the input dimensions are different. SPP performs multiple pooling operations of different sizes on the input image to obtain multiple feature maps of different sizes, so that the neural network can process input images of any size and obtain outputs of the same dimension. SPP can be added to the last layer of a convolutional neural network, and its input is a fixed-size feature map. The SPP layer first performs multiple pooling operations of different sizes and strides on the input feature map. Each pooling operation obtains a fixed-size feature map, and the feature maps are concatenated in the channel dimension to obtain a fixed-size feature map with a variable number of channels as output.

[0029] Atrous convolution spatial pyramid pooling

[0030] Atrous convolution spatial pyramid pooling (ASPP) is a type of atrous convolution structure, which is used to improve the receptive field of the model and enhance the ability to extract features. ASPP combines multiple atrous convolutions with different sampling rates and global pooling to extract multi-scale information, thus adapting to the semantic segmentation task of objects of different sizes.

[0031] The ASPP module consists of a global pooling layer and multiple parallel dilated convolutions. Different dilated convolutions have different sampling rates, usually 6, 12, and 18. These sampling rates are achieved by dilated convolutions on the original convolution kernels, thereby increasing the receptive field of the convolution kernels. Finally, ASPP concatenates feature maps at different sampling rates to increase the receptive field and preserve the spatial resolution of the feature map. Finally, the output feature map of ASPP is adjusted to the same size as the input feature map through a convolution layer. ASPP can effectively improve the network's receptive field and acquisition of semantic information on large-scale objects, thereby improving the performance of semantic segmentation.

[0032] One of the core innovations of the present invention is that the designed convolutional attention mechanism module CBAM is combined with the ASPP module to obtain the ASPP module combined with the attention mechanism. The hole convolution may cause discontinuity of spatial information. The convolutional attention mechanism module CBAM is introduced into the juxtaposition structure for extracting features. The role of the attention mechanism is to focus on the noticed target pixels and enhance their weights. The hole convolution in ASPP can obtain contextual information of different scales. Adding an attention mechanism can make the features of related categories more aggregated, thereby effectively reducing the phenomenon of empty features. The convolutional attention mechanism module CBAM is used to learn contextual information and capture internal associations. The five branches of the juxtaposition structure obtain more contextual information by collecting features from different domains, and combining this information with the refined feature map to determine the dependencies between pixels and the differences between categories.

[0033] A2: Upsampling

[0034] Upsampling is used to expand a low-resolution image or part of an image to a high resolution. In deep learning, upsampling is often used to restore the size of downsampled feature maps so that these feature maps can match the size of the original input image. Upsampling methods include nearest neighbor interpolation, bilinear interpolation, bicubic interpolation, and deconvolution. Among them, bilinear interpolation can perform linear interpolation in both directions to create new pixels, thereby enlarging the image to the desired size. Deconvolution restores the size of the downsampled feature map through a reverse convolution operation. In semantic segmentation, upsampling can be used to restore low-resolution feature maps to the size of the original image and improve the accuracy of semantic segmentation by merging feature maps from other layers.

[0035] Bilinear interpolation

[0036] Bilinear interpolation is an image scaling technique, also known as bilinear interpolation, which is usually used to enlarge a low-resolution image to a high-resolution image while maintaining the smoothness and details of the image. It can be used to enlarge or reduce an image to any size while maintaining the smoothness and details of the image. This technique uses the weighted average of the grayscale values ​​of four adjacent pixels to estimate the new pixel value.

[0037] The formula is:

[0038] f(x,y)=(1-u)(1-v)f(x1,y1)+u(1-v)f(x2,y1)

[0039] +(1-u)vf(x1,y2)+uvf(x2,y2)

[0040] Among them, u and v are the horizontal and vertical distances of the target pixel relative to the four nearest neighbors. The upsampling operation of the image can be achieved using bilinear interpolation. In bilinear interpolation, interpolation is performed in two directions. In this case, the first step of interpolation is to perform linear interpolation in the x-axis and y-axis directions respectively. When calculating the scaling of the image, the image of the target size can be obtained by performing multiple bilinear interpolations on the original image. In each interpolation, the interpolation factors u and v will be reduced accordingly to maintain the smoothness of the image;

[0041] Deconvolution

[0042] Deconvolution is an operation in convolutional neural networks. The deconvolution operation can restore a low-dimensional feature map to a high-dimensional feature map, thereby achieving the upsampling function. Usually, in a neural network, a pooling layer is used to downsample the size of the feature map during the encoding process, and then a deconvolution layer is used to upsample the size of the feature map during the decoding process.

[0043] The process of deconvolution operation is similar to that of convolution operation, but the difference between convolution and deconvolution lies in the size of the convolution kernel and the object of operation. The size of the convolution kernel used in deconvolution is usually larger than that used in normal convolution, and the output feature map is padded to expand it to the size of the original image. The essence of deconvolution is to restore the information loss during convolution downsampling by reversing the process of convolution operation, thereby restoring the size and accuracy of the feature map.

[0044] Unlike traditional convolutional layers, the convolution kernel size of deconvolutional layers is usually larger because deconvolution requires inserting a large number of zero values ​​in the output feature map for upsampling. In convolution operations, each input pixel corresponds to only one output pixel, while in deconvolution operations, one input pixel usually corresponds to multiple output pixels.

[0045] Anti-pooling

[0046] In convolutional neural networks, pooling layers are used for downsampling to reduce the size of the image and reduce the amount of computation. However, in some applications, the image needs to be restored to its original resolution, in which case the unpooling layer is needed. The unpooling operation can be seen as the reverse operation of pooling, and its purpose is to expand the output in the pooling layer and restore it to the size of the original image. In the unpooling process, the pooled values ​​are assigned back to each pixel position in the original image using an interpolation method in order to obtain a high-resolution image.

[0047] Taking reverse maximum pooling as an example, the reverse pooling process is introduced. Before reverse maximum pooling, it is necessary to record the position of the maximum value of each sliding window. Then, according to the step size and scanning method of the pooling operation, the feature value before pooling is restored to its original position, thereby obtaining the feature map after reverse pooling.

[0048] A3: Pixel Classification

[0049] Use convolutional neural networks to process and classify images. The input of the network is an image, and the output is a segmentation map of the same size as the input image, where each pixel is classified into a certain category.

[0050] Semantic segmentation refers to classifying each pixel in an image to achieve semantic understanding of the entire image. Pixel classification is one of the core tasks of semantic segmentation. Its goal is to assign a semantic category to each pixel in the image and mark each pixel in the image as belonging to a certain category. In pixel classification, convolutional neural networks are usually used to process and classify images. The input of the network is an image, and the output is a segmentation map of the same size as the input image, in which each pixel is classified into a certain category. According to the classification results, it can be determined which category each pixel in the image belongs to. Subsequently, the pixels can be counted to determine the number of different pixel categories, and the area occupied by the category can be determined based on the number of pixels. Lay the foundation for the division of connected domains in the following text.

[0051] Furthermore, the convolutional attention mechanism module CBAM in step A1 includes a channel attention module CAM and a spatial attention module SAM, which obtain the channel attention weight and the spatial attention weight in turn.

