Urban road and form integrated intelligent generation method based on heterogeneous graph network

Abstract

The invention discloses an urban road and form integrated intelligent generation method based on a heterogeneous graph network. The method comprises the following steps: collecting a target design site and road, plot and building data of a city where the target design site is located; selecting a planning background data range; constructing a heterogeneous graph structure adjacent matrix, and performing feature extraction and model pre-training to obtain a pre-trained heterogeneous graph neural network based on a road structure and an urban form; importing the planning background data into the pre-training model, and predicting road and form integrated generation of the target design site; planning background data is mapped to a potential space, road and plot morphological parameters are learned, the parameters are optimized in combination with a cross entropy loss function and a back propagation algorithm, and a road network and spatial morphological scheme matched with the plot features is generated; and finally, realizing visual display of the scheme through a UMAP method and a three-dimensional rendering engine, and outputting professional drawings and interactive three-dimensional models containing parameters such as road grades, land use functions, building density and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of urban planning, and specifically to an integrated intelligent generation method for urban roads and forms based on a heterogeneous graph network. Background Art

[0002] The collaborative design of urban road networks and spatial forms is a key link in urban planning, directly affecting urban traffic efficiency, functional layout, and spatial quality. Traditional design methods usually rely on manual experience and phased independent design models, and the road planning and design are often separated from the land use layout and building form design. Most existing technologies use two-dimensional drawings or simple parametric tools, making it difficult to effectively integrate multi-source heterogeneous data such as oblique photography and planning indicators, resulting in a low matching degree between site feature analysis and planning constraints. At the same time, the dynamic coupling relationships between elements such as road connectivity and plot function compatibility lack quantitative representation, and the generation of solutions is easily limited by subjective experience, making it difficult to achieve multi-objective collaborative optimization. In addition, traditional visualization means are mostly limited to static drawings or simplified models, lacking the ability for dynamic interactive analysis and multi-dimensional verification of attributes such as road alignment, traffic flow, and development intensity, restricting the scientific evaluation and rapid iteration of solutions. How to break through the data barrier, establish an intelligent association model for spatial elements, and achieve integrated dynamic generation has become an important challenge for improving the efficiency and scientificity of urban planning and design. Summary of the Invention

[0003] Aiming at the deficiencies of the existing technology, the present invention provides an integrated intelligent generation method for urban roads and forms based on a heterogeneous graph network, which solves the problems of the separated design of road networks and spatial forms and the difficulty in quantitatively mining the coupling relationships of spatial elements in traditional methods.

[0004] To achieve the above objectives, the present invention is realized through the following technical solutions:

[0005] An integrated intelligent generation method for urban roads and forms based on a heterogeneous graph network, comprising the following steps:

[0006] (1) Collect oblique photography data of the target site through a drone aerial photography device, and extract the site information of the road network and building form to be generated; retrieve the regulatory detailed planning data and urban design scheme of the local planning bureau through a data interface, and extract the road data, land use data, and building data of the target design site in the city.

[0007] (2) Expand the planning scope of the target design site in the site information of the road network and building form to be generated obtained in step (1) by 1 kilometer as the selection range of the planning background data, and intercept the road data, land use data, and building data of the target design site in the city obtained in step (1) as the planning background data.

[0008] (3) Establish a heterogeneous graph neural network based on road structure and urban form; First, according to the objectives in step (1), design the road data, land use data, and building data of the city where the target design site is located to obtain the spatial positions of the urban roads and plots in the city where the target design site is located, and construct the adjacency matrix of the heterogeneous graph structure of the urban roads and plots in the city where the target design site is located; Second, extract features from the road data, land use data, and building data of the city where the target design site is located, and use the extracted features as the input features of the graph neural network; Finally, use the graph convolutional neural network (GCN), train the heterogeneous graph neural network model through multiple convolutional layers, and optimize the network parameters according to the training results to improve the model prediction ability, and finally obtain the pre-trained heterogeneous graph neural network model;

[0009] (4) Convert the planning background data obtained in step (2) into a graph data structure through the adjacency matrix of the heterogeneous graph structure in step (3), standardize and preprocess it, import the pre-trained heterogeneous graph neural network model obtained in step (3), learn the structure and features of the graph, initially learn the coupling relationship between the road and the spatial form integration, and predict information such as the road distribution, building form, and plot layout of the target site, and predict and generate the heterogeneous graph network of the target design site;

[0010] (5) Realize the integrated intelligent generation of the road and form of the target design site based on the graph variational autoencoder; First, learn the road and plot shape parameters of the target design site based on the planning background data, and further combine the shape parameters with the heterogeneous graph network of the target design site obtained in step (4), and through the mapping and regeneration of the feature data of the target design site, realize the generation of the road and form scheme of the target design site; And in the process, measure the difference between the generated data and the input data by calculating the cross-entropy loss; Through the backpropagation algorithm, optimize the model parameters to minimize the loss, and then output the integrated scheme of the optimized road and form of the target design site;

[0011] (6) Establish a scheme visualization platform, connect to the scheme output system, and perform visualization processing on the road and form layout data of the target design site generated by the UMAP method; The visualization platform will express the generated urban road and spatial form scheme and data in professional drawings, and display the intuitive spatial effect through a 3D model combined with augmented reality and virtual reality technologies; The generated visualization scheme includes detailed information such as road grade, road length, road width, traffic flow, land use form, land use function, land use floor area ratio, land use height limit, and building density, which is convenient to evaluate the spatial effect and rationality of the scheme from multiple angles.

