A method for assessing the potential of urban building rooftop photovoltaic applications based on open-source satellite maps and deep learning
By combining high-resolution satellite maps and deep learning technology with instance segmentation models and multi-round semi-supervised clustering algorithms, the problem of data acquisition difficulties has been solved, enabling accurate assessment of the usable area of photovoltaic rooftops in cities. This method is applicable to the assessment of photovoltaic application potential in various types of cities.
Patent Information
- Application Number
- CN202311160215.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-09-11
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2043-09-11
AI Technical Summary
Existing methods for assessing the available area of urban rooftop photovoltaics based on open-source satellite imagery suffer from difficulties in data acquisition and insufficient three-dimensional information, leading to inaccurate assessment results, especially in small and medium-sized cities and underdeveloped areas.
Using high-resolution satellite maps combined with deep learning technology, this study extracts building roof features and evaluates roof area through an instance segmentation model and a multi-round semi-supervised clustering algorithm. The process includes data preprocessing, instance segmentation model training, roof depth feature extraction, and cluster optimization, ultimately calculating the proportion of rooftop photovoltaic usable area.
It enables accurate assessment of urban building rooftop photovoltaic installation capacity and production capacity under readily available data conditions. It is applicable to all types of cities, does not rely on indirect building information, and the results are targeted and persuasive.
Smart Images

Figure CN118097445B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a method for assessing the solar energy potential of urban building rooftops, specifically a method for assessing the potential of urban building rooftop photovoltaic applications based on open-source satellite maps and deep learning. Background Technology
[0002] To conserve energy and reduce carbon emissions, many countries have set carbon neutrality goals in recent years. As the most promising renewable energy source, large-scale utilization of solar energy is a crucial means to achieve carbon neutrality. Cities consume 75% of the world's energy and emit 70% of global greenhouse gases. Meanwhile, the numerous building rooftops in cities are ideal locations for installing photovoltaics. Therefore, many countries and cities have proposed the concept of "solar cities" and development strategies such as "photovoltaic roadmaps."
[0003] Assessing the potential of urban rooftop photovoltaics (PV) involves three main steps. First, the solar energy resources of the study area are estimated, and rooftop solar radiation is calculated using simulation tools. Then, a crucial step is to assess the available urban rooftop area for PV installation. Building rooftops are often not fully utilized for PV installations because they are often occupied by building elements such as chimneys, elevator shafts, or HVAC equipment. Finally, based on the available rooftop PV area, the installed PV capacity and power generation are calculated. For assessing the available rooftop PV area at the urban scale, the most common method is to first obtain the total rooftop area. Then, one or a set of rooftop PV available area coefficients (R²) are calculated. a Multiply the two parameters to obtain the usable area of rooftop photovoltaics.
[0004] There are generally three methods to obtain the total roof area of a city. First, some cities have conducted relatively detailed statistics on their building stock, such as total building area, number of buildings, and number of floors. Based on this data, the total roof area can be calculated. Second, some cities have built 3D city building models or collected digital surface model (DSM) data or LiDAR data for the entire city. Based on this data, building roofs can be easily extracted to calculate the city's roof area. The data sources required for these two methods are typically obtained in three ways: through cooperation with local governments, by purchasing from data vendors, or by downloading from open-source websites such as OpenStreetMap. However, the first two methods limit data availability, while the last method cannot guarantee data quality for some cities. Therefore, with the development of deep learning technology, many researchers choose to extract building roofs based on open-source satellite maps or aerial imagery (such as Google Earth satellite imagery) and then calculate the city's roof area.
[0005] Rooftop photovoltaic available area factor (R) aThere are three main methods for evaluating R. The first method is to set R based on the policies and standards of the city being studied. a The second method is to directly use R from cities similar to the city being studied. a The research results may be modified after considering the urban and architectural characteristics of the study area. Compared to Method 1, Method 2 is more targeted, but the selection criteria for similar cities (in principle, no two cities are exactly the same) need further exploration, and the selected similar cities may not have existing research results. Therefore, a more commonly used and targeted method is based on building classification and sampling calculation methods, including: ① classifying the buildings in the study city, ② sampling the roofs of each type of building and calculating R. a ③ R a It applies to the entire city. Buildings can be classified from multiple perspectives, including neighborhood density, population density, building type, and building shape.