[0052] The convolutional attention mechanism module CBAM operates as follows:

[0053] The channel attention module CAM and the spatial attention module SAM are multiplied by the original feature map A∈R C×H×W , adaptively adjust the features, increase the weight of the target features, and generate refined features A∈R C×H×W :

[0054]

[0055] where represents element-wise multiplication; if the two operands have different dimensions, the values ​​are replicated in such a way that the spatial attention values ​​are broadcast along the channel dimension and the channel attention values ​​are broadcast along the spatial dimension;

[0056] In the channel attention module CAM, the spatial information of the feature map is aggregated using average pooling and maximum pooling operations to generate two different spatial context descriptors: F cavg and F cmax ; Then forward both to the shared neural network to generate output features, and then merge them using element addition; Finally, send the merged sum to the sigmoid function σ to get the channel attention map F C ∈R C×1×1; In order to reduce parameter resources, the hidden layer size is set to R C×1×1 , where r is defined as the reduction ratio; the channel attention calculation formula is as follows:

[0057] F C =σ(W1(W0(F cavg ))+W1(W0(F cmax )))

[0058] Among them, W0∈R C / r×C and W1∈R C×C / r is the weight, shared by two inputs, and W0 is followed by the ReLU activation function;

[0059] In the spatial attention module SAM, the channel information is aggregated using average pooling and maximum pooling operations to generate two two-dimensional maps: F cavg ∈R 1×H×W and F smax ∈R 1×H×W , respectively represent the average pooling features and the maximum pooling features across channels, and they are connected and convolved through the convolution layer to generate a two-dimensional spatial attention map F S ∈R 1×H×W ; The spatial attention map is calculated as follows:

[0060] F S =σ(f 7×7 (F savg ; F smax ))

[0061] Among them, σ is the sigmoid function, f 7×7 Represents a convolution operation with a filter size of 7×7.

[0062] Furthermore, the calculation method of the area of ​​different types of roofs in step S4 is: factors such as roof area, light intensity, weather conditions, and inclination angle need to be considered. The roof area of ​​urban buildings is calculated based on the semantic segmentation results of urban buildings. The roof area is calculated by counting the number of pixels in the connected domain of the roof part and multiplying it by the pixel scale. The calculation formula is:

[0063] A i =r 2 ×N i

[0064] A i represents the area of ​​the i-th connected domain, where r is the pixel scale, N i is the number of pixels in the i-th connected domain. Due to the occlusion of urban buildings, the effective roof area of ​​urban buildings needs to be calculated, that is, the actual area where photovoltaic panels can be installed. The calculation formula is:

[0065] A′i =A i (1-B)

[0066] Where B is the shielding coefficient of urban buildings, A' i In order to consider the location factor and convert the roof area after shading, the shading coefficients of different types of buildings are determined according to the building attributes.

[0067] The analysis of construction conditions and influencing factors of photovoltaic power generation on the roof of buildings mainly analyzes the potential of photovoltaic power generation on the roof from two aspects. On the one hand, the analysis is based on the different roof structures and types, and on the other hand, the analysis is based on the different locations and environments of the roofs and the different shading coefficients.

[0068] The analysis steps considering different factors are as follows:

[0069] According to the characteristics of each roof, taking into account the different slopes of flat roofs and sloping roofs, considering roofs of different materials such as colored steel tiles and cement, considering the installation development rate of rooftop photovoltaics under different types such as industrial and commercial and residential buildings, and comprehensively considering the characteristics of each roof, an assessment of the rooftop photovoltaic conditions is carried out.

[0070] Based on the above considerations, the present invention provides a method for analyzing the location environment and shading coefficient conditions of the roof, wherein the influence of various objects on the degree of solar radiation and the effective building roof area after shading are analyzed as follows:

[0071] Climate condition analysis, according to the "solar energy resource assessment method" and existing research, shows that the number of days with sunshine hours greater than 6 hours in different months can reflect the utilization value of solar energy resources. If the sunshine hours in a day are less than 6 hours, it is considered that solar energy has no utilization value. The stability of solar energy is expressed by the ratio K of the maximum and minimum number of days with sunshine hours greater than 6 hours in each month of the year.

[0072] Analysis of wind speed conditions: Since photovoltaic panels are arranged outdoors, they usually have a certain inclination angle, resulting in a larger windward area. The impact of wind load on the battery component bracket needs to be considered.

[0073] The monthly variation of solar radiation is analyzed by finding the pattern based on the characteristics of solar radiation in each region. It generally shows periodic variation according to the seasons.

[0074] Analysis of monthly changes in sunshine hours shows that in addition to seasonal changes, sunshine hours in different regions also vary. For example, the sunshine hours in the southern coastal areas are generally longer, while the sunshine hours in inland areas are relatively shorter. At the same time, weather conditions will also affect the sunshine hours, and the sunshine hours in rainy weather will generally decrease.

[0075] The shielding coefficient refers to the degree of influence of shielding objects such as buildings, trees, and billboards on solar radiation in the environment where the roof is located. These shielding objects will block the direct and scattered solar radiation, so that the intensity of solar radiation reaching the ground is reduced, thereby affecting the collection efficiency of solar energy. Therefore, when evaluating solar energy resources, it is necessary to consider the impact of the urban shielding coefficient on solar radiation. The value of the urban shielding coefficient is usually a decimal between 0 and 1, where 0 means there is no shielding object and 1 means it is completely blocked. This coefficient can be measured or estimated according to actual conditions, including the influence of factors such as building height, density, orientation, tree height and distribution, number and location of billboards, etc. The urban shielding coefficient has an important influence on the collection efficiency of solar energy. Therefore, when selecting the location of a solar energy project, it is necessary to fully consider the influence of the urban shielding coefficient and select a location with fewer shielding objects to improve the collection efficiency of solar energy. The shielding coefficient is an indicator to measure the degree of influence of a building or an area on solar radiation, and its size is related to factors such as building height, orientation, density, etc. The present invention obtains the shielding coefficient by the following method:

[0076] Three-dimensional model of the building: The three-dimensional model of the building can be used in combination with GIS and other software to simulate the impact of solar radiation on the building and calculate the shading coefficient.

[0077] Field observation: Measuring stations can be set up in the area to be built or around existing buildings to conduct field observations of solar radiation, and then the shielding coefficient can be calculated based on the measured data.

[0078] Furthermore, the photovoltaic potential assessment in step S4 is specifically as follows:

[0079] B1: Division of connected domains of urban buildings. For the results obtained after semantic segmentation, pixels need to be grouped into different connected domains for further processing and analysis.

[0080] A connected domain refers to an area composed of pixels with the same pixel value, and the pixels in this area can be connected to each other through continuous straight lines, curves or other means. For a connected domain, it is usually necessary to further analyze and process it, such as calculating the area, perimeter, center point position, etc. of the connected domain. Therefore, dividing the connected domain is an important step after semantic segmentation. An image connected domain refers to a set of pixels composed of adjacent pixels. Image connected domain detection is an important issue in image processing, which can be used to separate different objects, calculate the number and size of objects, etc. In an image, a connected domain is usually used to separate objects. For a binary image, a connected domain can be composed of pixels with a pixel value of 1. For a grayscale image, a connected domain can be composed of pixels with the same or very close pixel values. In image processing, connected domain labels are usually used to represent different connected domains, and each connected domain is assigned a unique label.

[0081] B2: Obtaining the names of urban buildings

[0082] After semantic segmentation of satellite images of urban buildings, in order to accurately locate the subject to which the building belongs, it is necessary to match the building with the subject name, and then obtain the subject name information. The subject name information is obtained through point of interest (POI) data. Point of interest is the subject registration place on the map, such as stores, gas stations, hospitals, train stations, etc. POI data can enable research on spatiotemporal behavior, urban planning, geographic information, etc.