[0012] Preferably, the site information for generating the road network and building form in step (1) includes: the spatial location of the site, the planning scope of the site, and the properties of existing buildings, roads, facilities, green spaces, and water bodies within the site; and the road data, land use data, and building data of the city where the target design site is located include: road grade, road length, road width, traffic flow, land form, land function, land volume ratio, land height limit, and building density, a total of 9 properties.

[0013] Table 1 Road, land and building data of the city where the target design site is located

[0014]

[0015] Preferably, the method for obtaining the planning background data in step (2) is to refer to the "Urban Residential Area Planning and Design Standard (GB50180-2018)" to expand the planning range of the target design site in the site information of the road network and building form to be generated in step (1) outward by 1 kilometer, and use it as the selection range of the planning background data. Then, the road data, land use data and building data of the city where the target design site obtained in step (1) is located are intercepted with the selection range of the planning background data to obtain the planning background data.

[0016] Preferably, the method for constructing the heterogeneous graph structure of urban roads and plots in step (3) is to extract the spatial location information of urban roads and plots in the city where the target design site is located based on the road data, land use data and building data of the city where the target design site is located in step (1), and thereby construct a heterogeneous graph structure adjacency matrix of urban roads and plots in the city where the target design site is located, wherein: nodes include road nodes representing road intersections and key road points, and plot nodes representing different land use units; edges are used to connect road nodes to represent road connectivity, connect plot nodes to represent plot proximity, and connect road nodes and plot nodes to represent road and plot proximity.

[0017] Preferably, in step (3), a pre-trained heterogeneous graph neural network model is constructed, and for each road node v r Extract feature vector h r , h r =[road_type,length,width,traffic_volume], where v r Represents a single road node, road_type, length, width, and traffic_volume correspond to the road grade, road length, road width, and traffic volume in the road data of the city where the target design site is located. l Extract feature vector h l , where vl Represents a single plot node, h l = [landuse, FAR, height, density], where landuse, FAR, height, and density correspond to the land use function, land use floor area ratio, land height limit, and building density in the land use data and building data of the city where the target design site is located, respectively; connect all features to form a feature matrix H, H = [h1, h2, …, h r , …, h l T , where T represents that the internal vectors form a matrix;

[0018] Then, a graph convolutional neural network is used to perform feature propagation and training on the heterogeneous graph structure of the urban roads and plots in the city where the target design site is located obtained in step (3); the propagation rule of the graph convolutional neural network is:

[0019]

[0020] Among them, H (l) is the feature matrix of the l-th layer, H (l+1) is the feature matrix of the next layer associated with the l-th layer, [[ID=2,4]]is the normalized adjacency matrix, W (l) is the trainable weight matrix, σ is the activation function ReLU;

[0021] Finally, semi-supervised learning is used, and the cross-entropy is used to calculate the loss function, and iterative training is performed until the loss converges, and finally a pre-trained heterogeneous graph neural network model is obtained. The loss function is:

[0022]

[0023] Among them, y i is the actual value of the i-th sample, is the predicted value output by the model of the i-th sample.

[0024] Preferably, the graph variational autoencoder (GraphVAE) in step (5) realizes the integrated intelligent generation of the roads and forms of the target design site, including an encoding process and a decoding process;

[0025] In the encoding process, the input data first includes the road and plot form data in the planning background data. According to the encoder, the road and plot shape parameters are mapped to the latent space matrix, the mean and variance of the latent variables are generated, and the latent variables are regularized to avoid the distribution of the latent variables being too complex or overfitting, and a point is sampled from the distribution of the latent variables using reparameterization;

[0026] ​In the decoding process, first, the road and plot form parameters learned from the planning background information are input into the decoder. According to the training results of the heterogeneous graph neural network model of the target site, the parameters are remapped within the spatial range of the target site. While preserving the topological relationship of the input data, the road alignment, plot scale, and spatial structure are dynamically adjusted. By calculating the cross-entropy loss, the difference between the decoder output and the original form data is measured. The greater the cross-entropy loss, the larger the gradient needs to be calculated using the backpropagation algorithm, and the network weights are adjusted to a greater extent to accelerate the reduction of the difference between the predicted output and the target data, so as to realize the iterative optimization of the model parameters. Through this form parameter optimization mechanism, adaptive adjustment can be achieved according to the site characteristics and planning requirements, and the adaptive generation of roads, plots, and buildings within the target site can be realized.

[0027] Preferably, in step (5), when calculating the cross-entropy loss, during the model training process, the difference between the prediction result and the planning background data is quantified through the cross-entropy loss function; the road grade, road length, road width, traffic flow, land use form, land use function, land use floor area ratio, land use height limit, building density output by the heterogeneous graph network of the target design site in step (4) are feature-aligned with the planning background data; the cross-entropy loss based on node features and edge features is calculated according to the following formula, and the formula is as follows:

[0028]

[0029] where N is the total number of nodes in the graph; C is the set of node feature categories; y i,ε is the true feature value of node i; is the predicted feature value of node i;

[0030]

[0031] where E is the total number of edges in the graph; is the set of node feature categories; y j,k is the true feature value of node j; is the predicted feature value of node j; k is a certain feature category of node j;

[0032] Then, calculate the cross-entropy loss of the predicted graph structure and the input graph structure in terms of topological association and attribute distribution, and the formula is as follows:

[0033]

[0034] where, N is the total number of nodes in the graph structure; A i,j is the connection relationship between node i and node u in the true adjacency matrix; A i,u is the connection probability between node i and node u in the predicted adjacency matrix; is a set of connection probabilities between node i and node u in the predicted adjacency matrix;

[0035] Combined with the KL divergence regularization term, calculate the overall graph structure cross-entropy loss, optimize the latent space distribution, and prevent the model from overfitting. The formula is as follows:

[0036] L 总和 = αL 图结构 + βL 节点 + γL 边 + δL 正则化项

[0037] Among them, α, β, and γ are weight coefficients used to balance the contributions of node, edge, and graph structure losses; L 图结构 is the cross-entropy loss of the graph structure; L 节点 is the cross-entropy loss of the graph structure nodes; L 边 is the cross-entropy loss of the graph structure edges; L 正则化项 is the regularization term; δ is the weight coefficient of the regularization term.