[0006] However, current methods for estimating the available area of urban rooftop photovoltaics based on open-source satellite imagery still have some shortcomings. Firstly, current R-squared methods based on building classification and sampling calculations... a The assessment methods are difficult to apply. Currently used building classification standards are based on indirect parameters beyond image information, such as street density, population density, building type, and building shape. This building information is difficult to obtain in urban studies, significantly diminishing the advantage of readily available open-source satellite imagery. Secondly, when estimating the total roof area, the lack of three-dimensional information in satellite maps leads to an underestimation of pitched roof areas. Summary of the Invention
[0007] To address the current difficulty in data acquisition hindering the assessment of the potential for photovoltaic applications on urban building rooftops, this invention proposes a method for assessing the potential for photovoltaic applications on urban building rooftops based on open-source satellite maps and deep learning. This method can assess the potential photovoltaic installed capacity and production capacity of urban building rooftops based on easily accessible open-source data, providing a reference for relevant researchers and policymakers.
[0008] The objective of this invention can be achieved through the following technical solutions:
[0009] A method for assessing the potential of urban building rooftop photovoltaic applications based on open-source satellite maps and deep learning, the method comprising the following steps:
[0010] S1. High-resolution satellite map acquisition: Acquire satellite imagery of the city under study, with a spatial resolution requirement of less than 1m / pixel;
[0011] S2. Data preprocessing:
[0012] S21. Image cropping: Cropping satellite images into sizes suitable for input to deep learning instance segmentation models;
[0013] S22, Image Quality Enhancement: Corrects overly dark or overexposed images, while enhancing the contrast of all images and reducing noise;
[0014] S3. Construct the training dataset for the instance segmentation model;
[0015] S31. Image Sampling and Labeling: Extract a certain proportion of sample images from the image set obtained in step S2, label the image set, and establish the original dataset for model training.
[0016] S32. Data Augmentation: Data augmentation methods are used to expand the original labeled dataset. The augmented dataset is then used as the training dataset for the instance segmentation model.
[0017] S33, Dataset Splitting: Split the dataset obtained in S32 into a training set, a cross-validation set, and a test set;
[0018] S4. Establish an instance segmentation model: Train the instance segmentation model using the dataset obtained in S3;
[0019] S5. Roof Instance Extraction: Use the instance segmentation model trained in S4 for inference to extract the building roofs from all satellite images obtained in S3. Store each roof as an image and record the pixel area and coordinates of each roof.
[0020] S6. Roof Depth Feature Extraction: Use the backbone network of the instance segmentation model trained in S4 to extract the depth features of each roof image;
[0021] S7. Multi-round semi-supervised roof clustering process based on the constrained seed k-means algorithm: Based on the depth features of all roofs extracted in S6, multi-round semi-supervised clustering optimization is performed:
[0022] S71. Based on the characteristics of building roofs, propose similarity criteria, including: ① The two roofs are of the same type (flat roof / sloping roof). ② The two roofs have similar orientations. When two roofs simultaneously satisfy conditions ① and ②, they are considered strongly similar; when they only satisfy condition ①, they are weakly similar. Otherwise, the two roofs are not similar.
[0023] S72. Pre-clustering: Based on the depth features of all roofs, the number of clusters is set, and the unsupervised k-means clustering algorithm is used for pre-clustering;
[0024] S73. Seed Class Selection: Referring to the roof similarity criteria in S71, select the class with the most similar roofs from the pre-clustering results as the seed class. Determine and record the characteristics of the most common similar roofs in each seed class (flat / sloping roof, similar / dissimilar orientation) as the class features. Then, remove roofs that do not meet the class similarity characteristics (non-seed roofs) from each seed class, and label the remaining roofs (seed roofs) as that class. Simultaneously record the number of roofs in each seed class after clustering (N). i ) and the number of filtered seed roofs (N) si ), calculate N si and N i The ratio P;
[0025] S74. Semi-supervised clustering: Semi-supervised clustering is performed based on seed class using the constrained seed k-means algorithm;
[0026] S75. Multi-round semi-supervised clustering and seed class optimization: The method in S73 is used to select seed classes from clusters outside the seed class (non-seed classes), and then seed roofs are filtered within the seed classes. When no new seed classes appear, and the p-value of all seed classes is greater than 0.5, the clustering results are output for subsequent analysis. Otherwise, S73 and S74 are repeated. When p ≥ 0.5, it indicates that the proportion of seed vertices within a seed class is greater than 50%, suggesting that the seed vertices within the seed class have strong similarity characteristics.