[0083] The present invention selects the map API to obtain POI data, and divides the POI into six categories: shopping malls, communities, factories, schools, governments, and hospitals. The contents obtained are coordinates and names;

[0084] B3: Calculate the photovoltaic power generation potential on the roofs of urban buildings

[0085] The power generation forecast of a photovoltaic power station should be calculated and determined based on the solar energy resources at the site, taking into account various factors such as the photovoltaic power station system design, photovoltaic array layout and environmental conditions.

[0086] The calculation formula for the annual average power generation Ep of a voltaic power station is as follows:

[0087]

[0088] Where: E p is the power generation (kWh); H A is the total annual solar radiation (kWh / m 2 );P Az is the system installation capacity (kW) (calculated after connected domain analysis); K1 is the photoelectric conversion efficiency coefficient; K2 is the system comprehensive efficiency coefficient; E s is the solar irradiance under standard test conditions, taking 1kW / m 2 .

[0089] Furthermore, the roof carbon reduction potential assessment in step S5 includes the following steps:

[0090] C1: Calculate the area on the roof of the building where photovoltaic panels can be installed;

[0091] Remote sensing technology, architectural drawings, etc. can be used to obtain the area of ​​the building's roof and calculate the area available for the installation of photovoltaic panels.

[0092] C2: Calculate the annual power generation of photovoltaic panels;

[0093] The annual power generation can be calculated using mathematical models based on local solar radiation data and parameters such as the power generation efficiency of photovoltaic panels.

[0094] C3: Calculate the carbon emissions of the photovoltaic power generation system over its entire life cycle;

[0095] The carbon emissions during the entire life cycle include the carbon emissions during manufacturing, transportation, installation, maintenance, etc. The life cycle assessment method can be used to evaluate each link of the photovoltaic power generation system and calculate the carbon emissions during the entire life cycle.

[0096] C4: Calculate the emission reduction of photovoltaic power generation system;

[0097] The emission reduction is equal to the difference between the annual power generation of photovoltaic panels multiplied by the carbon emission factor (carbon emissions per unit of power generation) and the annual carbon emissions of photovoltaics. The emission reduction of photovoltaic power generation systems can be calculated based on the carbon emission factors of countries or regions.

[0098] C5: Calculate the carbon reduction potential of photovoltaic power on rooftops of urban buildings;

[0099] The carbon reduction potential is equal to the annual emission reduction of building rooftop photovoltaics multiplied by the service life of the photovoltaic power generation system. The service life of the photovoltaic power generation system can be determined according to actual conditions, and the carbon reduction potential of urban building rooftop photovoltaics can be calculated.

[0100] C6: Calculate the carbon emissions of photovoltaic power generation throughout its life cycle;

[0101] Emission calculation refers to the total amount of carbon emissions from the entire process of photovoltaic system design, manufacturing, installation, use, maintenance, and removal. Usually, it includes energy consumption and carbon emissions in each link of photovoltaic module production, transportation, installation, operation, and removal;

[0102] The calculation of carbon emissions over the entire life cycle of a building rooftop photovoltaic system includes the following steps:

[0103] D1: Calculate the carbon emissions during the production of photovoltaic modules;

[0104] This process includes raw material mining, processing, manufacturing and transportation. The specific calculation method can refer to industry standards or relevant research reports.

[0105] D2: Calculate the carbon emissions during the transportation of PV panels;

[0106] This process includes the transportation of components from the manufacturing plant to the construction site, and the calculation method can refer to data such as transportation energy consumption and emission factors.

[0107] D3: Calculate the carbon emissions during the installation of photovoltaic modules;

[0108] This process involves the operation of manpower and mechanical equipment on the roof of the building, and the calculation method can refer to data such as energy consumption and emission factors during the installation process.

[0109] D4: Calculate the carbon emissions during the operation of photovoltaic modules;

[0110] This process includes the energy consumption of photovoltaic modules during operation and the emissions generated by operation. The calculation method can refer to data such as the power generation of the photovoltaic system and the efficiency of the photovoltaic modules.

[0111] D5: Calculate the carbon emissions during the dismantling of PV panels;

[0112] This process includes the dismantling and processing of photovoltaic modules. The calculation method can refer to data such as energy consumption and emission coefficient during the dismantling process.

[0113] C7: Calculate the carbon reduction capacity of rooftop photovoltaics:

[0114] The carbon reduction capacity of rooftop photovoltaics is calculated through the capacity, power generation, power generation efficiency and carbon factor of the solar power generation system; the total amount of electricity that can be generated each year is calculated using the capacity and power generation efficiency of the solar power generation system, and the carbon emissions that can be reduced by the photovoltaic system each year are calculated based on the difference between the carbon emissions corresponding to this power generation in the power grid and the carbon emissions of this power generation in the photovoltaic system; the carbon emission reduction of the photovoltaic system over its entire life cycle can be calculated based on the annual carbon emissions reduction of photovoltaics and the service life of the photovoltaic system;

[0115] The carbon factor refers to the carbon dioxide emissions per unit of electricity in the power grid. The size of the carbon factor depends on the production method and region of electricity. The size of the carbon factor is also affected by the source of electricity.

[0116] The annual carbon reduction potential and life cycle carbon reduction potential of rooftop photovoltaic systems are shown in the following formula:

[0117]

[0118]

[0119] In the formula, is the annual reduction in carbon emissions from the photovoltaic system, k is the electric carbon factor, It is the carbon emission reduction of the photovoltaic system throughout its life cycle.

[0120] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0121] 1. Data acquisition is convenient. Based on the open platform API, remote sensing satellite images are obtained for roof recognition and area calculation;

[0122] 2. The analysis is cost-effective, the data source is free and no other data is required to calculate the city's photovoltaic potential, and locate each building for photovoltaic potential calculation;

[0123] 3. Guaranteed accuracy: roof recognition is performed based on a highly accurate semantic segmentation model to distinguish different roof types, thus improving the accuracy of the assessment;

[0124] 4. It is more scalable. Unlike detailed building modeling, which is applicable to a single region, by expanding the data set, it is possible to conduct PV potential analysis in multiple provinces or even the whole country.

[0125] 5. The analysis time is shorter. Using the method of the present invention, it takes 8 hours to complete the training of the neural network, 3 seconds to complete the identification of a 1km2 urban block, and about 10 minutes to complete the entire city. The analysis of the city's photovoltaic potential can be completed quickly, and the evaluation efficiency has been greatly improved compared with the existing mainstream models. BRIEF DESCRIPTION OF THE DRAWINGS

[0126] Figure 1 is a flow chart of the method of the present invention;

[0127] Figure 2 It is a framework diagram of the method of the present invention;

[0128] Figure 3 It is a structural diagram of the CBAM module;

[0129] Figure 4 It is a schematic diagram of the combination of CBAM module and ASPP module;

[0130] Figure 5 This is a structural comparison diagram of the DeepLab v3 model and the DeepLab v3+CBAM model;

[0131] Figure 6 Coordinate diagram of the four nearest neighbor points in bilinear interpolation;

[0132] Figure 7 Schematic diagram of the deconvolution process. DETAILED DESCRIPTION

[0133] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent forms of modifications to the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0134] like Figure 1 and Figure 2 As shown, the present invention provides a method for evaluating photovoltaic developable roof resources and carbon reduction potential based on satellite image recognition, comprising the following steps:

[0135] S1: Dataset acquisition:

[0136] Download high-resolution building roof satellite image datasets from public maps and obtain roof dataset information based on satellite images;

[0137] S2: Preliminary classification of data:

[0138] According to the presence or absence of roof features, the dataset is cleaned and images with roofs are divided to form a satellite image dataset;

[0139] S3: Roof Identification:

[0140] The image semantic segmentation method based on deep learning extracts roof feature analysis, accurately identifies the roof area of ​​the building and classifies it according to the roof structure and material;

[0141] S4: Calculation and evaluation of photovoltaic potential

[0142] Divide the connected domain of the building, calculate the area of ​​different types of roofs based on the scale, and evaluate the photovoltaic potential based on the installed capacity per unit area rule;

[0143] S5: Conduct rooftop carbon reduction potential assessment based on the photovoltaic potential assessment results.

