Method and system for evaluating building integrated photovoltaic potential in high-density urban area

Abstract

The invention provides a method and system for evaluating the integrated photovoltaic potential of buildings in a high-density urban area, and the method comprises the steps: obtaining the color point cloud data of a city through the photographing and surveying of an unmanned plane, introducing the physical priori knowledge through algorithm design, and assisting a neural network in extracting a feature vector for representing the characteristics of a building group from the data; comprising the steps of point cloud sampling, normal estimation, window-wall ratio estimation and geographic orientation angle calculation, so that the model can learn implicit relation between the urban form and a photovoltaic potential prediction result; aiming at a scene of comprehensively realizing photovoltaic integration popularization of medium and large-scale buildings, the power generation efficiency of building facades and photovoltaic modules mounted at different positions of windows and walls is considered, feature vectors are input into a regression prediction framework formed by a one-dimensional point cloud neural network, evaluation errors are reduced, systematic deviation of resolution is avoided, and the evaluation accuracy is improved. Data acquisition and prediction are facilitated, and the function of predicting the maximum photovoltaic power generation potential after the building is comprehensively paved with the BIPV is realized.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of urban photovoltaic energy potential evaluation, and particularly relates to a method and system for evaluating building-integrated photovoltaic potential in high-density urban areas. BACKGROUND

[0002] In the field of urban photovoltaic energy potential evaluation, it is usually necessary to evaluate the maximum building-integrated photovoltaic potential of a medium-to-large scale region to provide a scientific basis for urban-level solar deployment strategies. When a point cloud neural network model directly predicts the maximum building-integrated photovoltaic (BIPV) potential by inputting city medium-to-large scale point cloud data, there are three technical problems caused by the inherent defects of point cloud data: 1) It is difficult to learn the geographical orientation information of buildings, and the prediction of building light radiation is disordered; 2) It is difficult to quickly extract the building surface attributes, and the efficiency of photovoltaic components in different installation areas is difficult to distinguish; 3) The spatial entity continuity is insufficient, and the spatial occlusion relationship modeling is ineffective.

[0003] The disadvantages of the prior art include: 1) Morphological parameter dependency problem. Traditional methods (such as multiple linear regression MLR) rely on expert subjective selection of morphological parameters (such as FAR, BD), resulting in large differences in method effectiveness in different regions, and the same index system is difficult to generalize; Since traditional regression prediction (such as multiple linear regression MLR) requires a large number of samples for index screening, suitable urban morphological indicators are selected by comparison. This method has strong artificial dependence, and the selected indicators have regional suitability and do not have universal generalization performance; and the index screening is highly subjective and lacks objective quantitative mechanisms. At the same time, a single urban morphological index system cannot comprehensively describe complex urban morphological changes.

[0004] 2) The input of the existing point cloud neural network (such as PointNet and Point Transformer) is only a set of spatial coordinate points, which usually uses T-Net or voxel-based spatial encoding to eliminate the disorder of point cloud and achieve rotation invariance. However, this method is commonly used for segmentation and classification tasks, and cannot enable the point cloud neural network to learn the spatial occlusion and geographical orientation features of buildings in the regression prediction of BIPV potential; T-Net structure learns the spatial transformation matrix of point cloud to realize the alignment and standardization of point cloud data, thereby improving the robustness of neural network to disorder and rotation changes. The voxel-based spatial encoding introduces local spatial relationship information for each point, enhancing the model's understanding of local structure and relative position in point cloud. These methods perform well in point cloud classification, segmentation, and other tasks, but have limited effect in BIPV photovoltaic maximum potential regression prediction. The reason is that point cloud data only describes the position of discrete coordinate points and cannot express the continuous geometric boundary and spatial topological relationship of building surface. Photovoltaic potential prediction relies on accurate description of the spatial occlusion relationship between buildings and the relative orientation of the surface to the sun, which cannot be simply reflected by spatial transformation or encoding. There is no entity occlusion between points in point cloud data, and overall rotation does not change the orientation information of buildings in real geographic space, so existing methods cannot reflect the actual physical environment, affecting the accuracy of regression prediction.

[0005] 3) There is a difference in efficiency between photovoltaic modules installed on walls and windows, and the area and distribution of building surfaces are different, making it difficult to quantify by a single index.

[0006] Point cloud data is essentially a collection of unordered points, and point cloud neural networks cannot directly learn these unordered points and the building surfaces and physical properties they represent, making it difficult for point cloud neural networks to distinguish between effective light-receiving surface area and photovoltaic components in different installation locations. Current photovoltaic potential evaluation methods usually use the average window-to-wall ratio (WWR) obtained under different building formats as a normalization parameter (for example, commercial office buildings are empirically assigned a window-to-wall ratio of 0.4-0.8). This approach ignores the differences in building morphology. For example, buildings with all-glass curtain wall structures and brick-concrete structures are assigned the same window-to-wall ratio parameter due to the same building format, but there are significant differences in surface area distribution and photovoltaic suitability.

[0007] Existing techniques for quickly predicting building performance by using convolutional neural networks (CNN) to learn building geometric representation: First, voxelize the building block, i.e., construct a three-dimensional voxel grid within the space range of the wrapped building body. Mark the voxel units covered by the building, 0 for blank areas and 1 for building entity areas. Then, obtain a set of voxel data containing voxel spatial positions (x, y, z) and their corresponding label information (0 or 1) as the input of the three-dimensional convolutional neural network (3D CNN) and for subsequent prediction of building performance indicators. The neural network model used is a three-dimensional convolutional neural network that can directly learn and represent the structural characteristics of the building body in three-dimensional space.

[0008] However, the prediction accuracy of the above method is closely related to the voxel resolution. Specifically, when the target building volume is large, if a smaller voxel unit is selected to ensure high resolution, the calculation cost will increase significantly; if a larger voxel unit is used, the prediction accuracy may decrease. Therefore, this method is suitable for performance prediction of single buildings, but it is not effective for building groups. Kiti et al. converted LiDAR data into a digital surface model and used the solar radiation model tool in ArcMap software to evaluate the maximum photovoltaic potential of roof buildings in Auckland (located in New Zealand). The principle of this method is to calculate the total solar radiation received by each pixel in the grid within a specified time period based on shadow analysis. The disadvantage is similar to voxelization, which is that the simulation accuracy is strongly related to the resolution, and it can only evaluate the roof radiation, without considering the building facade. Chen et al. used a deep semantic segmentation network based on satellite images to extract the building base surface from the images, used the ALOS World 3D-30 m (AW3D30) dataset to interpolate the building base surface to generate a three-dimensional building block, and finally used a hemispherical view algorithm to estimate the solar radiation received by the roof and the photovoltaic power generation potential. This algorithm also does not consider the evaluation of facade photovoltaic, and due to the low accuracy of the dataset used for interpolating building height, the building details are not enough, which will also cause evaluation errors. SUMMARY

[0009] The technical problem to be solved by the present application is to provide a method and system for evaluating the integrated photovoltaic potential of buildings in high-density urban areas, which can be used to predict the maximum photovoltaic power generation potential after the buildings are fully covered with BIPV.