[0027] S8. Total Roof Area Assessment and Sloping Roof Area Correction: Based on the pixel area of each roof obtained in S5, the actual roof area is calculated; simultaneously, based on the clustering results in S7, the classes with the characteristic of sloping roofs are corrected. The roof area calculation method is as follows:
[0028]
[0029] In the formula, S pixel and S roof,mask S1 represents the area represented by a single pixel in the high-resolution satellite image, and S8 represents the roof pixel area extracted from the satellite image. θ is the pitch angle of the sloping roof.
[0030] S9. Calculate the usable area ratio of rooftop photovoltaics (R) a ): Sample each type of roof in the clustering results and calculate the proportion of rooftop photovoltaic usable area:
[0031] S91. Calculation method: First, for all roofs, the area occupied by building elements is excluded; second, for pitched roofs, the parts facing east-south-west are the usable parts.
[0032] S92. Sampling Method: For non-seed classes and seed classes with P ≥ 0.9, first use systematic sampling to extract 12 roofs, and then use the method in S91 to calculate R for these roofs. a Remove R a Find the maximum and minimum values of the roofs, and calculate the average R of all remaining roofs. a For seed classes with p < 0.9, the systematic sampling method is first used to extract... One seed roof (Group A) and For the non-seed roofs (Group B), calculate the R of these roofs using the method in S91. a Remove R from each group a Find the maximum and minimum values of the roofs, and calculate the average R value for all remaining roofs. a ;
[0033] S10. Calculation of photovoltaic installation area and production capacity:
[0034] S101. The formulas for calculating the rooftop photovoltaic installation area and installed capacity are as follows:
[0035] S install =N cell ×S cell (2)
[0036] P install =N cell ×P cell (3)
[0037]
[0038] D m =W cell ×cosβ+L m (5)
[0039] S ava =S roof ×R a (6)
[0040] In the formula, N cell S cell and P cell These represent the number of photovoltaic modules, the area of a single module, and its peak power, respectively. a and D m The spacing between the front and rear rows of the photovoltaic array is to prevent them from shading each other, and also to allow for maintenance. L cell and W cell For the length and width of a single photovoltaic module. β and L m For the tilt angle of the PV modules and the spacing of the pedestrian walkways used for maintenance. S ava This refers to the area of the roof that can be used for photovoltaic installations;
[0041] S102, The formula for calculating photovoltaic capacity is:
[0042] E = S install ×R annual ×eff cell ×eff sys (7)
[0043] In the formula, R annual It studies the intensity of solar radiation on a horizontal plane in the area where the city is located. cell and eff sys These refer to photovoltaic module efficiency and photovoltaic system efficiency, respectively.
[0044] Compared with the prior art, the advantages of the present invention are:
[0045] 1. The data required for the implementation of this invention is more readily available. High-resolution satellite maps are now available through various channels. Furthermore, this invention does not rely on additional urban building information, such as street density, population density, building type, and building shape, required in previous methods when assessing the proportion of usable rooftop photovoltaic area.
[0046] 2. The method of this invention for assessing the proportion of usable rooftop photovoltaic area is more direct. Compared to previous methods that used indirect urban or building information as the classification method for the proportion of usable rooftop photovoltaic area, this invention directly classifies building rooftops based on their direct characteristics, namely the depth characteristics of rooftop images, resulting in more targeted and convincing results.
[0047] 3. This invention has broader applicability. Previous methods for assessing the usable area of rooftop photovoltaics required additional urban building information, which is severely lacking in some small and medium-sized cities and underdeveloped areas. Satellite maps, however, have wide coverage, therefore this invention is applicable to assessing the potential of rooftop photovoltaic applications in all types of cities.