[0038] Preferably, the backpropagation algorithm in step (5); calculate the prediction errors of node features, edge features, and topological structure according to the cross-entropy loss function, and use the chain rule of differentiation to calculate the weight parameter gradients of each neural network layer layer by layer from the decoder to the encoder, clarify the contribution degree of the weight parameters of each neural network layer to the total error, and then guide the update direction of the heterogeneous graph neural network model parameters of the target design site. If the weight parameter gradient is a positive gradient function, the weight needs to be reduced. Conversely, increase its weight. The formula is as follows:

[0039]

[0040] Among them, l is the loss function, W l is the weight matrix of the l-th layer, represents the derivative of a certain variable in a multivariate function, represents the partial derivative of the loss function L with respect to the weight parameter, that is, only considering the influence of the change on, while other variables remain unchanged; is the normalized adjacency matrix; H l is the feature matrix of the l-th layer; T represents the transpose of the matrix, that is, the rows and columns of the original matrix are interchanged; δ l+1 is the error gradient matrix of the (l + 1)-th layer;

[0041] Introduce the Adam optimizer to calculate the smoothed historical gradient direction of the first moment momentum value, avoid frequent changes in the gradient direction due to the dynamics of road and plot data, and at the same time ensure that the data requires the gradient to remain consistent for a long time due to strong standardization, so as to strengthen the continuity of the planning intention and accelerate the convergence of the model. The specific formula is as follows:

[0042] m t= β1·m t-1 + (1 - β1)·g t

[0043] Where m t is the updated momentum value, t is the t-th iteration, t - 1 is the number of the (t - 1)-th iteration, m t-1 is the momentum value at the (t - 1)-th iteration, g t is the gradient of the current step, β1 is to control the decay rate of the historical gradient;

[0044] According to the momentum calculation result, calculate the updated weight parameters for the next iteration of each layer of the neural network, and perform iterative calculations to continuously optimize the morphological parameters of the heterogeneous graph neural network generated by the target site design, forming an optimized road and space form scheme. The specific formula is as follows:

[0045]

[0046] Where W t+1 is the updated weight parameter at the (t + 1)-th iteration, W t is the weight parameter at the t-th iteration, τ is the adaptive learning rate term, ρ is the adaptive learning rate term, ∈ is a constant to prevent division by zero error when the denominator ρ approaches zero;

[0047] At the same time, limit the gradient update range through gradient clipping technology, constrain the gradient amplitude within a preset threshold to avoid gradient explosion or gradient disappearance problems during training; optimize the semantic mapping ability of the heterogeneous graph neural network of the target design site in the latent space by iteratively updating the weight matrices of the encoder and decoder, ensuring that the generated road alignment and plot form parameters match the topological features and semantic consistency of the planning background data.

[0048] Preferably, the UMAP method in step (6) is as follows: Input the node features and edge features in the target design field plan generated by the heterogeneous graph network into the UMAP algorithm. By calculating the similarity of nodes in the high-dimensional space and the similarity of low-dimensional space coordinates, and by minimizing the difference between the two similarities, the low-dimensional space coordinate relationship is made infinitely close to the high-dimensional space structure relationship, generating spatial coordinates. The coordinates retain the local and global structure features of the original high-dimensional data, providing a geometric basis for 3D model construction. The specific formula is as follows:

[0049]

[0050] Where p ij is the high-dimensional space similarity value, d(x i , x j ) is the Euclidean distance between node i and node j, μ iis the distance from node i to its nearest neighbor node, σ i is the scale parameter;

[0051] Randomly assign initial three-dimensional coordinates (x, y, z) to each node, calculate the distances between nodes in the three-dimensional space using the initial three-dimensional coordinates, and calculate the similarity of the low-dimensional space coordinates. The specific formulas are as follows:

[0052]

[0053] where q ij is the low-dimensional space similarity value, a and b are hyperparameters that control the distribution shape, and di,j is the coordinate distance between node i and node j;

[0054]

[0055] where M is the minimum value of the high-dimensional space similarity and the low-dimensional space similarity;

[0056] Continuously iterate the initial three-dimensional coordinates assigned to each node through the calculation result of the M value until the high-dimensional space similarity and the low-dimensional space similarity are nearly the same. Output the three-dimensional coordinates of each node, and use the three-dimensional point cloud rendering technology to generate an interactive virtual scene model.

[0057] Beneficial effects:

[0058] (1) The urban road and form integrated intelligent generation method based on the heterogeneous graph network of the present invention integrates oblique photography data, regulatory planning data, and urban design plans through the dual-source data acquisition technology of drone aerial photography and planning data interface, constructs a multi-dimensional site information dataset, and realizes the deep integration of geographical spatial information and planning constraint conditions, significantly improving the integrity of data coverage and the accuracy of feature extraction.