[0010] The technical solution adopted by the present application to solve the above technical problem is a method for evaluating the integrated photovoltaic potential of buildings in high-density urban areas, comprising the following steps: S1: obtaining city point cloud through surveying and mapping; S2: separating building point cloud from city point cloud using RANSAC algorithm, generating three-dimensional model of building based on building footprint and building point cloud; S3: extracting data features from three-dimensional model through point cloud sampling, normal estimation, geographic orientation calculation and window-wall ratio calculation, and fusing into feature vector; S4: dividing point cloud data and space range, inputting feature vector into regression prediction model composed of one-dimensional point cloud neural network, and outputting predicted value of photovoltaic power generation potential of regional BIPV; S5: obtaining radiation intensity of building surface through solar radiation simulation, and calculating evaluation value of photovoltaic power generation potential; S6: evaluating the performance of the regression prediction model through different evaluation indexes and reference models.

[0011] According to the above scheme, in the step S3, the specific steps are: S31: Surface sampling is performed with the surface area of the three-dimensional model as the weight, a minimum value of the number of surface sampling points is set, and point cloud distribution is performed on the remaining points to obtain the spatial coordinates of the point cloud; S32: Normal estimation is performed on each point on the surface of the three-dimensional model to obtain the normal vector of the point cloud, and building entity information is retained; S33: The geographic orientation angle of the building is extracted based on the three-dimensional model, and the point cloud distribution is fused through spatial connection; S34: The window-wall ratio of the building unit is extracted using a Gaussian mixture model combined with the building point cloud; S35: The data features obtained in steps S31 to S34 are fused in a spatially related manner to form a feature vector.

[0012] Further, in the step S32, the specific steps are: S321: Neighborhood search is performed, and KDTree algorithm based on binary tree principle is used to define the neighborhood information of the spatial points by setting the neighborhood radius and the number of nearest neighbors of the point cloud; S322: Covariance matrix is constructed, the spatial vector mean of the neighborhood point set is calculated for each spatial point and the point set in its neighborhood, the covariance matrix is constructed based on the mean, and the eigenvalue and eigenvector are obtained through covariance matrix calculation; the smallest eigenvalue is screened out and the corresponding eigenvector is calculated, the direction away from the model is set as the normal direction, which is used as the normal vector; the normal vector of all point clouds is obtained by traversing the above steps.

[0013] Further, in the step S33, the specific steps are: S331: (0, 1, 0) is set as the north direction vector, and the direction from the north direction vector to the normal vector in the XY plane is set as the angle calculation direction, and the angle value range is [0, 360); the angle is divided into 16 partitions with 22.5° as the boundary; S332: The facade surface of the three-dimensional model is separated, and the normal vector of the centroid of the facade surface is calculated; the normal vector is projected onto the XY plane, and the included angle between the normal vector and the north direction vector and the angle partition to which the angle belongs are calculated; S333: The facade surface area in the angle partition is aggregated, the angle partition with the largest area is screened out, and the average angle in the partition is calculated; the average angle is scaled to [-90, 90] to obtain the main orientation angle of the building, which is used as an index of the geographic orientation angle of the building.

[0014] Further, in the step S34, the specific steps are: S341: Separate the building point cloud by building footprint contour, and divide the building group into independent buildings; S342: Extract the RGB color and spatial coordinate values in the building point cloud; S343: Combine the RGB color and spatial coordinate values of the same point; use Gaussian mixture model to cluster the combined data, divide the building point cloud into window point cloud and non-window point cloud, and calculate the area of the building window element.

[0015] According to the above scheme, in step S4, the specific steps are: S41: Scale the corresponding features of different data features, so that the data range is reduced to the [-1, 1] interval; S42: After inputting the data, perform feature transformation including T-Net and matrix multiplication operation, specifically including orthogonal regularizer, dot product model and multi-layer perceptron; Nonlinear transformation is performed in the convolution layer and dense layer of the regression prediction model; Use tanh and LeakyReLu as activation functions respectively; S43: Align the features, improve the dimensions and capture the complex patterns and relationships in the data by applying Transform and MLP twice; use Global Max Pooling to summarize all data features into a global feature vector; S44: Convert and compress the global feature vector through MLP; set the Dropout regularizer to randomly discard part of the neurons to prevent overfitting; use linear activation function in the output layer to regress and predict the photovoltaic power generation potential.

[0016] According to the above scheme, in step S5, the specific steps are: S51: Simulate solar radiation by Perez scattering radiation model and ray tracing algorithm to obtain the radiation intensity of the building surface; S52: Calculate the photovoltaic power generation potential of the photovoltaic device deployed at different positions of the building according to the radiation intensity; S53: Screen the morphological factors that significantly affect the prediction, and build a multiple linear regression model representing the influence relationship between the morphological factors and the photovoltaic power generation potential.

[0017] According to the above scheme, in step S6, the specific steps are: S61: Use the root mean square error and the determination coefficient to evaluate the performance of the model; the root mean square error is used as the error rate index of the trained model, and the root mean square error and the determination coefficient are used to evaluate the prediction performance of the trained model; S62: Use the multiple linear regression model with urban morphological factors as independent variables and annual photovoltaic power generation potential as dependent variables as the reference model; select the stepwise regression method to screen significant explanatory variables.