[0048] 4. This invention has strong scalability. Researchers in related fields can extend this invention, for example, by adopting new instance segmentation models, methods for calculating the proportion of usable photovoltaic area on building roofs, methods for calculating photovoltaic installation area, and methods for calculating photovoltaic production capacity. Attached Figure Description
[0049] The invention will now be further described with reference to the accompanying drawings.
[0050] Figure 1 This is the overall flowchart of the present invention;
[0051] Figure 2 This is a location diagram of Yangpu District, Shanghai, according to an embodiment of the present invention;
[0052] Figure 3 This is a schematic diagram of the multi-round semi-supervised roof clustering process based on the constrained seed k-means algorithm in an embodiment of the present invention;
[0053] Figure 4 The rooftop photovoltaic usable area ratio (R) of this invention embodiment a A schematic diagram of the calculation method;
[0054] Figure 5 R represents all roofs of the 12 seed classes and 3 non-seed classes in this embodiment of the invention. a Distribution;
[0055] Figure 6 R is the cluster of 100 roofs in this embodiment of the invention. a Distribution map;
[0056] Figure 7 This is a schematic diagram of the installation of a photovoltaic module array on a flat roof according to an embodiment of the present invention. Detailed Implementation
[0057] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0058] To more clearly and understandably illustrate the embodiments of the present invention, the present invention selects Yangpu District, Shanghai for the description of the embodiments, as shown in the appendix. Figure 2 As shown. The example is Yangpu District, Shanghai.
[0059] like Figure 1 As shown, this embodiment uses a method for assessing the potential of urban building rooftop photovoltaic applications based on open-source satellite maps and deep learning, as proposed in this invention. The method includes the following steps:
[0060] S1. High-resolution satellite map acquisition: Satellite imagery of Yangpu District, Shanghai, was acquired using the TuSimple Earth 4 software, with a spatial resolution of 0.255m / pixel. Figure 2 As shown;
[0061] S2. Data preprocessing:
[0062] S21. Image cropping: Using the osgeo library in Python, the satellite images in the example are cropped into 512×512 pixel images, totaling 5270 images;
[0063] S22. Image quality enhancement: Use the cv2 library in Python to perform gamma correction on overly dark or overexposed images. Then, based on the cv2 library, use the Contrast Limited Adaptive Histogram Equalization (CLAHE) algorithm to perform histogram equalization on all images to enhance image contrast while suppressing noise.
[0064] S3. Construct the training dataset for the instance segmentation model:
[0065] S31. Image Sampling and Labeling: 10% of the images obtained in S2 were sampled at equal intervals, with additional sampling performed on high-density built-up areas, resulting in a total of 634 sample images. LabelMe software was used to label the image set, creating the original dataset for model training.
[0066] S32. Data Augmentation: Nine data augmentation methods from Python's imgaug library were used to augment the original labeled dataset, including cloud simulation, cropping, flipping, snowflake simulation, Gaussian blur, multiply, rotation, fast snowy landscape simulation, and translation, resulting in a dataset containing 6340 images. The augmented images and annotation files were saved in MSCOCO dataset format.
[0067] S33. Dataset Splitting: The dataset obtained in S32 is split based on random sampling, with the training set, cross-validation set, and test set accounting for 70%, 15%, and 15% of the total dataset, respectively.
[0068] S4. Establishing an instance segmentation model: In this embodiment of the invention, a Mask R-CNN instance segmentation model based on a deep residual network (ResNet50) is established using the PyTorch library. The model is trained using the dataset obtained in S3. The optimizer is stochastic gradient descent (SGD), with a learning rate of 0.00025 and a momentum of 0.9.