[0144] The specific requirements for data collection in step S1 are as follows:

[0145] For the purpose of building roof recognition, satellite images of building roofs are collected. To ensure the accuracy of the recognition results, the following standards must be followed: the satellite images must be as clear as possible, and when the data set is annotated, the data set volume must account for no less than 10% of the original data, and the number of each category must be balanced; geographic location information must be included to facilitate roof location positioning; satellite images must be of consistent scale and size to facilitate stitching and overall visualization.

[0146] The method for cleaning the data set in step S2 is: classifying the acquired satellite images into images with buildings and images without buildings, and marking them with labels.

[0147] In step S2, image data is collected and labels are prepared. Labels generally refer to the category to which an image belongs. The purpose of the model of the present invention is to separate buildings with roofs from buildings without roofs, so the image classification model is divided into two categories, images with buildings and images without buildings. The acquired satellite images are classified, wherein images containing buildings are often rich in color and the image occupies a large memory size, so the images are pre-classified, and the process of the image occupying a memory size in the middle part is manually labeled.

[0148] In step S2, the satellite images are classified using the trained deep learning model Resnet50.

[0149] In step S3, the present invention studies the roofs of urban buildings and uses deep learning and convolutional neural networks to build an end-to-end semantic segmentation model to identify the roofs of urban buildings. Figure 5 As shown, the DeepLab v3+ model is improved in the present invention, and an attention mechanism is added to the encoding-decoding structure to form the DeepLab v3+CBAM model, and its key steps can be divided into three stages: feature extraction, upsampling, and pixel classification. In the encoder, ResNet-101 is used as the backbone network to extract semantic information to obtain the underlying features, and then the ASPP module with attention mechanism is connected to obtain the encoder features.

[0150] The recognition of the building roof area using the constructed DeepLab v3+CBAM model specifically includes the following steps:

[0151] A1: Feature Extraction

[0152] Use convolutional neural networks to extract features. Through multi-layer convolution and pooling operations, the features of the image are gradually abstracted. In the process of feature extraction, the convolution kernel is used to scan the entire image to obtain a feature map. The first few layers of the network are used to extract local details of the image. As the convolution layer goes deeper, the abstract and complex features in the image are extracted.

[0153] The structure of the convolutional neural network in the present invention includes:

[0154] Convolutional Layer

[0155] The convolution layer is one of the core layers in the convolutional neural network and one of the main components used to extract image features. The convolution layer generates an output feature map by sliding a set of convolution kernels on the input image and performing convolution operations on each position. In the convolutional neural network, the weights of the convolution kernels can automatically learn to extract the features of the input image. The core idea of ​​the convolution layer is local connection and weight sharing. Local connection means that each neuron in the convolution layer is only connected to a local part of the pixels on the input image, which can reduce the number of weights to be learned and reduce the complexity of the model. Weight sharing means that each convolution kernel in the convolution layer shares the same weight on the entire input image, which can further reduce the number of weights to be learned and make the convolution layer translationally invariant, that is, for different positions on the input image, the features extracted by the same convolution kernel are the same. The convolution layer usually includes multiple convolution kernels, each of which can extract different features, and the number of channels of the output feature map is equal to the number of convolution kernels. Convolution layers usually use activation functions to introduce nonlinear characteristics to enhance the expressive power of the model.

[0156] The pooling layer is a commonly used hierarchical structure in convolutional neural networks. It is mainly used to reduce the size and number of feature maps, reduce the number of parameters, and alleviate the overfitting of the model. It reduces the parameters that need to be learned in the network by downsampling the spatial size of the feature map, and can also improve the robustness of the model to position and size. The operation of the pooling layer is to take a statistical value, such as the maximum value, average value, etc., for each sub-region in the feature map, and use the statistical value as the output of the sub-region. Usually, the kernel function size and step size of the pooling layer are smaller than those of the convolution layer, thereby reducing the size of the feature map.

[0157] The main functions of the pooling layer are: reducing the size and number of feature maps, reducing model parameters, and alleviating model overfitting; increasing the robustness of the model to transformations such as translation and rotation; improving the model's feature extraction capabilities, which is conducive to detecting the location of the target in the image; reducing the amount of calculation and accelerating model training and prediction. Common pooling layers include the maximum pooling layer and the average pooling layer.

[0158] Spatial Pyramid Pooling

[0159] Spatial Pyramid Pooling (SPP) is used to solve the problem of different sizes of output feature maps after convolutional neural network processing for input images of different sizes. Traditional fully connected layers are designed based on fixed input dimensions, so manual adjustment is required when the input dimensions are different. SPP performs multiple pooling operations of different sizes on the input image to obtain multiple feature maps of different sizes, so that the neural network can process input images of any size and obtain outputs of the same dimension. SPP can be added to the last layer of a convolutional neural network, and its input is a fixed-size feature map. The SPP layer first performs multiple pooling operations of different sizes and strides on the input feature map. Each pooling operation obtains a fixed-size feature map, and the feature maps are concatenated in the channel dimension to obtain a fixed-size feature map with a variable number of channels as output.

[0160] Atrous convolution spatial pyramid pooling

[0161] Atrous convolution spatial pyramid pooling (ASPP) is a type of atrous convolution structure, which is used to improve the receptive field of the model and enhance the ability to extract features. ASPP combines multiple atrous convolutions with different sampling rates and global pooling to extract multi-scale information, thus adapting to the semantic segmentation task of objects of different sizes.

[0162] The ASPP module consists of a global pooling layer and multiple parallel dilated convolutions. Different dilated convolutions have different sampling rates, usually 6, 12, and 18. These sampling rates are achieved by dilated convolutions on the original convolution kernels, thereby increasing the receptive field of the convolution kernels. Finally, ASPP concatenates feature maps at different sampling rates to increase the receptive field and preserve the spatial resolution of the feature map. Finally, the output feature map of ASPP is adjusted to the same size as the input feature map through a convolution layer. ASPP can effectively improve the network's receptive field and acquisition of semantic information on large-scale objects, thereby improving the performance of semantic segmentation.

[0163] like Figure 4 As shown, one of the core innovations of the present invention is that the designed convolutional attention mechanism module CBAM is combined with the ASPP module to obtain the ASPP module combined with the attention mechanism, and the hole convolution may cause discontinuity of spatial information. The convolutional attention mechanism module CBAM is introduced into the juxtaposition structure for extracting features. The role of the attention mechanism is to focus on the noticed target pixels and enhance their weights. The hole convolution in ASPP can obtain contextual information of different scales. Adding an attention mechanism can make the features of related categories more aggregated, thereby effectively reducing the phenomenon of empty features. The convolutional attention mechanism module CBAM is used to learn contextual information and capture internal associations. The five branches of the juxtaposition structure obtain more contextual information by collecting features from different domains, and combining this information with the refined feature maps to determine the dependencies between pixels and the differences between categories.

[0164] The convolutional attention mechanism module CBAM includes a channel attention module CAM and a spatial attention module SAM, which obtain the channel attention weight and the spatial attention weight in turn.