[0018] System for evaluating building integrated photovoltaic potential in high-density urban areas, a point cloud sub-module for obtaining urban point cloud data through aerial surveying and mapping; a modeling sub-module for separating building point cloud from urban point cloud using a RANSAC algorithm, and generating a three-dimensional model of the building based on the building footprint and the building point cloud; a feature sub-module for extracting data features from the three-dimensional model through point cloud sampling, normal estimation, geographic orientation calculation and window-wall ratio calculation, and fusing the features into a feature vector; a prediction sub-module for delimiting the point cloud data and the spatial range, inputting the feature vector into a regression prediction model composed of a one-dimensional point cloud neural network, and outputting a predicted value of the photovoltaic power generation potential of the regional BIPV; a simulation sub-module for obtaining the radiation intensity of the building surface through solar radiation simulation, and calculating an evaluation value of the photovoltaic power generation potential; an evaluation sub-module for evaluating the performance of the regression prediction model through different evaluation indexes and reference models.

[0019] A computer memory having stored therein a computer program executable by a computer processor, the computer program performing a method for evaluating building integrated photovoltaic potential in high-density urban areas.

[0020] The present application has the following advantages: 1. The method and system for evaluating building integrated photovoltaic potential in high-density urban areas of the present application obtain urban color point cloud data through aerial surveying and mapping by unmanned aerial vehicle photogrammetry, and extract feature vectors representing the characteristics of the building group from the data; input the feature vectors into a regression prediction framework composed of a one-dimensional point cloud neural network to realize the function of predicting the maximum photovoltaic power generation potential after the building is fully paved with BIPV.

[0021] 2. The present application automatically executes the algorithm for extracting feature vectors based on the color point cloud data obtained by aerial surveying and mapping by unmanned aerial vehicle, and helps the neural network to learn and extract features by introducing physical prior knowledge through algorithm design, including point cloud sampling, normal estimation, window-wall ratio estimation, and geographic orientation angle calculation. The data structure of the point cloud combined with the normal can clearly express the entity information of the building itself, i.e. the direction indicated by the normal vector is the outside of the building, and the reverse is the inside; the estimation of the window-wall ratio helps the model to understand the difference in photovoltaic potential under different window-wall ratios, because the efficiency of window photovoltaic components is generally low, while the efficiency of wall components is higher; the calculation of the geographic orientation angle helps the model to identify the physical relationship between the prediction result and the orientation of the building through algorithm; the point cloud sampling with surface area as weight is to strengthen the model's learning of the distribution characteristics of the point cloud, so that the model can learn the implicit relationship between the urban form and the photovoltaic potential prediction result.

[0022] 3.The application is aimed at the scene of full photovoltaic integration of medium and large scale (such as 400*400m range) buildings, and proposes an algorithm framework for predicting the maximum photovoltaic power generation potential in the region, considering the building facade and the power generation efficiency of photovoltaic components installed in different positions of windows and walls, which is consistent with the actual application deployment, greatly reducing the evaluation error. The algorithm synthesis framework is based on neural network configuration, and the one-dimensional convolutional neural network in the algorithm can avoid the systematic resolution bias caused by voxelization or three-dimensional convolutional neural network. The point cloud data volume is small, and the algorithm uses point cloud data for prediction, which can greatly facilitate data acquisition and the convenience of prediction. The problems of entity missing, geographical orientation missing and building surface information missing of point cloud data in input neural network for photovoltaic potential regression prediction are solved, and the problems of distorted regression prediction results and fluctuating prediction range deviation from the actual situation are avoided.

[0023] Of course, implementing any product of the application does not necessarily require all the advantages described above. BRIEF DESCRIPTION OF DRAWINGS

[0024] In order to more clearly illustrate the technical solutions of the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are some embodiments of the present application, and those skilled in the art can also obtain other drawings according to these drawings without creative labor.

[0025] Figure 1 is a flowchart of the embodiment of the present application.

[0026] Figure 2 is a general framework diagram of the embodiment of the present application.

[0027] Figure 3 is a configuration diagram of the neural network of the embodiment of the present application.

[0028] Figure 4 is a roof solar radiation distribution diagram of the embodiment of the present application.

[0029] Figure 5 is a diagram of the influence of urban form factors on the overall photovoltaic power generation potential of urban blocks of the embodiment of the present application.

[0030] Figure 6 is a model training result comparison diagram of the embodiment of the present application. DETAILED DESCRIPTION

[0031] In order to make the purpose, technical scheme and advantages of the present application more clear, the following will further describe the present application in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, and are not used to limit the present application.

[0032] Embodiment 1 Reference Figure 1 The specific steps of the method for evaluating the building-integrated photovoltaic potential of high-density urban areas are as follows: S1: Obtain city point cloud by mapping; S2: Separate building point cloud from city point cloud using RANSAC algorithm, generate three-dimensional model of building based on building footprint and building point cloud; S3: Extract data features from three-dimensional model through point cloud sampling, normal estimation, geographic orientation calculation and window-wall ratio calculation, and fuse into feature vector; S4: Draw point cloud data and space range, input feature vector into regression prediction model composed of one-dimensional point cloud neural network, and output predicted value of photovoltaic power generation potential of regional BIPV; S5: Obtain radiation intensity of building surface through solar radiation simulation, and calculate evaluation value of photovoltaic power generation potential; S6: Evaluate the performance of the regression prediction model through different evaluation indexes and reference models.

[0033] Further, in step S3, the specific steps are as follows: S31: Surface sampling is performed with the surface area of the three-dimensional model as the weight, the minimum value of the number of surface sampling points is set, and the remaining number of points is distributed to the point cloud to obtain the spatial coordinates of the point cloud; S32: Normal estimation is performed on each point on the surface of the three-dimensional model to obtain the normal vector of the point cloud, and the building entity information is retained; S33: Extract the geographic orientation angle of the building based on the three-dimensional model, and fuse the point cloud distribution through spatial connection; S34: Use Gaussian mixture model to extract the window-wall ratio of the building monomer combined with the building point cloud; S35: Use the data features obtained in steps S31 to S34 to fuse and form a feature vector in a spatially related manner.

[0034] Further, in step S32, the specific steps are as follows: S321: Perform neighborhood search, and use KDTree algorithm based on binary tree principle to define neighborhood information of spatial points by setting neighborhood radius and number of nearest neighbors of point cloud; S322: Construct covariance matrix, calculate spatial vector mean of neighborhood point set for each spatial point and point set in its neighborhood; construct covariance matrix based on the mean, and obtain eigenvalue and eigenvector through covariance matrix calculation; select the smallest eigenvalue and calculate the corresponding eigenvector, set the direction away from the model as the normal direction, and use it as the normal vector; obtain the normal vector of all point clouds through the above steps.