[0069] In this embodiment of the invention, the average precision (AP) is used as the performance evaluation metric for the instance segmentation model. The formulas for calculating the precision and recall of roof instance segmentation are as follows:
[0070]
[0071]
[0072] In the formula, N TP N P and N GT TP represents the number of predicted and ground truth boxes with an intersection-union ratio (IU) greater than a specified threshold, the total number of predicted boxes, and the total number of ground truth boxes, respectively. IU is the ratio of the area of intersection between predicted and ground truth boxes to the union of the two boxes. The AP calculation method for roof instance segmentation is as follows: ① Sort the predicted boxes by confidence level; ② Add a predicted box each time and calculate the cumulative recall and precision; ③ Plot the recall and precision of each step into a PR curve, and the area enclosed by the curve and the coordinate axis is the AP. This embodiment of the invention uses the AP calculation method of the MSCOCO dataset, interpolating the recall between 0 and 1 with an interval of 0.01. The AP calculation method is as follows:
[0073]
[0074]
[0075] In the formula, p interp(r) The precision at each interpolated recall is the maximum precision obtained from the direction of increasing the current recall. In this embodiment of the invention, AP50 is selected as the evaluation metric, that is, the intersection-union ratio (IoU) between the predicted bounding box and the ground truth bounding box for the roof is greater than 0.5, indicating that the detection is correct. After 150 rounds of training, the Mask R-CNN instance segmentation model achieved an AP50 of 0.92 for both object detection and instance segmentation tasks on the test set.
[0076] S5. Roof Instance Extraction: The Mask R-CNN instance segmentation model trained in S4 was used for inference to extract building roofs from all satellite images obtained in S3. Each roof was stored as an image with a black background, and the pixel area and coordinates of each roof were recorded. Only roof prediction boxes with a confidence score greater than 0.9 were extracted during the roof extraction process. A total of 23,597 roof instances were extracted.
[0077] S6. Roof Depth Feature Extraction: The backbone network of the Mask R-CNN instance segmentation model trained in S4 is used to extract the depth features of each roof image. The extracted features are pooled using a 1×1 dimensional MaxAdaptivePool2d algorithm, and the combination of features from layers 1, 2, and 3 of the backbone output is selected as the depth feature of each image.
[0078] S7. Multi-round semi-supervised roof clustering process based on constrained seed k-means algorithm: Based on the depth features of all roofs extracted in S6, multi-round semi-supervised clustering optimization is performed, as follows: Figure 3 As shown:
[0079] S71. Determine similarity criteria: Based on the characteristics of building roofs, propose similarity criteria, including: ① The two roofs are of the same type (flat roof / sloping roof). ② The two roofs have similar orientations. When two roofs simultaneously satisfy conditions ① and ②, they are considered strongly similar; when they only satisfy condition ①, they are weakly similar. Otherwise, the two roofs are not similar.
[0080] S72. Pre-clustering: Based on the depth features of all roofs, the number of clusters is set to 100, and the unsupervised k-means clustering algorithm is used for pre-clustering.
[0081] S73. Seed Class Selection: Referring to the roof similarity criteria in S71, select the class with the most similar roofs from the pre-clustering results as the seed class. Determine the characteristics of the most common similar roofs in each seed class (flat / sloping roof, similar / dissimilar orientation) as the class features and record them. Then, remove roofs that do not meet the class similarity characteristics (non-seed roofs) from each seed class, and label the remaining roofs (seed roofs) as that class. Simultaneously record the number of roofs in each seed class after clustering (N). i ) and the number of filtered seed roofs (N) si ), calculate N si and N i The ratio P;
[0082] S74. Semi-supervised clustering: Based on the seed class, the number of clusters is set to 100, and the constrained seed k-means algorithm is used for semi-supervised clustering.
[0083] S75. Multi-round semi-supervised clustering and seed class optimization: The method in S73 is used to screen seed classes from clusters outside the seed class (non-seed classes), and seed roofs are filtered from the seed classes. When no new seed classes appear and the p-value of all seed classes is greater than 0.5, the clustering results are output for subsequent analysis. Otherwise, S73 and S74 are repeated. When p ≥ 0.5, it indicates that the proportion of seed vertices in the seed class is greater than 50%, indicating that the seed vertices in the seed class have strong similarity characteristics. In the embodiment of the present invention, after 8 rounds of semi-supervised clustering and optimization (excluding pre-clustering), a total of 85 seed classes and 15 non-seed classes are obtained. The p-value of all seed classes is greater than 0.6, and the p-value of 46 seed classes is greater than 0.9. At the same time, in the 8th round, the number of seed roofs and all roofs in the seed classes reached 16,600 and 19,773, respectively, accounting for 70.3% and 83.8% of the total number of extracted roofs. The results show that using the roof clustering method proposed in this invention, the vast majority of roofs are well classified, and the roofs in each seed class have good similarity.