[0165] Reference Figure 3 As shown, the convolutional attention mechanism module CBAM operates as follows:

[0166] The channel attention module CAM and the spatial attention module SAM are multiplied by the original feature map A∈R C×H×W , adaptively adjust the features, increase the weight of the target features, and generate refined features A∈R C×H×W :

[0167]

[0168] where represents element-wise multiplication; if the two operands have different dimensions, the values ​​are replicated in such a way that the spatial attention values ​​are broadcast along the channel dimension and the channel attention values ​​are broadcast along the spatial dimension;

[0169] In the channel attention module CAM, the spatial information of the feature map is aggregated using average pooling and maximum pooling operations to generate two different spatial context descriptors: F cavg and F cmax ; Then forward both to the shared neural network to generate output features, and then merge them using element addition; Finally, send the merged sum to the sigmoid function σ to get the channel attention map F C ∈R C×1×1 ; In order to reduce parameter resources, the hidden layer size is set to R C×1×1 , where r is defined as the reduction ratio; the channel attention calculation formula is as follows:

[0170] F C =σ(W1(W0(F cavg ))+W1(W0(F cmax )))

[0171] Among them, W0∈R C / r×C and W1∈R C×C / r is the weight, shared by two inputs, and W0 is followed by the ReLU activation function;

[0172] In the spatial attention module SAM, the channel information is aggregated using average pooling and maximum pooling operations to generate two two-dimensional maps: F cavg ∈R 1×H×W and F smax ∈R 1×H×W , respectively represent the average pooling features and the maximum pooling features across channels, and they are connected and convolved through the convolution layer to generate a two-dimensional spatial attention map F S ∈R 1×H×W ; The spatial attention map is calculated as follows:

[0173] F S =σ(f 7×7 (F savg ; F smax ))

[0174] Among them, σ is the sigmoid function, f 7×7 Represents a convolution operation with a filter size of 7×7.

[0175] A2: Upsampling

[0176] Upsampling is used to expand a low-resolution image or part of an image to a high resolution. In deep learning, upsampling is often used to restore the size of downsampled feature maps so that these feature maps can match the size of the original input image. Upsampling methods include nearest neighbor interpolation, bilinear interpolation, bicubic interpolation, and deconvolution. Among them, bilinear interpolation can perform linear interpolation in both directions to create new pixels, thereby enlarging the image to the desired size. Deconvolution restores the size of the downsampled feature map through a reverse convolution operation. In semantic segmentation, upsampling can be used to restore low-resolution feature maps to the size of the original image and improve the accuracy of semantic segmentation by merging feature maps from other layers.

[0177] Bilinear interpolation

[0178] Bilinear interpolation is an image scaling technique, also known as bilinear interpolation, which is usually used to enlarge a low-resolution image to a high-resolution image while maintaining the smoothness and details of the image. It can be used to enlarge or reduce an image to any size while maintaining the smoothness and details of the image. This technique uses the weighted average of the grayscale values ​​of four adjacent pixels to estimate the new pixel value.

[0179] For the coordinates of the target pixel, the coordinates of its four nearest neighbor points are as follows: Figure 6 As shown, the formula is:

[0180] f(x,y)=(1-u)(1-v)f(x1,y1)+u(1-v)f(x2,y1)

[0181] +(1-u)vf(x1,y2)+uvf(x2,y2)

[0182] Among them, u and v are the horizontal and vertical distances of the target pixel relative to the four nearest neighbors. The upsampling operation of the image can be achieved using bilinear interpolation. In bilinear interpolation, interpolation is performed in two directions. In this case, the first step of interpolation is to perform linear interpolation in the x-axis and y-axis directions respectively. When calculating the scaling of the image, the image of the target size can be obtained by performing multiple bilinear interpolations on the original image. In each interpolation, the interpolation factors u and v will be reduced accordingly to maintain the smoothness of the image;

[0183] Deconvolution

[0184] Deconvolution is an operation in convolutional neural networks. The deconvolution operation can restore a low-dimensional feature map to a high-dimensional feature map, thereby achieving the upsampling function. Usually, in a neural network, a pooling layer is used to downsample the size of the feature map during the encoding process, and then a deconvolution layer is used to upsample the size of the feature map during the decoding process.

[0185] The process of deconvolution operation is similar to that of convolution operation, but the difference between convolution and deconvolution lies in the size of the convolution kernel and the object of operation. The size of the convolution kernel used in deconvolution is usually larger than that used in normal convolution, and the output feature map is padded to expand it to the size of the original image. The essence of deconvolution is to restore the information loss during convolution downsampling by reversing the process of convolution operation, thereby restoring the size and accuracy of the feature map.

[0186] like Figure 7 As shown in the figure, unlike the traditional convolution layer, the convolution kernel size of the deconvolution layer is usually larger, because deconvolution needs to insert a large number of zero values ​​in the output feature map for upsampling. In the convolution operation, each input pixel corresponds to only one output pixel, while in the deconvolution operation, one input pixel usually corresponds to multiple output pixels.

[0187] Anti-pooling

[0188] In convolutional neural networks, pooling layers are used for downsampling to reduce the size of the image and reduce the amount of computation. However, in some applications, the image needs to be restored to its original resolution, in which case the unpooling layer is needed. The unpooling operation can be seen as the reverse operation of pooling, and its purpose is to expand the output in the pooling layer and restore it to the size of the original image. In the unpooling process, the pooled values ​​are assigned back to each pixel position in the original image using an interpolation method in order to obtain a high-resolution image.

[0189] Taking reverse maximum pooling as an example, the reverse pooling process is introduced. Before reverse maximum pooling, it is necessary to record the position of the maximum value of each sliding window. Then, according to the step size and scanning method of the pooling operation, the feature value before pooling is restored to its original position, thereby obtaining the feature map after reverse pooling.

[0190] A3: Pixel Classification

[0191] Use convolutional neural networks to process and classify images. The input of the network is an image, and the output is a segmentation map of the same size as the input image, where each pixel is classified into a certain category.

[0192] Semantic segmentation refers to classifying each pixel in an image to achieve semantic understanding of the entire image. Pixel classification is one of the core tasks of semantic segmentation. Its goal is to assign a semantic category to each pixel in the image and mark each pixel in the image as belonging to a certain category. In pixel classification, convolutional neural networks are usually used to process and classify images. The input of the network is an image, and the output is a segmentation map of the same size as the input image, in which each pixel is classified into a certain category. According to the classification results, it can be determined which category each pixel in the image belongs to. Subsequently, the pixels can be counted to determine the number of different pixel categories, and the area occupied by the category can be determined based on the number of pixels. Lay the foundation for the division of connected domains in the following text.

[0193] The calculation method of different types of roof areas in step S4 is: factors such as roof area, light intensity, weather conditions, and inclination angle need to be considered. The roof area of ​​urban buildings is calculated based on the semantic segmentation results of urban buildings. The roof area is calculated by counting the number of pixels in the connected domain of the roof part and multiplying it by the pixel scale. The calculation formula is:

[0194] A i =r 2 ×N i

[0195] A i represents the area of ​​the i-th connected domain, where r is the pixel scale, N i is the number of pixels in the i-th connected domain. Due to the occlusion of urban buildings, the effective roof area of ​​urban buildings needs to be calculated, that is, the actual area where photovoltaic panels can be installed. The calculation formula is:

[0196] A′ i =A i (1-B)

[0197] Where B is the shielding coefficient of urban buildings, A' i In order to consider the location factor and convert the roof area after shading, the shading coefficients of different types of buildings are determined according to the building attributes.