[0035] Further, in step S33, the specific steps are: S331: Set (0, 1, 0) as the positive north direction vector, and set the direction from the positive north direction vector to the normal vector clockwise on the XY plane as the angle calculation direction, with the angle value ranging from [0, 360); divide the angle into 16 partitions with 22.5° as the boundary; S332: Separate the facade surface of the three-dimensional model, and calculate the normal vector of the centroid of the facade surface; project the normal vector to the XY plane, and calculate the included angle between the normal vector and the positive north direction vector and the angle partition to which the angle belongs; S333: Aggregate the facade surface area in the angle partition, select the angle partition with the largest area, and calculate the average angle in the partition; scale the average angle to [-90, 90] to obtain the main orientation angle of the building, which is used as an index of the geographic orientation angle of the building.

[0036] Further, in step S34, the specific steps are: S341: Separate the building point cloud by the building footprint contour to divide the building group into independent buildings; S342: Extract the RGB color and spatial coordinate values in the building point cloud; S343: Combine the RGB color and spatial coordinate values of the same point; use the Gaussian mixture model to cluster the combined data, divide the building point cloud into window point cloud and non-window point cloud, and calculate the area of the building window element based on this.

[0037] In step S4, the specific steps are: S41: Scale the corresponding features of different data features to reduce the data range to the [-1, 1] interval; S42: After inputting the data, perform feature transformation including T-Net and matrix multiplication operation, specifically including orthogonal regularizer, dot product model and multi-layer perceptron; perform nonlinear transformation in the convolution layer and dense layer of the regression prediction model; use tanh and LeakyReLu as activation functions respectively; S43: Align the features, improve the dimension, and capture the complex patterns and relationships in the data by applying Transform and MLP twice; use Global Max Pooling to summarize all data features into a global feature vector; S44: Convert and compress the global feature vector through MLP; set the Dropout regularizer to randomly discard part of the neurons to prevent overfitting; use the linear activation function in the output layer to regress and predict the photovoltaic power generation potential.

[0038] In step S5, the specific steps are: S51: simulate solar radiation by Perez scattering radiation model and ray tracing algorithm to obtain the radiation intensity of the building surface; S52: calculate the photovoltaic power generation potential of the photovoltaic device arranged at different positions of the building according to the radiation intensity; S53: screen morphological factors that significantly affect the prediction, and build a multiple linear regression model representing the influence relationship between the morphological factors and the photovoltaic power generation potential.

[0039] In step S6, the specific steps are: S61: use the root mean square error and the determination coefficient to evaluate the performance of the model; the root mean square error is used as an error rate index of the trained model, and the root mean square error and the determination coefficient are used to evaluate the prediction performance of the trained model; S62: use the multiple linear regression model with the urban morphological factors as the independent variable and the annual photovoltaic power generation potential as the dependent variable as the reference model; select the stepwise regression method to screen the significant explanatory variables.

[0040] The embodiment obtains urban color point cloud data by unmanned aerial vehicle photogrammetry, extracts feature vectors for representing the characteristics of the building group in the data, inputs the feature vectors into a regression prediction framework composed of one-dimensional point cloud neural networks, and realizes the function of predicting the maximum photovoltaic power generation potential after the building is fully paved with BIPV.

[0041] Embodiment 2 The steps of this embodiment are the same as those of embodiment 1, and the difference is that each step is applied to a specific example. Referring to Figure 2 , specifically comprising the following steps: S1: obtain urban color point cloud data by unmanned aerial vehicle photogrammetry; S2: based on the elevation and color information of the urban color point cloud data, use the RANSAC algorithm to separate the building point cloud from the urban point cloud; use the City3D model to generate a three-dimensional model based on the building footprint obtained from an open source website and the building point cloud data.

[0042] S3: obtain the spatial coordinates of the point cloud, the normal vector of the point cloud, the building geographical orientation angle, and the window-wall ratio of the building monomer from the three-dimensional model by point cloud sampling, normal estimation, geographical orientation calculation, and window-wall ratio calculation; use the spatial correlation to fuse the data features to form a feature vector.

[0043] S31: point cloud sampling; when surface sampling is performed on the three-dimensional model, the surface area is used as the weight, refer to formulas (1) to (3); before sampling, the minimum value of the surface sampling point number is set to ensure the learning of the building detail features, and then the remaining point number is distributed.

[0044] (1) (2) (3) Let be the area of the i-th triangle, be the three vertices of the i-th triangle, and N be the total number of triangles; , , be the cumulative distribution function of the i-th triangle; be the i-th triangle; be the sampling point, and be in the range of [0, 1] and ; by generating a random number , find the corresponding interval and triangle index , and uniformly sample inside the triangle.

[0045] S32: Normal estimation; surface point cloud sampling with three-dimensional model surface area as weight, combined with normal estimation to retain building entity information, to solve the unordered problem of point cloud convolution; S321: Neighborhood search, using KDTree algorithm based on binary tree principle to define the neighborhood information of spatial points by setting the neighborhood radius and the number of nearest neighbors. This method not only considers the geometric density of the neighborhood, but also avoids the calculation cost caused by too large neighborhood. This algorithm is widely used in normal estimation of three-dimensional point cloud because it can quickly and efficiently determine the neighboring points of a given point in three-dimensional space; S322: Constructing the covariance matrix, for each spatial point and the point set in its neighborhood, calculate the spatial vector mean of the neighborhood point set (4), which reflects the center position of the local point cloud. Based on the mean, construct the covariance matrix (5), and obtain the eigenvalue and eigenvector (6-7) through covariance matrix calculation. Select the smallest eigenvalue and calculate the corresponding eigenvector, set the direction away from the model as the normal direction, and take it as the normal vector. Through the above steps, all point cloud normal vectors are obtained.

[0046] (4) (5) (6) For the neighborhood points of the spatial point , ​​​​​is the number of neighborhood points, is the covariance matrix, is the eigenvalue, is the unit vector, is the eigenvector.