[0059] (2) The urban road and form integrated intelligent generation method based on the heterogeneous graph network of the present invention constructs a modeling method for the adjacency matrix of the road-lot heterogeneous graph structure. Through the topological connection of road nodes, lot nodes, and three types of edge relationships, it completely represents the multi-scale coupling relationship of urban spatial elements, and combines the multi-layer feature propagation mechanism of the GCN graph convolutional network to effectively mine the implicit association rules between road grades - land use functions - development intensities.

[0060] (3) The urban road and form integrated intelligent generation method based on the heterogeneous graph network of the present invention uses the Graph Variational Autoencoder (GraphVAE) to realize the collaborative generation of road networks and spatial forms. Through the potential space parameter mapping and feature reconstruction technology, under the constraints of specifications such as the "Urban Residential Area Planning and Design Standard", it automatically generates a scheme with a reasonable topological structure and meeting the requirements of land use compatibility, supporting multi-modal scheme comparison and interactive optimization.

[0061] (4) The present invention is based on an intelligent method for the integrated generation of urban roads and morphology based on heterogeneous graph networks, constructs a three-level loss function system of node-edge-graph structure, calculates cross entropy based on nine types of node features such as road grade and land use volume ratio and three types of edge features such as road connectivity, and combines it with KL divergence regularization constraints to achieve joint optimization of road alignment, land parcel morphology and planning control indicators, ensuring that the generated plan meets the requirements of both traffic organization efficiency and spatial morphology rationality.

[0062] (5) The present invention is based on an intelligent method for integrating urban roads and forms in heterogeneous graph networks, integrating the UMAP dimensionality reduction algorithm and visualization engine, and converting vector features such as road width and building height in heterogeneous graph networks into interactive three-dimensional spatial models. It supports dynamic adjustment of planning indicators and real-time detection of spatial conflicts, and achieves seamless superposition of scheme data and real-life space through augmented reality technology, significantly improving the immersiveness of scheme review and the scientific nature of decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 Flow chart of the method of the present invention;

[0064] Figure 2 This is the logic for generating the road and space solutions for the target design site in the embodiment of the present invention. DETAILED DESCRIPTION

[0065] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0066] Example

[0067] like Figure 1 and Figure 2 As shown, the technical solution of the present invention will be described in detail below by taking the urban design of a certain area in a certain city as an example.

[0068] (1) Collect road, land, and building data for the target design site and the city where it is located, including:

[0069] (1.1) Oblique photography data of the target site is collected using drones, and site information for the road network and building forms to be generated is extracted, including the spatial location of the site, the planned scope of the site, and the properties of existing buildings, roads, facilities, green spaces, and water bodies within the site;

[0070] (1.2) Retrieve the regulatory detailed planning data and urban design plans of the local planning bureau through the data interface, and extract the road data, land use data, and building data of the city where the target design site is located, specifically including: road grade, road length, road width, traffic flow, land use form, land use function, land use plot ratio, land use height limit, and building density, a total of 9 attributes;

[0071] Table 1: Road, land use, and building data of the city where the target design site is located

[0072]

[0073] (2) Expand 1 kilometer from the planning scope of the target design site in the site information of the road network and building form to be generated obtained in step (1) as the selection range of the planning background data, and intercept the road data, land use data, and building data of the city where the target design site is located obtained in step (1) as the planning background data, specifically including: [[ID=...]]

[0074] (2.1) Refer to the "Urban Residential Area Planning and Design Standard (GB50180 - 2018)" to expand the planning scope of the target design site in the site information of the road network and building form to be generated obtained in step (1) by 1 kilometer, and use it as the selection range of the planning background data;

[0075] (2.2) Intercept the road data, land use data, and building data of the city where the target design site is located obtained in step (1) with the selection range of the planning background data to obtain the planning background data;

[0076] (3) Establish a heterogeneous graph neural network based on road structure and urban form, specifically including:

[0077] (3.1) Obtain the spatial positions of the urban roads and plots in the city where the target design site is located according to the road data, land use data, and building data of the city where the target design site is located in step (1).

[0078] (3.2) Construct the adjacency matrix of the heterogeneous graph structure of the urban roads and plots in the city where the target design site is located. The nodes in the adjacency matrix of the heterogeneous graph structure include road nodes representing road intersections and key road points, and plot nodes representing different land use units; the edges are used to connect between road nodes, indicating road connectivity, connect between plot nodes, indicating plot proximity relationships, and connect between road nodes and plot nodes, indicating the proximity relationship between roads and plots.

[0079] (3.3) For the pre-trained heterogeneous graph neural network model constructed in step (3), for each road node v r Extract the feature vector h r , h r= [road_type, length, width, traffic_volume], where v r represents a single road node, and road_type, length, width, traffic_volume correspond to the road grade, road length, road width, and traffic volume in the road data of the city where the target design site is located in sequence; for each plot node v l extract the feature vector h l , where v l represents a single plot node, and h l = [landuse, FAR, height, density], and landuse, FAR, height, density correspond to the land use function, land use floor area ratio, land use height limit, and building density in the land use data and building data of the city where the target design site is located in sequence; connect all features to form a feature matrix H, H = [h1, h2, …, h r , …, h l T , where T represents that the internal vectors form a matrix form.

[0080] (3.4) Use a graph convolutional neural network to perform feature propagation and training on the heterogeneous graph structure of the urban roads and plots in the city where the target design site is obtained in step (3); the propagation rule of the graph convolutional neural network is:

[0081]

[0082] Among them, H (l) is the feature matrix of the l-th layer, H (l+1) is the next-layer feature matrix associated with the l-th layer, is the normalized adjacency matrix, W (l) is the trainable weight matrix, and σ is the activation function ReLU;

[0083] Finally, use semi-supervised learning, calculate the loss function using cross-entropy, and iteratively train until the loss converges to finally obtain a pre-trained heterogeneous graph neural network model. The loss function is:

[0084]

[0085] Among them, y i is the actual value of the i-th sample, is the predicted value output by the model for the i-th sample.