[0198] The analysis of construction conditions and influencing factors of photovoltaic power generation on the roof of buildings mainly analyzes the potential of photovoltaic power generation on the roof from two aspects. On the one hand, the analysis is based on the different roof structures and types, and on the other hand, the analysis is based on the different locations and environments of the roofs and the different shading coefficients.

[0199] The analysis steps considering different factors are as follows:

[0200] According to the characteristics of each roof, taking into account the different slopes of flat roofs and sloping roofs, considering roofs of different materials such as colored steel tiles and cement, considering the installation development rate of rooftop photovoltaics under different types such as industrial and commercial and residential buildings, and comprehensively considering the characteristics of each roof, an assessment of the rooftop photovoltaic conditions is carried out.

[0201] Based on the above considerations, the present invention provides a method for analyzing the location environment and shading coefficient conditions of the roof, wherein the influence of various objects on the degree of solar radiation and the effective building roof area after shading are analyzed as follows:

[0202] Climate condition analysis, according to the "solar energy resource assessment method" and existing research, shows that the number of days with sunshine hours greater than 6 hours in different months can reflect the utilization value of solar energy resources. If the sunshine hours in a day are less than 6 hours, it is considered that solar energy has no utilization value. The stability of solar energy is expressed by the ratio K of the maximum and minimum number of days with sunshine hours greater than 6 hours in each month of the year.

[0203] Analysis of wind speed conditions: Since photovoltaic panels are arranged outdoors, they usually have a certain inclination angle, resulting in a larger windward area. The impact of wind load on the battery component bracket needs to be considered.

[0204] The monthly variation of solar radiation is analyzed by finding the pattern based on the characteristics of solar radiation in each region. It generally shows periodic variation according to the seasons.

[0205] Analysis of monthly changes in sunshine hours shows that in addition to seasonal changes, sunshine hours in different regions also vary. For example, the sunshine hours in the southern coastal areas are generally longer, while the sunshine hours in inland areas are relatively shorter. At the same time, weather conditions will also affect the sunshine hours, and the sunshine hours in rainy weather will generally decrease.

[0206] The shielding coefficient refers to the degree of influence of shielding objects such as buildings, trees, and billboards on solar radiation in the environment where the roof is located. These shielding objects will block the direct and scattered solar radiation, so that the intensity of solar radiation reaching the ground is reduced, thereby affecting the collection efficiency of solar energy. Therefore, when evaluating solar energy resources, it is necessary to consider the impact of the urban shielding coefficient on solar radiation. The value of the urban shielding coefficient is usually a decimal between 0 and 1, where 0 means there is no shielding object and 1 means it is completely blocked. This coefficient can be measured or estimated according to actual conditions, including the influence of factors such as building height, density, orientation, tree height and distribution, number and location of billboards, etc. The urban shielding coefficient has an important influence on the collection efficiency of solar energy. Therefore, when selecting the location of a solar energy project, it is necessary to fully consider the influence of the urban shielding coefficient and select a location with fewer shielding objects to improve the collection efficiency of solar energy. The shielding coefficient is an indicator to measure the degree of influence of a building or an area on solar radiation, and its size is related to factors such as building height, orientation, density, etc. The present invention obtains the shielding coefficient by the following method:

[0207] Three-dimensional model of the building: The three-dimensional model of the building can be used in combination with GIS and other software to simulate the impact of solar radiation on the building and calculate the shading coefficient.

[0208] Field observation: Measuring stations can be set up in the area to be built or around existing buildings to conduct field observations of solar radiation, and then the shielding coefficient can be calculated based on the measured data.

[0209] The photovoltaic potential assessment in step S4 is specifically as follows:

[0210] B1: Division of connected domains of urban buildings. For the results obtained after semantic segmentation, pixels need to be grouped into different connected domains for further processing and analysis.

[0211] A connected domain refers to an area composed of pixels with the same pixel value, and the pixels in this area can be connected to each other through continuous straight lines, curves or other means. For a connected domain, it is usually necessary to further analyze and process it, such as calculating the area, perimeter, center point position, etc. of the connected domain. Therefore, dividing the connected domain is an important step after semantic segmentation. An image connected domain refers to a set of pixels composed of adjacent pixels. Image connected domain detection is an important issue in image processing, which can be used to separate different objects, calculate the number and size of objects, etc. In an image, a connected domain is usually used to separate objects. For a binary image, a connected domain can be composed of pixels with a pixel value of 1. For a grayscale image, a connected domain can be composed of pixels with the same or very close pixel values. In image processing, connected domain labels are usually used to represent different connected domains, and each connected domain is assigned a unique label.

[0212] B2: Obtaining the names of urban buildings

[0213] After semantic segmentation of satellite images of urban buildings, in order to accurately locate the subject to which the building belongs, it is necessary to match the building with the subject name, and then obtain the subject name information. The subject name information is obtained through point of interest (POI) data. Point of interest is the subject registration place on the map, such as stores, gas stations, hospitals, train stations, etc. POI data can enable research on spatiotemporal behavior, urban planning, geographic information, etc.

[0214] The present invention selects the map API to obtain POI data, and divides the POI into six categories: shopping malls, communities, factories, schools, governments, and hospitals. The contents obtained are coordinates and names;

[0215] B3: Calculate the photovoltaic power generation potential on the roofs of urban buildings

[0216] The power generation forecast of a photovoltaic power station should be calculated and determined based on the solar energy resources at the site, taking into account various factors such as the photovoltaic power station system design, photovoltaic array layout and environmental conditions.

[0217] The calculation formula for the annual average power generation Ep of a voltaic power station is as follows:

[0218]

[0219] Where: E p is the power generation (kWh); H A is the total annual solar radiation (kWh / m 2 );P Az is the system installation capacity (kW) (calculated after connected domain analysis); K1 is the photoelectric conversion efficiency coefficient; K2 is the system comprehensive efficiency coefficient; E s is the solar irradiance under standard test conditions, taking 1kW / m 2 .

[0220] The average photoelectric conversion efficiency of the common polycrystalline silicon solar panels on the market is between 15% and 20%, while the high-end monocrystalline silicon solar panels can reach an efficiency of more than 22%. To simplify the calculation, the photoelectric conversion efficiency in this invention is 20%. The total power generation efficiency of the photovoltaic power station system is the product of the efficiencies of all system products. The power generation efficiency of general photovoltaic projects is about 70% to 80%.

[0221] The rooftop carbon reduction potential assessment in step S5 includes the following steps:

[0222] C1: Calculate the area on the roof of the building where photovoltaic panels can be installed;

[0223] Remote sensing technology, architectural drawings, etc. can be used to obtain the area of ​​the building's roof and calculate the area available for the installation of photovoltaic panels.

[0224] C2: Calculate the annual power generation of photovoltaic panels;

[0225] The annual power generation can be calculated using mathematical models based on local solar radiation data and parameters such as the power generation efficiency of photovoltaic panels.

[0226] C3: Calculate the carbon emissions of the photovoltaic power generation system over its entire life cycle;

[0227] The carbon emissions during the entire life cycle include the carbon emissions during manufacturing, transportation, installation, maintenance, etc. The life cycle assessment method can be used to evaluate each link of the photovoltaic power generation system and calculate the carbon emissions during the entire life cycle.