[0047] S33: Building geographic orientation calculation; building geographic orientation information is extracted based on the three-dimensional model established by point cloud, and the point cloud distribution is fused through spatial connection to strengthen the individual characteristics of the building group point cloud and solve the point cloud rotation invariance problem; S331: Set (0, 1, 0) as the north direction vector, set the direction from the north direction vector to the normal vector on the XY plane as the angle calculation direction, and the angle value range is [0, 360); divide the angle into 16 partitions with 22.5° as the boundary; S332: Separate the facade surface of the three-dimensional model, and calculate the normal vector of the centroid of the facade surface; project the normal vector to the XY plane, calculate the included angle between the normal vector and the north direction vector, and the angle partition to which the angle belongs; S333: Aggregate the facade surface area in the angle partition, select the angle partition with the largest area and calculate the average angle in the partition . Scale the average angle to [-90, 90] through (7) to obtain the main orientation angle of the building , which is used as an index of the building geographic orientation angle;

[0048] S34: Window-wall ratio calculation: use Gaussian mixture model (GMM) combined with building point cloud data (including color and spatial coordinate information) to extract building window-wall ratio to strengthen the individual characteristics of the building group point cloud; S341: Separate the building point cloud by the building footprint contour to divide the building group into independent buildings; S342: Extract the RGB color and spatial coordinate information in the building point cloud; S343: Combine the RGB and XYZ values respectively, and use Gaussian mixture model to cluster the data; The purpose of using Gaussian mixture model is to divide the building point cloud into window point cloud and non-window point cloud, and calculate the area of the building window element. Compared with the traditional k-means clustering algorithm, GMM considers the covariance of the feature vector (the covariance matrix allows each Gaussian component to have different shape and direction), which can adapt to the complex shape distribution of the data; on the other hand, the relative proportion of window and non-window point cloud data can be described by the mixing coefficient. Therefore, it can better meet the purpose of this step.

[0049] The Gaussian mixture model is completed iteratively using the log-likelihood function (8) and the EM algorithm (9)-(12). The entire iterative process includes initialization, expectation, and maximumization. Random initialization is used for initialization. The building point cloud is set to consist of window point cloud and non-window point cloud, so the number of sub-models of the Gaussian mixture model is set to 2. The mixing coefficients of the Gaussian distribution of window elements under the maximum log-likelihood function are obtained through iteration. Finally, the window area of ​​a single building is calculated by combining the three-dimensional model with formulas (13)-(15); (8) (9) (10) (11) (12) (13) (14) (15) set up for , as well as The set of parameters; The set of observed data points Each It is an observation sample; It is the total number of data points; It is the number of Gaussian distribution sub-models; It is the first The mixing coefficient of a Gaussian distributed component represents the proportion of that component in the total mixture; It is the first The probability density function of a Gaussian distribution, given the mean Covariance Matrix Under these conditions, data points The probability of; It represents data points From the first The posterior probability of the Gaussian component reflects, under the current parameter estimation, the likelihood of... Belonging to The probability of; It's a latent variable, just a data point. Whether or not a case is filed depends on the Gaussian component. ; , , Represents the new value after iteration; and These are the building window area and the building surface area, respectively. The building surface area is obtained through a 3D model.

[0050] S35: Calculate the spatial coordinates of the point cloud, the normal vector of the point cloud, the geographical orientation angle of the building, and the window-to-wall ratio of the individual building using the data features obtained through the above steps; finally, use spatial association to fuse the data features to form a feature vector.

[0051] S4: Define the point cloud data and spatial range, input the feature vector into a regression prediction framework composed of a one-dimensional point cloud neural network, and predict the photovoltaic power generation potential of the regional BIPV. A neural network consists of an input layer, hidden layers, and an output layer. The input layer is the receiver of the neural network, used to read the input data. The hidden layer extracts salient features from the data through a series of linear and nonlinear transformations. The output layer transforms the extracted features into output data through an activation function. See also... Figure 3 By configuring a neural network, a feature vector with eight dimensions and its corresponding photovoltaic potential label are obtained.

[0052] S41: To avoid gradient explosion or vanishing during training, feature scaling is applied to the feature data to reduce the data range to the [-1, 1] interval; appropriate feature scalers are set for different data features: 1. Point cloud spatial coordinates: The center of the bottom surface of the research unit is set as the origin of the coordinate system, and the data ranges of x, y, and z are [-200, 200], [-200, 200], and [0, 300], respectively; 2. Set the building window-to-wall ratio and the building's geographical orientation angle to [0, 1] and [-90, 90], respectively.

[0053] After setting the data range, the data is scaled according to formula (16). The normal vector has a range of [-1, 1], so it is not changed.

[0054] (16) in, The original value, These are scaled values.

[0055] S42: After inputting the data, a feature transformation (Transform) is performed, which includes the T-Net module and matrix multiplication operations. The T-Net uses an orthogonal regularizer internally to maintain the independence of the weight matrix, helping the model to distinguish the significant features of different data. Using the dot product allows the model to transform the feature vectors to maintain consistency of the features at different angles and positions, thereby improving the robustness and accuracy of feature processing. The purpose of setting up the multi-layer perceptron (MLP) is to convert the original features into high-dimensional features, enhancing the expression ability of the features. Non-linear transformation is performed within the convolutional and dense layers to extract and enhance the complexity and discriminability of the features, enabling the neural network to learn complex features. The tanh and LeakyReLu activation functions are used respectively to speed up the iteration and prevent the occurrence of gradient vanishing problems.

[0056] S43: Repeat the application of Transform and MLP to further align the features while increasing the dimension and capturing complex patterns and relationships in the data. After two Transform and MLP processes, use Global MaxPooling to aggregate the features of all data into a global feature to capture the most significant overall features.

[0057] S44: Convert and compress the global feature vector through MLP to enhance its expression ability and discriminability. Set up the Dropout regularizer to prevent overfitting by randomly discarding some neurons, improving the generalization ability of the model. In the output layer, use the linear activation function to realize the regression prediction of photovoltaic power generation potential.

[0058] S5: Evaluate the solar photovoltaic potential of the research unit by simulating solar radiation to obtain the radiation intensity of the building surface, and calculate the solar photovoltaic power generation potential based on the radiation intensity values of the building facade and roof surface; S51: Solar radiation simulation; Solar radiation simulation is completed using the Ladybug Tools tool in the Rhino & Grasshopper platform. Ladybug Tools is a powerful environmental analysis tool that supports EnergyPlus weather files and uses the Radiance module to simulate solar radiation. The Perez scattering radiation model applied in the Radiance module is suitable for building surface radiation calculation in urban environments. In addition, based on real physical principles, the ray tracing algorithm is used to simulate the propagation and interaction of light in three-dimensional space, effectively calculating the solar radiation on the roof and facade of the building. During the simulation, the simulation period is set to one year, the building surface is divided into 1x1 meter grids, and the annual cumulative solar radiation of each grid is calculated. The radiation values of the roof and building facade surface grids are aggregated respectively to obtain the annual cumulative solar radiation of the building facade and roof surface.