[0086] (4) Import the planning background data into the pre-trained model to predict the road and form heterogeneous graph network of the target design site.

[0087] ​(4.1) Convert the planned background data obtained in step (2) into a graph data structure through the heterogeneous graph structure adjacency matrix in step (3), and perform standardization and preprocessing on it.

[0088] (4.2) Import the pre-trained heterogeneous graph neural network model obtained in step (3), learn the structure and features of the graph, initially learn the coupling relationship between the road and the spatial form integration, and predict information such as the road distribution, building form, and plot layout of the target site, and predict and generate the heterogeneous graph network of the target design site.

[0089] (5) Realize the integrated intelligent generation of the road and form of the target design site. Specifically, it includes:

[0090] (5.1) Learn the road and plot shape parameters of the target design site based on the planned background data.

[0091] (5.2) Further combine the shape parameters with the heterogeneous graph network of the target design site obtained in step (4), and realize the generation of the road and form scheme of the target design site through the mapping and regeneration of the feature data of the target design site.

[0092] (5.3) By calculating the cross-entropy loss, during the model training process, quantify the difference between the prediction result and the planned background data through the cross-entropy loss function; align the road grade, road length, road width, traffic flow, land use form, land use function, land use floor area ratio, land use height limit, building density output by the heterogeneous graph network of the target design site in step (4) with the planned background data; calculate the cross-entropy loss based on node features and edge features according to the following formula, and the formula is as follows:

[0093]

[0094] Among them, N is the total number of nodes in the graph; C is the set of node feature categories; y i,ε is the true feature value of node i; is the predicted feature value of node i;

[0095]

[0096] Among them, E is the total number of edges in the graph; is the set of node feature categories; y j,k is the true feature value of node j; is the predicted feature value of node j; k is a certain feature category of node j;

[0097] Then calculate the cross-entropy loss between the predicted graph structure and the input graph structure in terms of topological association and attribute distribution, and the formula is as follows:

[0098]

[0099] Among them, N is the total number of nodes in the graph structure; A i,j is the connection relationship between node i and node u in the real adjacency matrix; A i,u is the connection probability between node i and node u in the predicted adjacency matrix; is the set of connection probabilities between node i and node u in the predicted adjacency matrix;

[0100] Combined with the KL divergence regularization term, calculate the overall graph structure cross-entropy loss, optimize the latent space distribution, and prevent the model from overfitting. The formula is as follows:

[0101] L 总和 = αL 图结构 + βL 节点 + γL 边 + δL 正则化项

[0102] Among them, α, β, and γ are weight coefficients used to balance the contributions of node, edge, and graph structure losses; L 图结构 is the cross-entropy loss of the graph structure; L 节点 is the cross-entropy loss of the graph structure nodes; L 边 is the cross-entropy loss of the graph structure edges; L 正则化项 is the regularization term; δ is the weight coefficient of the regularization term.

[0103] (5.4) Backpropagation algorithm; calculate the prediction errors of node features, edge features, and topological structure according to the cross-entropy loss function, and use the chain rule of differentiation to calculate the weight parameter gradients of each neural network layer layer by layer from the decoder to the encoder, clarify the contribution degree of the weight parameters of each neural network layer to the total error, and then guide the update direction of the heterogeneous graph neural network model parameters of the target design site. If the weight parameter gradient is a positive gradient function, the weight needs to be reduced. Conversely, increase its weight. The formula is as follows:

[0104]

[0105] Among them, L is the loss function, W l is the weight matrix of the l-th layer, represents the derivative of a certain variable in a multivariate function, represents the partial derivative of the loss function L with respect to the weight parameter, that is, only considering the influence of the change on, while other variables remain unchanged; is the normalized adjacency matrix; H l is the feature matrix of the l-th layer; T represents the transpose of the matrix, that is, the rows and columns of the original matrix are interchanged; δ l+1 is the error gradient matrix of the (l + 1)-th layer;

[0106] The Adam optimizer is introduced to calculate the smoothed historical gradient direction of the first-order moment momentum value, avoiding frequent changes in the gradient direction due to the dynamics of road and plot data. At the same time, it ensures that the data requires the gradient to maintain consistency for a long time due to strong normalization to strengthen the continuity of the planning intention and accelerate model convergence. The specific formula is as follows:

[0107] m t = β1·m t-1 +(1 - β1)·g t

[0108] Among them, m t is the updated momentum value, t is the t-th iteration, t - 1 is the number of the (t - 1)-th iteration, m t-1 is the momentum value of the (t - 1)-th iteration, g t is the gradient of the current step, β1 is to control the decay rate of the historical gradient;

[0109] According to the calculation result of the momentum, calculate the updated weight parameters for the next iteration of each layer of the neural network, and calculate iteratively to continuously optimize the morphological parameters of the heterogeneous graph neural network generated by the target site design, forming an optimized road and spatial form scheme. The specific formula is as follows:

[0110]

[0111] Among them, W t+1 is the updated weight parameter for the (t + 1)-th iteration, W t is the weight parameter for the t-th iteration, τ is the adaptive learning rate term, ρ is the adaptive learning rate term, and ∈ is a constant to prevent division by zero error when ρ approaches zero in the denominator.