[0228] C4: Calculate the emission reduction of photovoltaic power generation system;

[0229] The emission reduction is equal to the difference between the annual power generation of photovoltaic panels multiplied by the carbon emission factor (carbon emissions per unit of power generation) and the annual carbon emissions of photovoltaics. The emission reduction of photovoltaic power generation systems can be calculated based on the carbon emission factors of countries or regions.

[0230] C5: Calculate the carbon reduction potential of photovoltaic power on rooftops of urban buildings;

[0231] The carbon reduction potential is equal to the annual emission reduction of building rooftop photovoltaics multiplied by the service life of the photovoltaic power generation system. The service life of the photovoltaic power generation system can be determined according to actual conditions, and the carbon reduction potential of urban building rooftop photovoltaics can be calculated.

[0232] C6: Calculate the carbon emissions of photovoltaic power generation throughout its life cycle;

[0233] Emission calculation refers to the total amount of carbon emissions from the entire process of photovoltaic system design, manufacturing, installation, use, maintenance, and removal. Usually, it includes energy consumption and carbon emissions in each link of photovoltaic module production, transportation, installation, operation, and removal;

[0234] The calculation of carbon emissions over the entire life cycle of a building rooftop photovoltaic system includes the following steps:

[0235] D1: Calculate the carbon emissions during the production of photovoltaic modules;

[0236] This process includes raw material mining, processing, manufacturing and transportation. The specific calculation method can refer to industry standards or relevant research reports.

[0237] D2: Calculate the carbon emissions during the transportation of PV panels;

[0238] This process includes the transportation of components from the manufacturing plant to the construction site, and the calculation method can refer to data such as transportation energy consumption and emission factors.

[0239] D3: Calculate the carbon emissions during the installation of photovoltaic modules;

[0240] This process involves the operation of manpower and mechanical equipment on the roof of the building, and the calculation method can refer to data such as energy consumption and emission factors during the installation process.

[0241] D4: Calculate the carbon emissions during the operation of photovoltaic modules;

[0242] This process includes the energy consumption of photovoltaic modules during operation and the emissions generated by operation. The calculation method can refer to data such as the power generation of the photovoltaic system and the efficiency of the photovoltaic modules.

[0243] D5: Calculate the carbon emissions during the dismantling of PV panels;

[0244] This process includes the dismantling and processing of photovoltaic modules. The calculation method can refer to data such as energy consumption and emission coefficient during the dismantling process.

[0245] The sum of the carbon emissions from all the above links is the carbon emissions of the entire life cycle of photovoltaic power generation on the roof of the building. In the production process of photovoltaic panels, the materials used are becoming more and more environmentally friendly, and the energy consumption of the production process is gradually decreasing. At the same time, with the continuous development and application of photovoltaic technology, the power output of photovoltaic power generation has increased year by year, and the carbon emissions of photovoltaic power generation in the entire life cycle have also been reduced. According to the statistical report of the International Energy Agency, the carbon emissions of photovoltaic power generation in the entire life cycle have dropped from about 130gCO2e / kWh in 2010 to about 35-45gCO2e / kWh in 2020. According to the data in the report, the median carbon emissions of photovoltaic power generation in the life cycle are about 48gCO2e / kWh, that is, for every 1kWh of photovoltaic power generation, an average of 48gCO2e of greenhouse gases will be emitted during the entire life cycle.

[0246] CO2e / kWh: The carbon dioxide equivalent emitted per kilowatt-hour of electricity generated. By calculating the annual power generation of the photovoltaic system, the annual carbon emissions of the photovoltaic system are obtained through the carbon emission equivalent per kWh. Combined with the life cycle of the photovoltaic system, the carbon emissions of the photovoltaic system over its entire life cycle are obtained. The annual carbon emissions of the photovoltaic system and the carbon emissions of the photovoltaic system over its life cycle are shown in the following formula.

[0247] C year =Q×C kWh

[0248] C LCA =K×C year

[0249] In the formula, C year is the annual carbon emissions of the photovoltaic system, Q is the annual power generation of the photovoltaic system, C kWh is the carbon dioxide equivalent emitted per kilowatt-hour of electricity, C LCA is the carbon emission of the photovoltaic system during its life cycle, and L is the photovoltaic system life cycle.

[0250] C7: Calculate the carbon reduction capacity of rooftop photovoltaics:

[0251] The carbon reduction capacity of rooftop photovoltaics is calculated through the capacity, power generation, power generation efficiency and carbon factor of the solar power generation system; the total amount of electricity that can be generated each year is calculated using the capacity and power generation efficiency of the solar power generation system, and the carbon emissions that can be reduced by the photovoltaic system each year are calculated based on the difference between the carbon emissions corresponding to this power generation in the power grid and the carbon emissions of this power generation in the photovoltaic system; the carbon emission reduction of the photovoltaic system over its entire life cycle can be calculated based on the annual carbon emissions reduction of photovoltaics and the service life of the photovoltaic system;

[0252] The carbon factor refers to the carbon dioxide emissions per unit of electricity in the power grid. The size of the carbon factor depends on the production method and region of electricity. The size of the carbon factor is also affected by the source of electricity.

[0253] The annual carbon reduction potential and life cycle carbon reduction potential of rooftop photovoltaic systems are shown in the following formula:

[0254]

[0255]

[0256] In the formula, is the annual reduction in carbon emissions from the photovoltaic system, k is the electric carbon factor, It is the carbon emission reduction of the photovoltaic system throughout its life cycle.

[0257] In order to evaluate the performance of the model and determine the optimization effect, a verification scheme was developed in this embodiment to compare with the current mainstream model, and the predicted data was used for testing. The accuracy of the model is evaluated by the ratio of the intersection and union between the predicted label and the true label, mIoU. The average mIoU value of the evaluation method proposed in the present invention is as high as 81.04%. The average mIoU values ​​of other mainstream models such as FCN-8S, SegNet, and Deeplab v2 are all lower than 80%. It can be seen that the evaluation method of the present invention has better evaluation performance.

Claims

1. A method for evaluating photovoltaic rooftop resources and carbon reduction potential based on satellite image recognition, characterized in that: The steps include: S1: Download high-resolution building roof satellite image datasets from public maps and obtain roof dataset information based on satellite images; S2: Clean the dataset according to whether there is a roof feature, and divide the images with roofs into satellite image datasets; S3: Extract roof feature analysis based on deep learning image semantic segmentation method, use DeepLab v3+CBAM model to identify the roof area of ​​the building and classify it according to the roof structure and material; The CBAM module is combined with the ASPP module of the convolutional neural network. The CBAM module includes the channel attention module CAM and the spatial attention module SAM; S4: Calculation and evaluation of photovoltaic potential Divide the connected domain of the building, calculate the area of ​​different types of roofs based on the scale, and evaluate the photovoltaic potential based on the installed capacity per unit area rule; Photovoltaic potential assessment includes: 1) Urban building connected domain division, grouping pixels into different connected domains; 2) Obtain the names of urban buildings. Select the map API to obtain POI data. The POI categories are divided into six categories: shopping malls, communities, factories, schools, governments, and hospitals. The obtained contents are coordinates and names. 3) Calculate the photovoltaic power generation potential of urban building roofs; S5: Conduct rooftop carbon reduction potential assessment based on the photovoltaic potential assessment results, including the following steps: C1: Calculate the area on the roof of the building where photovoltaic panels can be installed; C2: Calculate the annual power generation of photovoltaic panels; C3: Calculate the carbon emissions of the photovoltaic power generation system over its entire life cycle; C4: Calculate the emission reduction of photovoltaic power generation system; C5: Calculate the carbon reduction potential of photovoltaic power on rooftops of urban buildings; C6: Calculate the carbon emissions of photovoltaic power generation throughout its life cycle; C7: Calculate the carbon reduction capacity of rooftop photovoltaics.