[0059] The reliability of the simulation data is verified by comparing the model; see Figure 4 To verify the reliability of the simulation data of Ladybug Tools, the DSM data of M City is used, and the Area Solar Radiation tool in ARCGIS is used to calculate the solar radiation per square meter of the roof in the same area, and it is compared with the roof radiation value simulated by Ladybug Tools. The Area Solar Radiation tool takes into account both climate characteristics and topographic parameters, and has been applied in many studies to estimate the solar power generation potential and has been proven to be reliable. By applying the default settings, the Area Solar Radiation tool will calculate the annual solar radiation of each grid on the roof according to the neighborhood DSM grid, and the annual radiation value distribution of the building roof in M City is obtained by simulation calculation.

[0060] Subsequently, the total annual roof solar radiation of each unit obtained by Ladybug Tools and the Area Solar Radiation tool is imported into SPSS for bivariate correlation analysis, using the PEARSON coefficient and two-tailed significance test. The test results show that the Pearson correlation is 0.791**, and the significance (two-tailed) is <0.001 (Table 1). The results show that the simulation of solar radiation on buildings using Ladybug Tools is reliable.

[0061] Table 1 Bivariate correlation analysis of simulation results of Ladybug Tools and Area Solar Radiation tool

[0062] S52: Since there are differences in the application of photovoltaic devices deployed on the roof, wall and window of the building, the detailed parameters of photovoltaic devices in different positions are considered respectively. Finally, the photovoltaic power generation potential is evaluated and calculated by the following formula:

[0063]

[0064] G is the average value of the annual average solar radiation per unit area ( ). 、 and represent the window area, wall area and building surface area of the building, respectively; and are the surface areas obtained directly by separating the roof surface of the three-dimensional model ; Representing the conversion efficiency of a photovoltaic module The efficiency of photovoltaic modules is set at 21% for roof modules, 17% for BIPV wall modules and 15% for BIPV window modules, respectively. PR represents the performance ratio of BIPV. PR is set at 80% .

[0065] S53: Screening morphological factors that significantly affect the prediction, building a multiple regression equation representing the influence relationship between morphological factors and power generation potential, see Figure 5 For related urban morphological parameters, see Table 2; Table 2: Brief introduction of related urban morphological parameters

[0066] Through multiple regression analysis, it is found that BD (building density), BH_cv (building height coefficient of variation), SA_mean (building surface area mean), OA_sd (standard deviation of azimuth angle), and WWR_mean (window-to-wall ratio mean) have a significant positive effect on the overall solar power generation potential of the study unit. The increase of BD and SA_mean increases the installable space of building roof or facade BIPV devices, thus promoting the increase of overall power generation potential. The positive effect of BH_cv on power generation potential is reflected through reasonable building layout, in which the staggered building layout effectively reduces the shading between buildings and improves the effective illumination area of the overall building. The increase of OA_sd means the decrease of the consistency of the orientation of buildings in the unit, which allows buildings to adjust their orientation according to their morphological characteristics to maximize the effective area of receiving solar radiation. In the building integrated photovoltaic (BIPV) system, the photoelectric conversion efficiency of window photovoltaic devices is generally lower than that of wall photovoltaic devices. There is a significant positive correlation between BD, FAR (floor area ratio), BH_mean (building height mean), and WWR_mean. This indicates that in high-density areas, high window-to-wall ratio phenomenon is common, mainly due to the fact that these areas are dominated by high-rise office buildings, which usually have large glass curtain wall areas in design to enhance natural light utilization and visual aesthetics. These areas provide more extensive BIPV installation space, thus having higher overall power generation potential. Therefore, the model shows that WWR_mean has a significant positive effect on the overall power generation potential (Total_PP). However, the small VIF and P_value less than 0.001 from the model results emphasize the significance of WWR_mean in explaining the overall power generation potential and its importance in the construction of the prediction model.

[0067] BEI (Building expansibility index) and SVF sd (Standard deviation of sky view factor) have significant negative effects on the overall solar power generation potential of the unit. The increase of BEI means that the utilization of longitudinal space in the unit by buildings increases, which leads to the aggravation of inter-building shading, thus reducing the effective light area of the whole. The increase of SVF sd indicates that the polarization of SVF is serious, and in the case of the same land area, the buildings are densely arranged and there are large open areas such as parks. The dense arrangement of buildings means that the shading between them is serious, which reduces the overall BIPV power generation potential. Through the above analysis, a multiple regression equation is built:

[0068] S6: Evaluate the performance of the regression prediction model through different evaluation indexes and reference models; S61: Performance evaluation index; five different indexes are used for evaluation, which are root mean square error (RMSE) and determination coefficient (R 2 ), the specific formulas are (20) and (21); among them, RMSE is used as the error rate index when training the CNN model, and RMSE and R 2 are used to evaluate the prediction performance of the trained multi-dimensional single-channel CNN model. Since they are independent of the size of the data, these indexes have been widely used as the standard for performance evaluation, which can directly show the performance of the prediction model after training. R 2 close to 1 indicates that the prediction accuracy of the regression model is high, and the smaller the values of the remaining indexes, the more accurate the model prediction.

[0069] (20) (21) is the true value of the first, that is, the actual observation value of the target variable predicted by the model; is the predicted value of the first, that is, the result predicted by the model according to the input data and the learned parameters; is the arithmetic mean of all true values , which represents the average level of the true value of the target variable in the data set; is the total number of data. is the standard deviation of the true value.

[0070] ​​S62: Reference model; Multivariate linear regression (MLR) was used as a traditional model for comparison analysis. MLR is a statistical method for modeling the relationship between input variables and output variables. Urban form factors were used as independent variables, and annual solar PV potential was used as the dependent variable. Stepwise regression was used to select significant explanatory variables. Specifically, this method uses F-test to evaluate whether the introduction of a new variable significantly improves the explanatory power of the model. If the variable passes the F-test and shows statistical significance, it is retained in the model, and t-test is performed on all variables already included in the model. As new variables are introduced into the model, if some previously introduced variables are no longer significant in subsequent t-tests, these variables are removed from the model. The above process is repeated until no new significant variables can be added, and there are no longer non-significant variables in the model. The equation is described by (22).

[0071] (22) is the annual PV potential, is a constant, is the coefficient of the impact factor, is the urban form impact factor.