[0084] S8. Total Roof Area Assessment and Sloping Roof Area Correction: Based on the pixel area of each roof obtained in S5, the actual roof area is calculated; simultaneously, based on the clustering results in S7, the classes with the characteristic of sloping roofs are corrected. The roof area calculation method is as follows:
[0085]
[0086] In the formula, S pixel and S roof,mask S1 represents the area represented by a single pixel in the high-resolution satellite image, and S8 represents the roof pixel area extracted from the satellite image. θ is the pitch angle of the sloping roof. Based on the provisions of GB 50693-2011 "Technical Specification for Sloping Roof Engineering", the pitch angle of the sloping roof is set to 30°, and its area is corrected accordingly. The calculated roof area of the example building (Yangpu District, Shanghai) is 11.53 km². 2 Of which, the flat roof area is 7.63 km². 2 The pitched roof area is 3.9 km². 2 0.52km 2 This represents the increased area after the correction.
[0087] S9. Calculate the usable area ratio of rooftop photovoltaics (R) a ): Sample each type of roof in the clustering results and calculate the proportion of rooftop photovoltaic usable area:
[0088] S91. Calculation Method: First, for all roofs, the area occupied by building elements is excluded; second, for pitched roofs, the portion facing east-south-west is the usable portion, such as... Figure 4 As shown;
[0089] S92. Sampling Method: For non-seed classes and seed classes with P ≥ 0.9, first use systematic sampling to extract 12 roofs, and then use the method in S91 to calculate R for these roofs. a Remove R a Find the maximum and minimum values of the roofs, and calculate the average R value for all remaining roofs. a For seed classes with p < 0.9, the systematic sampling method is first used to extract... One seed roof (Group A) and For the non-seed roofs (Group B), calculate the R of these roofs using the method in S91. a Remove R from each group a Find the maximum and minimum values of the roofs, and calculate the average R value for all remaining roofs. a .
[0090] To verify the reliability of the roof clustering results of this invention, i.e., the R value of each roof category... aTo assess similarity, using the p-value as a reference standard, three seed classes were randomly selected at 0.1 intervals, for a total of 12 seed classes. Simultaneously, three non-seed classes were randomly selected. The R-value for all roofs within these seed classes was calculated. a The calculation results are as follows Figure 5 The results showed that, within the seed classes, the difference between the upper and lower quartiles of all 11 seed classes was less than 0.1. However, for non-seed classes, the difference between the upper and lower quartiles reached 0.4. Therefore, it can be considered that using the roof clustering method proposed in this invention, the R-value of roofs in the seed classes is... a The similarity is relatively strong, while the similarity of roofs in non-seed classes is relatively weak.
[0091] Meanwhile, in order to test the R proposed in S9 a To assess the accuracy of the sampling calculation method, this method was applied to the aforementioned 15 sample classes, and the deviation of the calculation results was compared, as shown in Table 1. The results show that, compared to calculating R for all roofs in each class... a In comparison, the R obtained by the sampling calculation method a The deviation is no greater than 12%, and the R-values for 10 classes are... a The deviation is less than 6%. Therefore, the sampling calculation method proposed in this invention can be applied to evaluate the R-squared value of roof clustering at the urban scale. a .
[0092] Table 1. Proportion of Available Rooftop Photovoltaic Area (R) a Results and biases of sampling calculation methods
[0093]
[0094]
[0095] The R-values for all 100 roof clusters in this embodiment of the invention were obtained using a sampling calculation method. a ,like Figure 6 As shown. This series of R... a Applied to 100 types of rooftops, the calculated usable area of photovoltaic power on building rooftops in the example (Yangpu District, Shanghai) is 7.64 km². 2 ;
[0096] S10. Calculation of photovoltaic installation area and production capacity:
[0097] S101. The formulas for calculating the rooftop photovoltaic installation area and installed capacity are as follows:
[0098] S install =N cell ×S cell (6)
[0099] P install =N cell ×P cell (7)
[0100]
[0101] D m =W cell ×cosβ+L m (9)
[0102] S ava =S roof ×R a (10)
[0103] In the formula, N cell S cell and P cell These represent the number of photovoltaic modules, the area of a single module, and its peak power, respectively. a and D m The spacing between the front and rear rows of the photovoltaic array is to prevent them from shading each other, and also to allow for maintenance. L cell and W cell For the length and width of a single photovoltaic module. β and L m For the tilt angle of the PV modules and the spacing of the pedestrian walkways used for maintenance. S ava It refers to the area of the roof that can be used for photovoltaic installation, such as Figure 7 As shown.