[0112] (5.6) Limit the gradient update range through the gradient clipping technique, and constrain the gradient amplitude within a preset threshold to avoid the problems of gradient explosion or gradient disappearance during the training process;

[0113] (5.7) Optimize the semantic mapping ability of the heterogeneous graph neural network of the target design site in the latent space by iteratively updating the weight matrices of the encoder and decoder, ensuring that the generated road alignment and plot morphological parameters match the topological features and semantic consistency of the planning background data.

[0114] (6) Output and visualize the integrated generation scheme of urban roads and spatial forms. Specifically, it includes:

[0115] (6.1) The solution visualization platform interfaces with the solution output system and processes the road and morphological layout data of the target design site generated through visualization using the UMAP method. The node features and edge features in the target design site solution generated by the heterogeneous graph network are input into the UMAP algorithm. By calculating the similarity of nodes in the high-dimensional space and the similarity of low-dimensional space coordinates, and minimizing the difference between the two similarities, the coordinate relationship in the low-dimensional space is made to approach the structural relationship in the high-dimensional space infinitely, generating spatial coordinates. The coordinates retain the local and global structural features of the original high-dimensional data, providing a geometric basis for the construction of the 3D model. The specific formula is as follows:

[0116]

[0117] Where p ij is the high-dimensional space similarity value, d(x i , x j ) is the Euclidean distance between node i and node j, μ i is the distance from node i to its nearest neighbor node, and σ i is the scale parameter.

[0118] Randomly assign initial 3D coordinates (x, y, z) to each node, calculate the distance between nodes in the 3D space using the initial 3D coordinates, and calculate the similarity of the low-dimensional space coordinates. The specific formula is as follows:

[0119]

[0120] Where q ij is the low-dimensional space similarity value, a and b are hyperparameters that control the distribution shape, and di,j is the coordinate distance between node i and node j.

[0121]

[0122] Where M is the minimized value of the high-dimensional space similarity and the low-dimensional space similarity.

[0123] Continuously iterate the initial 3D coordinates assigned to each node based on the calculation result of the M value until the high-dimensional space similarity and the low-dimensional space similarity are nearly the same. Output the 3D space coordinates of each node, and use the 3D point cloud rendering technology to generate an interactive virtual scene model.

[0124] (6.2) The visualization platform makes a professional drawing expression of the generated urban road and spatial form solutions and data, and demonstrates an intuitive spatial effect through the 3D model combined with augmented reality and virtual reality technologies.

[0125] (6.3) The generated visualization solutions include details of road grades, road lengths, road widths, traffic flows, land use patterns, land use functions, land use plot ratios, land use height limits, and building densities, facilitating the evaluation of the spatial effects and rationality of the solutions from multiple perspectives.

[0126] The above shows and describes 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 by the above embodiments. What is described in the above embodiments and the specification only illustrates the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and all these changes and improvements fall within the scope of the present invention claimed.

Claims

1. An integrated intelligent generation method for urban roads and forms based on heterogeneous graph networks, characterized in that It includes the following steps: (1) Collect the oblique photography data of the target site through a drone aerial photography device, and extract the site information of the road network and building form to be generated; Retrieve the regulatory detailed planning data and urban design plan of the local planning bureau through the data interface, and extract the road data, land use data, and building data of the city where the target design site is located; (2) Expand the planning scope of the target design site in the site information of the road network and building form obtained in step (1) by 1 km as the selection range of the planning background data, and intercept the road data, land use data, and building data of the city where the target design site is located obtained in step (1) as the planning background data; (3) Establish a heterogeneous graph neural network based on the road structure and urban form; First, obtain the spatial positions of the urban roads and plots in the city where the target design site is located according to the road data, land use data, and building data of the city where the target design site is located in step (1), and construct an adjacency matrix of the heterogeneous graph structure of the urban roads and plots in the city where the target design site is located; Secondly, extract the features of the road data, land use data, and building data of the city where the target design site is located, and use the extracted features as the input features of the graph neural network; Finally, use a graph convolutional neural network to train the heterogeneous graph neural network model through multiple convolutional layers, and optimize the network parameters according to the training results to improve the model prediction ability, and finally obtain a pre-trained heterogeneous graph neural network model; (4) Convert the information of the planning background data obtained in step (2) into a graph data structure through the adjacency matrix of the heterogeneous graph structure in step (3), and standardize and preprocess it, and import the pre-trained heterogeneous graph neural network model obtained in step (3) to learn the structure and features of the graph, initially learn the coupling relationship between the road and the spatial form integration, and predict the road distribution, building form, plot layout, etc. of the target site, and predict and generate the heterogeneous graph network of the target design site; (5) Realize the integrated intelligent generation of the road and form of the target design site based on the graph variational autoencoder; First, learn the road and plot shape parameters of the target design site based on the planning background data, and further combine the shape parameters with the heterogeneous graph network of the target design site obtained in step (4) to realize the generation of the road and form scheme of the target design site through the mapping and regeneration of the feature data of the target design site; And calculate the cross-entropy loss during the process to measure the difference between the generated data and the input data; Through the backpropagation algorithm, optimize the model parameters to minimize the loss, and then output the integrated scheme for optimizing the road and form of the target design site. (6) Establish a scheme visualization platform, connect it to the scheme output system, and obtain the road and morphological layout data of the target design site generated through visualization processing by the UMAP method; the visualization platform will express the generated urban road and spatial form scheme and data in professional drawings, and display the intuitive spatial effect through a 3D model combined with augmented reality and virtual reality technologies; the generated visualization scheme includes details such as road grade, road length, road width, traffic flow, land use form, land use function, land use plot ratio, land use height limit, and building density, which is convenient for evaluating the spatial effect and rationality of the scheme from multiple perspectives.