[0072] The deep learning model and its training history are shown in Figure 6 . The Training Loss curve shows a gradual downward trend. From the 0th generation, the curve trend gradually decreases, accompanied by large fluctuations. Around the 170th generation, the fluctuation amplitude significantly decreases, and the curve downward trend gradually slows down. Finally, around the 270th generation, it tends to be stable, indicating that the model training has basically converged, as shown in Figure 6 (1) represents the training process. The final results show that the model R 2 is 0.937, and the RMSE is 2415.547 kWh / year, which indicates that the prediction model performs relatively well, with an average error of only 2415.547 kWh / year in predicting the solar potential of the 75 study units. In Figure 6 (2) represents the comparison between the predicted value and the actual value. It can be seen that the predicted curve closely fits the true value curve, with only a few units having a large error, making the overall model evaluation difference large and the average error high. Figure 6The comparison between the predicted values of the CNN and the traditional regression model in different three-dimensional models represented by (3) is the comparison between the algorithm model of this embodiment and the traditional multiple regression model. It can be seen that the prediction of the solar photovoltaic potential by the CNN is more accurate overall. The traditional regression model is only close to the prediction in one unit, and the solar photovoltaic potential in some units appears negative. It can be seen from the multiple regression equation that BEI and SVF sd have a significant negative impact on the solar photovoltaic potential. In units 5-7, on the one hand, the building density is relatively dispersed, indicating that SVF sd is relatively large. On the other hand, due to the small difference between the building density and the building height, the BH_cv and BD terms are small, and BEI has a large coefficient. This makes the predicted value small or even negative. There are some open areas in unit 8, and the building density in other areas is high, indicating that SVF sd is relatively large. At the same time, there are some high-rise buildings in the unit, which makes the negative term in the equation larger and the equation smaller, resulting in a negative value. It can be seen from 1-3 that the prediction of the traditional regression model is not stable, and the predicted values in different units are large and negative. However, the CNN model shows good prediction performance in units with different urban forms, which shows that the algorithm framework can better capture the influence mechanism of complex urban forms on solar photovoltaic potential.

[0073] In this embodiment, the data label used for training is changed to adapt to various scenarios flexibly. In addition to predicting photovoltaic potential, the algorithm structure remains unchanged and can be effectively used to predict building performance evaluation affected by building group urban form, such as outdoor thermal comfort, building group facade temperature average, etc. The algorithm framework can be used for building group form multi-objective optimization, street early urban design performance evaluation, etc.

[0074] It should be understood that the size of the serial number of each step in the above embodiment does not mean the order of execution. The execution order of each process should be determined according to its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0075] Embodiment 3 This embodiment is used to realize the principle of the above-mentioned method embodiment to build a system for evaluating the photovoltaic potential of high-density urban area buildings. The system includes a point cloud submodule, a modeling submodule, a feature submodule, a prediction submodule, a simulation submodule, and an evaluation submodule.

[0076] The point cloud submodule is used to obtain city point clouds through surveying and mapping. The modeling submodule is used to separate building point clouds from city point clouds using the RANSAC algorithm, and generate a three-dimensional model of the building based on the building footprint and the building point cloud. The feature submodule is configured to extract data features from the three-dimensional model through point cloud sampling, normal estimation, geographic orientation calculation, and window-wall ratio calculation, and fuse the features into a feature vector. The prediction submodule is configured to delimit the point cloud data and a spatial range, input the feature vector into a regression prediction model composed of a one-dimensional point cloud neural network, and output a predicted value of the photovoltaic power generation potential of the region BIPV. The simulation submodule is configured to obtain the radiation intensity of the building surface through solar radiation simulation, and calculate an evaluation value of the photovoltaic power generation potential. The evaluation submodule is configured to evaluate the performance of the regression prediction model through different evaluation indexes and reference models.

[0077] Each submodule is mainly configured to implement each step of the method embodiment, and will not be described here.

[0078] It should be noted that, according to the needs of implementation, each step / component described in the present application can be split into more steps / components, or two or more steps / components or part of the operations of the steps / components can be combined into a new step / component, to achieve the purpose of the present application.

[0079] The embodiment also includes a processor, a communication interface, a memory, and a communication bus; the processor, the communication interface, and the memory complete communication with each other through the communication bus; the memory stores a computer program, and when the program is executed by the processor, the processor executes the steps of the method for evaluating the photovoltaic potential of high-density urban area building integration.

[0080] The embodiment also provides a computer readable storage medium, which stores executable instructions, and the instructions make the processor implement the method for evaluating the photovoltaic potential of high-density urban area building integration when executed by the processor.

[0081] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can adopt a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects.

[0082] Moreover, the present application can adopt the form of a computer program product implemented on one or more computer usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer usable program code.

[0083] The present application is described with reference to the flowcharts of the method and the computer program product according to the embodiment 1 of the present application and the block diagrams of the devices (systems) of the embodiment 3. It should be understood that each flow or block in the flowcharts or block diagrams, and the combination of the flows or blocks in the flowcharts or block diagrams can be realized by computer program instructions.

[0084] These computer program instructions can be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart block or blocks Figure 1 Figure 1 The functions described can be implemented in one or more of the following:

[0085] These computer program instructions can also be stored in a computer readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable memory produce an article of manufacture including instructions which implement the function specified in the flowchart block or blocks Figure 1 Figure 1 The functions described can be implemented in one or more of the following:

[0086] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart block or blocks Figure 1 Figure 1 The functions described can be implemented in one or more of the following:

[0087] The above embodiments are only used to illustrate the design ideas and features of the present application, and the purpose is to make the skilled in the art understand the present application and to carry it out, the protection scope of the present application is not limited to the above examples. Therefore, any equivalent variation or modification made according to the principles disclosed herein is within the scope of the present application.​​​

Claims

1. A method for evaluating the potential of building-integrated photovoltaics in high-density urban areas, characterized by: Includes the following steps: S1: Obtain urban point cloud through surveying; S2: Use the RANSAC algorithm to separate building point clouds from urban point clouds, and generate 3D models of buildings based on building footprints and building point clouds; S3: Extract data features from the 3D model through point cloud sampling, normal estimation, geographic orientation calculation, and window-wall ratio calculation, and fuse them into a feature vector; S4: Define the point cloud data and spatial range, input the feature vector into the regression prediction model composed of a one-dimensional point cloud neural network, and output the predicted value of the photovoltaic power generation potential of the regional BIPV. S5: Obtain the radiation intensity of the building surface through solar radiation simulation and calculate the assessment value of photovoltaic power generation potential; S6: Evaluate the performance of the regression prediction model using different evaluation metrics and reference models.