[0104] This embodiment of the invention (Yangpu District, Shanghai) is located at 31°N latitude. Referring to data from the Global Solar Atlas (GSA), the optimal PV tilt angle for Shanghai is 25°. The photovoltaic product selected is the LONGi Hi-MO6 Explorer LR5-54HTH-430M. The photovoltaic system parameters are shown in Table 2.
[0105] Table 2 Main parameters of photovoltaic modules and systems in the embodiments
[0106]
[0107] In this embodiment of the invention, when installing photovoltaic modules on a flat roof, a shading prevention distance and a maintenance distance must be reserved for the photovoltaic array. However, no such distance is reserved for pitched roofs, and the photovoltaic modules are installed parallel to the roof surface. This refers to section D of GB 50797-2012 "Design Code for Photovoltaic Power Stations". a The calculation method is as follows:
[0108]
[0109] In the formula, L a It is the length of the tilted surface of the PV array. It studies the latitude of the city. The D-axis of the photovoltaic array on the flat roof is calculated. a The maintenance spacing L for flat roofs is set at 1.95m. mGiven a height of 1m, the rooftop photovoltaic installation area of the example (Yangpu District, Shanghai) is 4.24km². 2 The installed capacity is 936.67MW.
[0110] S102, The formula for calculating photovoltaic capacity is:
[0111] E = S install ×R annual ×eff cell ×eff sys (12)
[0112] In the formula, R annual It studies the intensity of solar radiation on a horizontal plane in the area where the city is located. cell and eff sys These represent the photovoltaic module efficiency and the photovoltaic system efficiency, respectively. The annual solar radiation in the example (Yangpu District, Shanghai) is 1314.72 kWh / year, and the calculated annual photovoltaic capacity of the building rooftop in the example (Yangpu District, Shanghai) is 1047.74 GWh / year.
[0113] In the description of this specification, references to terms such as "embodiment," "example," "by way of," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.
[0114] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.
Claims
1. A method for assessing the potential of urban building rooftop photovoltaic applications based on open-source satellite maps and deep learning, the method comprising the following steps: S1. High-resolution satellite map acquisition; S2, Data Preprocessing: The satellite images obtained in S1 are cut into sizes suitable for the input of the deep learning model, and the image quality is enhanced. S3. Construct the training dataset for the instance segmentation model: Sampling and roof labeling are performed on the satellite image processed in S2, and data augmentation is performed. The augmented dataset is then split into a training set, a cross-validation set, and a test set. S4. Establish an instance segmentation model; build a Mask R-CNN instance segmentation model based on a deep residual network ResNet 50 using the PyTorch library, and train the model using the dataset obtained in S3; the optimizer is stochastic gradient descent, the learning rate is 0.00025, and the momentum is 0.