2. The method for evaluating photovoltaic rooftop resources and carbon reduction potential based on satellite image recognition according to claim 1 is characterized in that: The method for cleaning the data set in step S2 is: classifying the acquired satellite images into images with buildings and images without buildings, and marking them with labels.

3. The method for evaluating photovoltaic exploitable roof resources and carbon reduction potential based on satellite image recognition according to claim 2 is characterized in that: In step S2, the satellite images are classified using the trained deep learning model Resnet50.

4. The method for evaluating photovoltaic rooftop resources and carbon reduction potential based on satellite image recognition according to claim 1 is characterized in that: In step S3, the recognition of the building roof area using the constructed DeepLab v3+CBAM model specifically includes the following steps: A1: Feature Extraction Use convolutional neural networks to extract features. In the process of feature extraction, the convolution kernel is used to scan the entire image to obtain a feature map. The first few layers of the network are used to extract local details of the image. As the convolution layer goes deeper, abstract and complex features in the image are extracted. The designed convolutional attention mechanism module CBAM is combined with the ASPP module of the convolutional neural network to obtain the ASPP module combined with the attention mechanism. The convolutional attention mechanism module CBAM is introduced into the juxtaposition structure for extracting features. The convolutional attention mechanism module CBAM learns context information and captures internal associations. The five branches of the juxtaposition structure obtain more context information by collecting features from different domains, and combine this information with the refined feature map to determine the dependencies between pixels and the differences between categories. A2: Upsampling Upsampling is used to restore the low-resolution feature map to the size of the original image and improve the accuracy of semantic segmentation by merging it with feature maps from other layers; A3: Pixel Classification Use convolutional neural networks to process and classify images; the input of the network is an image, and the output is a segmentation map of the same size as the input image, in which each pixel is classified into a certain category.

5. The method for evaluating photovoltaic rooftop resources and carbon reduction potential based on satellite image recognition according to claim 4 is characterized in that: The convolutional attention mechanism module CBAM in step A1 includes a channel attention module CAM and a spatial attention module SAM, which obtain the channel attention weight and the spatial attention weight in turn.

6. The method for evaluating photovoltaic rooftop resources and carbon reduction potential based on satellite image recognition according to claim 5 is characterized in that: The convolutional attention mechanism module CBAM operates as follows: The channel attention module CAM and the spatial attention module SAM are multiplied by the original feature map A∈R C×H×W , adaptively adjust the features, increase the weight of the target features, and generate refined features A∈R C×H×W : in represents element-wise multiplication; if the two operands have different dimensions, the values ​​are replicated in such a way that the spatial attention values ​​are broadcast along the channel dimension and the channel attention values ​​are broadcast along the spatial dimension; In the channel attention module CAM, the spatial information of the feature map is aggregated using average pooling and maximum pooling operations to generate two different spatial context descriptors: F cavg and F cmax ; Then forward both to the shared neural network to generate output features, and then merge them using element addition; Finally, send the merged sum to the sigmoid function σ to get the channel attention map F C ∈R C×1×1 ; In order to reduce parameter resources, the hidden layer size is set to R C×1×1 , where r is defined as the reduction ratio; the channel attention calculation formula is as follows: F C =σ(W1(W0(F cavg ))+W1(W0(F cmax ))) Among them, W0∈R C / r×C and W1∈R C×C / r is the weight, shared by two inputs, and W0 is followed by the ReLU activation function; In the spatial attention module SAM, the channel information is aggregated using average pooling and maximum pooling operations to generate two two-dimensional maps: F cavg ∈R 1×H×W and F smax ∈R 1×H×W , respectively represent the average pooling features and the maximum pooling features across channels, and they are connected and convolved through the convolution layer to generate a two-dimensional spatial attention map F S ∈ R 1×H×W ; The spatial attention map is calculated as follows: F S =σ(f 7×7 (F savg ;F smax )) Among them, σ is the sigmoid function, f 7×7 Represents a convolution operation with a filter size of 7×7.

7. The method for evaluating photovoltaic rooftop resources and carbon reduction potential based on satellite image recognition according to claim 1 is characterized in that: The calculation method of the roof areas of different types in step S4 is: the roof area of ​​urban buildings is calculated according to the semantic segmentation result of urban buildings; the roof area is calculated by counting the number of pixels in the connected domain of the roof part and multiplying the scale of the pixels, and the calculation formula is: A i =r 2 ×N i A i represents the area of ​​the i-th connected domain, where r is the pixel scale, N i is the number of pixels in the i-th connected domain; due to the occlusion of urban buildings, the effective roof area of ​​urban buildings needs to be calculated, that is, the actual area where photovoltaic panels can be installed. The calculation formula is: A′ i =A i (1-B) Where B is the shielding coefficient of urban buildings, A' i In order to take the location factor into account and calculate the roof area after shading, the shading coefficients of different types of buildings are determined according to the building attributes.

8. The method for evaluating photovoltaic rooftop resources and carbon reduction potential based on satellite image recognition according to claim 1 is characterized in that: The photovoltaic potential assessment in step S4 is specifically as follows: B1: Urban building connected domain division, grouping pixels into different connected domains; B2: Obtain the names of urban buildings. Select the map API to obtain POI data. The POI categories are divided into six categories: shopping malls, communities, factories, schools, governments, and hospitals. The obtained contents are coordinates and names. B3: Calculate the photovoltaic power generation potential on the roofs of urban buildings. The annual average power generation Ep of the photovoltaic power station is calculated as follows: Where: E p is the power generation; H A is the total annual solar radiation; P Az is the system installation capacity; K1 is the photoelectric conversion efficiency coefficient; K2 is the system comprehensive efficiency coefficient; E s is the solar irradiance under standard test conditions.

9. The method for evaluating photovoltaic rooftop resources and carbon reduction potential based on satellite image recognition according to claim 1 is characterized in that: The calculation of photovoltaic full life cycle carbon emissions in step C6 includes the following steps: D1: Calculate the carbon emissions during the production of photovoltaic modules; D2: Calculate the carbon emissions during the transportation of PV panels; D3: Calculate the carbon emissions during the installation of photovoltaic modules; D4: Calculate the carbon emissions during the operation of photovoltaic modules; D5: Calculate the carbon emissions during the dismantling of PV panels; The calculation of the rooftop photovoltaic carbon reduction capacity in step C7 includes: The carbon reduction capacity of rooftop photovoltaics is calculated through the capacity, power generation, power generation efficiency and carbon factor of the solar power generation system; the total amount of power generation each year is calculated using the capacity and power generation efficiency of the solar power generation system, and the carbon emissions reduced by the photovoltaic system each year are calculated based on the difference between the carbon emissions corresponding to this power generation in the power grid and the carbon emissions of this power generation in the photovoltaic system; the carbon emission reduction of the photovoltaic system over its entire life cycle can be calculated based on the annual carbon emissions reduction of photovoltaics and the service life of the photovoltaic system; The carbon factor refers to the carbon dioxide emissions per unit of electricity in the power grid. The size of the carbon factor depends on the production method and region of electricity. The size of the carbon factor is also affected by the source of electricity. The annual carbon reduction potential and life cycle carbon reduction potential of rooftop photovoltaic systems are shown in the following formula: In the formula, is the annual reduction in carbon emissions from the photovoltaic system, k is the electric carbon factor, It is the carbon emission reduction of the photovoltaic system throughout its life cycle.

Images