2. The method for evaluating the potential of building-integrated photovoltaics in high-density urban areas according to claim 1, characterized in that: The specific steps in step S3 are as follows: S31: Surface sampling is performed using the surface area of ​​the 3D model as the weight. A minimum value for the number of surface sampling points is set, and the remaining points are allocated to the point cloud to obtain the spatial coordinates of the point cloud. S32: Estimate the normal vector of the point cloud by performing normal estimation on each point on the surface of the 3D model, and preserve the building entity information; S33: Extract the geographic orientation angle of the building based on the 3D model, and fuse the point cloud distribution through spatial connection; S34: Use Gaussian mixture model combined with building point cloud to extract the window and wall ratio of individual buildings; S35: Use spatial correlation to fuse the data features obtained in steps S31 to S34 to form a feature vector.

3. The method for evaluating the potential of building-integrated photovoltaics in high-density urban areas according to claim 2, characterized in that: The specific steps in step S32 are as follows: S321: Perform neighborhood search, using the KDTree algorithm based on the binary tree principle to define the neighborhood information of spatial points by setting the neighborhood radius of the point cloud and the number of nearest neighbors; S322: Construct the covariance matrix. For each spatial point and its neighborhood point set, calculate the mean of the spatial vector of the neighborhood point set. Construct a covariance matrix based on the mean, and calculate eigenvalues ​​and eigenvectors using the covariance matrix; select the smallest eigenvalue and calculate the corresponding eigenvector, and set the direction away from the model as the normal direction, using it as the normal vector; obtain the normal vectors of all point clouds by traversing through the above steps.

4. The method for evaluating the potential of building-integrated photovoltaics in high-density urban areas according to claim 2, characterized in that: The specific steps in step S33 are as follows: S331: Set (0,1,0) as the true north direction vector, and set the direction of angle calculation as the direction of counterclockwise rotation from the true north direction vector to the normal vector on the XY plane, with the angle value range being [0,360); divide the angle into 16 partitions with 22.5° as the boundary; S332: Separate the facade surface of the 3D model and calculate the normal vector of the centroid of the facade surface; project the normal vector onto the XY plane, calculate the angle between the normal vector and the north direction vector, and the angle zone to which the angle belongs; S333: Aggregate the surface area of ​​the facade within the angle partition, filter out the angle partition with the largest area and calculate the average angle within that partition; The average angle is scaled to [-90, 90] to obtain the building's main orientation angle, which serves as an indicator of the building's geographical orientation angle.

5. The method for evaluating the potential of building-integrated photovoltaics in high-density urban areas according to claim 2, characterized in that: The specific steps in step S34 are as follows: S341: Separate the building point cloud by the building footprint outline, so that the building group is divided into independent buildings; S342: Extract RGB color and spatial coordinate values ​​from building point cloud; S343: Combine the RGB color of the same point with its spatial coordinate values; The combined data is clustered using a Gaussian mixture model to divide the building point cloud into window point cloud and non-window point cloud, and the area of ​​the building window elements is calculated accordingly.

6. The method for evaluating the potential of building-integrated photovoltaics in high-density urban areas according to claim 1, characterized in that: The specific steps in step S4 are as follows: S41: Apply corresponding feature scaling to different data features to reduce the data range to the interval [-1, 1]. S42: After inputting the data, feature transformations including T-Net and matrix multiplication operations are performed, specifically including orthogonal regularizers, dot product models, and multilayer perceptrons; nonlinear transformations are performed within the convolutional and dense layers of the regression prediction model; tanh and LeakyReLu are used as activation functions respectively. S43: Apply Transform and MLP twice to align features, increase dimensionality, and capture complex patterns and relationships in the data; use Global Max Pooling to aggregate all data features into a single global feature vector; S44: Transform and compress global feature vectors using MLP; A Dropout regularizer is set up to randomly drop some neurons to prevent overfitting; a linear activation function is used in the output layer to regress and predict the photovoltaic power generation potential.

7. The method for evaluating the potential of building-integrated photovoltaics in high-density urban areas according to claim 1, characterized in that: The specific steps in step S5 are as follows: S51: The radiation intensity of building surfaces is obtained by simulating solar radiation using the Perez scattering radiation model and ray tracing algorithm; S52: Calculate the photovoltaic power generation potential of photovoltaic devices deployed at different locations on a building based on radiation intensity; S53: Screen out the morphological factors that significantly affect the prediction and build a multiple linear regression model to represent the influence relationship between morphological factors and photovoltaic power generation potential.

8. The method for evaluating the potential of building-integrated photovoltaics in high-density urban areas according to claim 1, characterized in that: The specific steps in step S6 are as follows: S61: Use root mean square error and coefficient of determination to evaluate the performance of the model; root mean square error is used as an error rate metric for training the model, and root mean square error and coefficient of determination are used to evaluate the predictive performance of the trained model. S 62: The multiple linear regression model with urban morphology factors as independent variables and annual photovoltaic power generation potential as dependent variable is used as the reference model; stepwise regression method is used to screen significant explanatory variables.

9. A system for evaluating the potential of building-integrated photovoltaics in high-density urban areas, characterized in that: The point cloud submodule is used to acquire urban point clouds through surveying and mapping. The modeling submodule is used to separate building point clouds from urban point clouds using the RANSAC algorithm, and generate 3D models of buildings based on building footprints and building point clouds. The feature submodule is used to extract data features from the 3D model through point cloud sampling, normal estimation, geographic orientation calculation and window-wall scale calculation, and fuse them into a feature vector; The prediction submodule is used to define the point cloud data and spatial range, input the feature vector into the regression prediction model composed of a one-dimensional point cloud neural network, and output the predicted value of the photovoltaic power generation potential of the region's BIPV. The simulation submodule is used to obtain the radiation intensity of building surfaces through solar radiation simulation and calculate the assessment value of photovoltaic power generation potential. The evaluation submodule is used to evaluate the performance of regression prediction models using different evaluation metrics and reference models.

10. A computer memory, characterized in that: It contains a computer program that can be executed by a computer processor, the computer program performing the method for evaluating the building-integrated photovoltaic potential in high-density urban areas as described in any one of claims 1 to 8.

Images