2. The integrated intelligent generation method for urban roads and forms based on the heterogeneous graph network according to claim 1, characterized in that The site information for generating the road network and building form in step (1) includes: the spatial location of the site, the planning scope of the site, the existing buildings, roads, facilities, green spaces, and water body attributes within the site; the road data, land use data, and building data of the city where the target design site is located include 9 attributes: road grade, road length, road width, traffic flow, land use form, land use function, land use plot ratio, land use height limit, and building density.

3. The method for integrally and intelligently generating urban roads and forms based on a heterogeneous graph network according to claim 2, wherein, The method for obtaining the planning background data in step (2) is as follows: Refer to the "Urban Residential Area Planning and Design Standard (GB50180-2018)" to expand the planning scope of the target design site in the site information for generating the road network and building form in step (1) outward by 1 kilometer, and use it as the selection range of the planning background data. Then, intercept the road data, land use data, and building data of the city where the target design site is located obtained in step (1) with the selection range of the planning background data to obtain the planning background data.

4. The method for integrally and intelligently generating urban roads and forms based on a heterogeneous graph network according to claim 3, wherein, The method for constructing the heterogeneous graph structure of urban roads and plots in step (3) is as follows: Based on the road data, land use data, and building data of the city where the target design site is located in step (1), extract the spatial location information of urban roads and plots in the city where the target design site is located, and construct the adjacency matrix of the heterogeneous graph structure of urban roads and plots in the city where the target design site is located. Among them: the nodes include road nodes representing road intersections and key road points, and plot nodes representing different land use units; the edges are used to connect road nodes to represent road connectivity, connect plot nodes to represent plot adjacency relationships, and connect road nodes and plot nodes to represent the adjacency relationship between roads and plots.

5. The integrated intelligent generation method for urban roads and forms based on heterogeneous graph network according to claim 3, characterized in that, In step (3), after the pre-trained heterogeneous graph neural network model is constructed, for each road node v r the feature vector h r is extracted, where h r = [road_type, length, width, traffic_volume], and v r represents a single road node. road_type, length, width, and traffic_volume respectively correspond to the road grade, road length, road width, and traffic flow in the road data of the city where the target design site is located; for each plot node v l the feature vector h l is extracted, where v l represents a single plot node, and h l = [landuse, FAR, height, density]. landuse, FAR, height, and density respectively correspond to the land use function, land use floor area ratio, land use height limit, and building density in the land use data and building data of the city where the target design site is located; all the features are connected to form a feature matrix H, H = [h1, h2, …, h r , …, h l T , where T represents the form in which the internal vectors form a matrix;​ Then, use a graph convolutional neural network to perform feature propagation and training on the heterogeneous graph structure of urban roads and plots in the city where the target design site is located obtained in step (3); the propagation rule of the graph convolutional neural network is: Among them, H (l) is the feature matrix of the l-th layer, and H (l+1) is the feature matrix of the next layer associated with the l-th layer, is the normalized adjacency matrix, W (l) is the trainable weight matrix, and σ is the activation function ReLU; Finally, use semi-supervised learning, calculate the loss function using cross-entropy, and iteratively train until the loss converges. Finally, obtain the pre-trained heterogeneous graph neural network model. The loss function is: where y i is the actual value of the i-th sample, and is the predicted value output by the i-th sample model.

6. The method for integrally and intelligently generating urban roads and forms based on a heterogeneous graph network according to claim 5, characterized in that: The graph variational autoencoder in step (5) realizes the integrated intelligent generation of the road and form of the target design site, including an encoding process and a decoding process; In the encoding process, the input data first includes the road and plot form data in the planned background data in step (2). According to the encoder, the road and plot shape parameters are mapped to the latent space matrix to generate the mean and variance of the latent variables, and the latent variables are regularized to avoid the distribution of the latent variables being too complex or overfitting. The reparameterization is used to sample a point from the distribution of the latent variables. In the decoding process, the road and plot form parameters learned from the planned background information are first input into the decoder. According to the training results of the heterogeneous graph neural network model of the target site, the parameters are remapped to the target site space range. On the basis of retaining the topological relationship of the input data, the road orientation, plot scale and spatial structure are dynamically adjusted. By calculating the cross-entropy loss, the difference between the decoder output and the original form data is measured. The greater the cross-entropy loss, the greater the gradient needs to be calculated using the backpropagation algorithm, and the greater the adjustment of the network weights, so as to accelerate the reduction of the difference between the predicted output and the target data, so as to realize the iterative optimization of the model parameters. Through this form parameter optimization mechanism, adaptive adjustment is realized according to the site characteristics and planning requirements, and the adaptive generation of roads, plots and buildings in the target site is realized.