9. S5. Roof instance extraction; S6. Roof depth feature extraction; S7. Multi-round semi-supervised roof clustering process based on constrained seed k-means algorithm; S8. Total roof area assessment and pitched roof area correction; S9. Calculate the usable area ratio of rooftop photovoltaic (PV) ); S10, Calculation of photovoltaic installation area and production capacity; The S6 roof depth feature extraction process requires the use of the backbone network of the instance segmentation model trained in S4 for extraction; The S7 multi-round semi-supervised roof clustering process based on the constrained seed k-means algorithm includes the following steps: S71. Based on the characteristics of building roofs, similarity criteria are proposed, including: ① the two roofs are of the same type, either flat or pitched; ② the two roofs have similar orientations. Two roofs are considered strongly similar if they satisfy both conditions ① and ②; weakly similar if they satisfy only condition ①; otherwise, they are not similar. S72. Pre-clustering: Based on the depth features of all roofs, the number of clusters is set, and the unsupervised k-means clustering algorithm is used for pre-clustering. S73. Seed Class Selection: Referring to the roof similarity criteria in S71, select the class with the most similar roofs from the pre-clustering results as the seed class; determine and record the characteristics of the most similar roofs in each seed class based on whether they are flat or pitched roofs, and whether their orientations are similar or dissimilar; then remove roofs that do not meet the similarity characteristics of each seed class, and label the remaining roof types as that class; simultaneously record the number N of roofs in the seed classes after clustering. i And the number of filtered seed roofs N si Calculate N si and N i The ratio P; S74. Semi-supervised clustering: Semi-supervised clustering is performed using the constrained seed k-means algorithm based on the seed class; S75. Multi-round semi-supervised clustering and seed class optimization: The method in S73 is used to select seed classes from non-seed classes and filter seed roofs from seed classes. When no new seed classes appear and the p-value of all seed classes is greater than 0.5, the clustering results are output for subsequent analysis; otherwise, S73 and S74 are repeated. When p ≥ 0.5, it means that the proportion of seed vertices in the seed class is greater than 50%, indicating that the seed vertices in the seed class have strong similarity characteristics. The S9 calculation method calculates the proportion of rooftop photovoltaic usable area ( The process includes: S91. Calculation method: For all roofs, the area occupied by building elements is excluded; secondly, for pitched roofs, the parts facing east-south-west are the usable parts. S92. Sampling Method: For non-seed classes and seed classes with P ≥ 0.9, first use systematic sampling to extract 12 roofs, and then use the method in S91 to calculate the p-values of these roofs. Remove Find the maximum and minimum values of the roof surfaces, and calculate the average of all remaining roof surfaces. For seed classes with P < 0.9, firstly, use systematic sampling to extract ⌈P*10⌉+2 seed roofs and ⌊(1-P)*10⌋+2 non-seed roofs, and then use the method in S91 to calculate the values of these roofs. Remove from each group Find the maximum and minimum values of the roof surfaces, and calculate the average of all remaining roof surfaces. .
2. The method for assessing the potential of urban building rooftop photovoltaic applications based on open-source satellite maps and deep learning as described in claim 1, characterized in that, The spatial resolution of the S1 high-resolution satellite map is less than 1 m / pixel.
3. The method for assessing the potential of urban building rooftop photovoltaic applications based on open-source satellite maps and deep learning as described in claim 1, characterized in that, In the S5 roof instance extraction process, the instance segmentation model trained by S4 is used to extract building roof instances and store them as images, while recording the pixel area and coordinates of each roof.
4. The method for evaluating the potential of urban building rooftop photovoltaic applications based on open-source satellite maps and deep learning, as described in claim 1, is characterized in that... The S8 total roof area assessment and pitched roof area correction process requires the various roof feature records obtained in S7. For roofs with the feature of pitched roofs, area correction is performed using the following formula: (1) In the formula, and These represent the area represented by a single pixel in the high-resolution satellite image in S1, and the roof pixel area extracted from the satellite image in S8, respectively. The angle of inclination of the sloping roof.
5. The method for evaluating the potential of urban building rooftop photovoltaic applications based on open-source satellite maps and deep learning, as described in claim 1, is characterized in that... The calculation of the S10 photovoltaic installation area and capacity includes: S101. The formulas for calculating the rooftop photovoltaic installation area and installed capacity are as follows: (2) (3) (4) (5) (6) In the formula, , and These represent the number of photovoltaic modules, the area of a single module, and its peak power, respectively. and The spacing between the front and rear rows of the photovoltaic array is to prevent them from shading each other, and to allow for maintenance. and The length and width of a single photovoltaic module; and The tilt angle of the PV modules and the spacing of the pedestrian walkways used for maintenance; This refers to the area of the roof that can be used for photovoltaic installations; S102, The formula for calculating photovoltaic capacity is: (7) In the formula, It studies the intensity of solar radiation on a horizontal plane in the area where the city is located. and These refer to photovoltaic module efficiency and photovoltaic system efficiency, respectively.
Citation Information
Patent Citations
Roof photovoltaic space-time prediction method
CN114332597A
Wind uplift resistance reliability analysis method for valve hall upright lockrand metal roof system
CN114707383A