7. The method for integrally and intelligently generating urban roads and forms based on a heterogeneous graph network according to claim 5, wherein: In step (5), the cross-entropy loss is calculated. During the model training process, the difference between the prediction result and the planned background data is quantified through the cross-entropy loss function; the road grade, road length, road width, traffic flow, land use form, land use function, land use plot ratio, land use height limit, building density output by the heterogeneous graph network of the target design site in step (4) are feature-aligned with the planned background data; the cross-entropy loss based on node features and edge features is calculated according to the following formula. The formula is as follows: where N is the total number of nodes in the graph; C is the set of categories of node features; y i,ε is the true feature value of node i; is the predicted feature value of node i; Among them, E is the total number of edges in the graph; is the set of categories of node features; y j,k is the true feature value of node j; is the predicted feature value of node j; k is a certain feature category of node j; Then calculate the cross-entropy loss between the predicted graph structure and the input graph structure in terms of topological association and attribute distribution. The formula is as follows: where N is the total number of nodes in the graph structure; A i,j is the connection relationship between node i and node u in the true adjacency matrix; A i,u is the connection probability between node i and node u in the predicted adjacency matrix; is the set of connection probabilities between node i and node u in the predicted adjacency matrix; Combined with the KL divergence regularization term, calculate the overall graph structure cross-entropy loss to optimize the latent space distribution and prevent the model from overfitting. The formula is as follows: L 总和 = αL 图结构 + βL 节点 + γL 边 + δγ 正则化项 Among them, α, β, and γ are weight coefficients used to balance the contributions of node, edge, and graph structure losses; L 图结构 is the cross-entropy loss of the graph structure; L 节点 is the cross-entropy loss of the graph structure nodes; L 边 is the cross-entropy loss of the graph structure edges; L 正则化项 is the regularization term; δ is the weight coefficient of the regularization term.

8. The method for integrally and intelligently generating urban roads and forms based on a heterogeneous graph network according to claim 5, wherein: In step (5), the backpropagation algorithm; calculate the prediction errors of node features, edge features and topological structure according to the cross-entropy loss function, and use the chain rule of differentiation to calculate the weight parameter gradients of each neural network layer layer by layer from the decoder to the encoder, clarify the contribution degree of the weight parameters of each neural network layer to the total error, and then guide the parameter update direction of the heterogeneous graph neural network model of the target design site. If the weight parameter gradient is a positive gradient function, the weight needs to be reduced, otherwise, increase its weight. The formula is as follows: Among them, L is the loss function, and W l is the weight matrix of the l-th layer, represents the derivative of a certain variable in a multivariate function, represents the partial derivative of the loss function L with respect to the weight parameter, that is, only considering the influence of the change on, while other variables remain unchanged; is the normalized adjacency matrix; H l is the feature matrix of the l-th layer; T represents the transpose of the matrix, that is, the rows and columns of the original matrix are interchanged; δ l+1 is the error gradient matrix of the (l + 1)-th layer; Introduce the Adam optimizer to calculate the smoothed historical gradient direction of the first-order moment momentum value, avoid frequent changes in the gradient direction due to the dynamics of road and plot data, and at the same time ensure that the gradient needs to be kept consistent for a long time due to strong normalization, so as to strengthen the continuity of the planning intention and accelerate the model convergence. The specific formula is as follows: m t = β1·m t-1 + (1 - β1)·g t where m t is the updated momentum value, t is the t-th iteration, t - 1 is the number of the (t - 1)-th iteration, m t-1 is the momentum value at the (t - 1)-th iteration, g t is the gradient of the current step, β1 controls the decay rate of the historical gradient; According to the momentum calculation results, calculate the updated weight parameters for the next iteration of each layer of the neural network, and perform successive iterative calculations to continuously optimize the morphological parameters of the heterogeneous graph neural network generated by the target site design, forming an optimized road and spatial form scheme. The specific formula is as follows: Among them, W t+1 is the updated weight parameter for the (t + 1)-th iteration, and W t is the weight parameter for the t-th iteration. τ is an adaptive learning rate term, ρ is an adaptive learning rate term, and ∈ is a constant to prevent a division-by-zero error when the denominator ρ approaches zero; At the same time, limit the gradient update range through gradient clipping technology, constrain the gradient amplitude within a preset threshold to avoid the problems of gradient explosion or gradient disappearance during training; optimize the semantic mapping ability of the heterogeneous graph neural network of the target design site in the latent space by iteratively updating the weight matrices of the encoder and decoder, and ensure that the generated road alignment and plot morphological parameters match the topological features and semantic consistency of the planning background data.

9. The method for integrally and intelligently generating urban roads and forms based on a heterogeneous graph network according to claim 6, wherein: The UMAP method in step (6) is as follows: Input the node features and edge features in the target design site scheme generated by the heterogeneous graph network into the UMAP algorithm. By calculating the similarity of nodes in the high-dimensional space and the similarity of low-dimensional space coordinates, and minimizing the difference between the two similarities, the coordinate relationship in the low-dimensional space is made infinitely close to the structural relationship in the high-dimensional space to generate spatial coordinates. The coordinates retain the local and global structural features of the original high-dimensional data, providing a geometric basis for the construction of the 3D model. The specific formula is as follows: Among them, p ij is the similarity value in the high-dimensional space, d(x i , x j ) is the Euclidean distance between node i and node j, μ i is the distance from node i to its nearest neighbor node, and σ i is the scale parameter; Randomly assign initial three-dimensional coordinates (x, y, z) to each node, calculate the distance between nodes in the three-dimensional space using the initial three-dimensional coordinates, and calculate the similarity of the low-dimensional space coordinates. The specific formula is as follows: where q ij is the similarity value in the low-dimensional space, a and b are hyperparameters controlling the distribution shape, and di,j is the coordinate distance between nodes i and j; where M is the minimum value of the high-dimensional space similarity and the low-dimensional space similarity; Continuously iterate the initial three-dimensional coordinates assigned to each node according to the calculation result of the M value until the high-dimensional space similarity and the low-dimensional space similarity are nearly the same. Output the three-dimensional space coordinates of each node, and use the three-dimensional point cloud rendering technology to generate an interactive virtual scene model.

Images