Generative artificial intelligence driven interchange design method

Abstract

The invention discloses a generative artificial intelligence-driven interchange design method, which comprises the following steps of: extracting data from multiple types of interchange by adopting a graph representation mode and a graph enhancement means, and converting the data into a graph structure to store design parameters so as to construct an interchange topology database; based on an encoder, a decoder and a condition vector embedding module, constructing an interchange structure generation model for realizing graph reconstruction and data distribution fitting, and learning from a database to generate an optimized interchange structure by embedding a design target into the model; the interchange structure generation model adopts a predictor to directly output an interchange structure, so that rapid screening of a complete and effective structure is realized; and performing quality evaluation on the generated interchange structure in combination with a graph attribute predictor and a graph similarity calculation module. The method is based on a graph convolution condition variational auto-encoder, and focuses on improving the design performance of the interchange in the aspects of throughput capacity and ramp length.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of road engineering design, artificial intelligence assisted planning and traffic facility automation design, and particularly relates to a generative artificial intelligence driven intercommunication interchange design method. BACKGROUND

[0002] As a key node connecting traffic flows in different directions, the expressway intercommunication interchange plays an important role in ensuring the efficiency and safety of the road network. The traditional interchange design process usually relies on manual operation, including data research, interchange selection, linear design and scheme comparison and selection, etc. The process is complex and time-consuming, and the design effect depends on the experience of engineers, which has the problems of limited exploration space and suboptimal scheme.

[0003] With the development of generative artificial intelligence (AIGC) technology, data-driven automatic structure generation methods have been gradually applied to complex structure design. In terms of graph structure modeling, models such as GraphVAE and GraphRNN can be used for structure generation, but direct application to interchange design still faces problems such as high degree of freedom, complex traffic target constraints and difficulty in evaluating generation quality. Therefore, there is an urgent need for a target-oriented graph generation method that can combine traffic performance indicators to automatically generate and evaluate structure design. SUMMARY

[0004] The present application aims to provide a generative artificial intelligence driven intercommunication interchange design method, which realizes the automatic generation and evaluation of intercommunication interchange structure through a target-oriented graph generation model, and improves the design efficiency and scheme quality.

[0005] To achieve the above functions, the present application designs a generative artificial intelligence driven intercommunication interchange design method, which executes the following steps S1-S3 for the interchange structure in the road network to complete the design of a new interchange structure that meets the preset target:

[0006] Step S1: Convert the interchange structure into a topological graph form for representation, construct an adjacency matrix and a node feature matrix according to the parameters affecting the plane design and the connection relationship between road segments, compose a feature matrix, and construct a design-oriented graph structure dataset after graph enhancement;

[0007] Step S2: Construct an interchange structure generation model based on a graph conditional variational autoencoder, including an encoder, a decoder and a conditional vector embedding module, wherein the encoder extracts the representation of the graph structure and the node features based on a multi-layer graph neural network, the interchange design target indicators are embedded into the model as a conditional vector, the decoder reconstructs the adjacency matrix and the node feature matrix of the graph based on a multi-layer perceptron, outputs the graph structure reconstruction result as the generated intercommunication interchange structure, and trains the interchange structure generation model;

[0008] Step S3: using the trained interchange structure generation model, generating a new interchange structure according to the interchange design target index, and combining the graph attribute predictor and the graph similarity calculation module to evaluate the quality of the generated interchange structure.

[0009] As a preferred technical solution of the present application: the specific steps of step S1 are as follows:

[0010] Step S1.1: convert the original interchange structure into a graph G={V,E}, where V={v1,v2,v3……vn} represents a node set, and the nodes in the set are road segments in the original interchange structure, where n represents the number of nodes; n} represents a node set, and the nodes in the set are road segments in the original interchange structure, where n represents the number of nodes; E={u,v} represents an edge set, and u and v represent two nodes in the graph; the edges in the set represent the connection relationship between adjacent road segments in the original interchange structure;

[0011] Step S1.2: for each node, select a plurality of parameters affecting the plane design to form a node feature matrix F of the graph where d represents the number of features; at the same time, the connection relationship between adjacent road segments in the original interchange structure is constructed as an adjacency matrix, and the undirected edges are constructed according to the connection relationship between the road segments, to form an adjacency matrix W∈{0,1} n×n ;

[0012] Step S1.3: use a graph enhancement method to expand the data distribution and construct a graph structure data set of the interchange structure.

[0013] As a preferred technical solution of the present application: the parameters affecting the plane design selected in step S1.2 include: road type, design speed, curvature, length, direction angle, number of lanes, acceleration section length and deceleration section length.

[0014] As a preferred technical solution of the present application: the graph enhancement method used in step S1.3 includes node-level perturbation or graph-level pruning method.

[0015] As a preferred technical solution of the present application: the specific steps of step S2 are as follows:

[0016] Step S2.1: construct an encoder, and use a multi-layer graph neural network to encode the graph for the input feature matrix G={W,F}, and use the following formula for node aggregation and update at each layer:

[0017]

[0018] wherein, wherein represents the hidden state of the kth layer node v1, σ represents an activation function, W kdenotes the learnable weight parameter of the k-th layer, N(v1) denotes the neighbor node set of node v1, v1 denotes the current target node in the multi-layer graph neural network, u1 denotes the neighbor node in the multi-layer graph neural network, AGG k denotes the aggregation function;

[0019] The encoder calculates the mean of the latent space through linear transformation and variance as follows:

[0020]

[0021]

[0022] In the formula, denotes the hidden state of the k-th layer node v1, K is the number of layers of the encoder, W μ , b μ denote the weight matrix and bias term of the output mean, respectively; LINEAR μ (·) denotes the calculation of the mean through the linear layer; denotes the diagonal elements of the extracted covariance matrix , forming a variance vector; denotes the calculation of the variance through the linear layer;

[0023] The latent variable is sampled from the standard normal distribution

[0024] Step S2.2: Based on the design target index of the interchange, a condition vector Y is constructed, which is spliced with the latent variable Z as the input of the decoder; the decoder recovers the node embedding representation through the multi-layer perception, and generates the adjacency matrix and node feature matrix of the graph through the following mapping:

[0025]

[0026] In the formula, denotes the adjacency matrix of the generated graph, denotes the node feature matrix of the generated graph; LINEAR W (·) denotes the linear transformation layer of the decoder, which is used to output the adjacency matrix; LINEAR F (·) denotes the linear transformation layer of the decoder, which is used to output the adjacency matrix and the node feature matrix; H K denotes the hidden state matrix obtained by the multi-layer perception at the k-th layer;

[0027] Step S2.3: Train the interchange structure generation model, and design the loss function as follows:

[0028]

[0029] In the formula, represents a loss function, ELBO represents the lower bound of evidence; the first item on the right side of the equation is the reconstruction loss, and the second item on the right side is the KL divergence loss; represents the expectation under the conditional probability distribution q φ (z|G,Y);p θ ((G|z,Y)) represents the probability distribution of the generation model given the latent variable z and the conditional vector Y; p(z|Y) represents the prior distribution of the latent space, which is set as a standard normal distribution; β represents the weight coefficient of the KL divergence term;

[0030] wherein the reconstruction loss is further split into a feature reconstruction term and a connection matrix reconstruction term, which are defined as follows:

[0031]

[0032] In the formula, represents the reconstruction loss of the node feature matrix, F M represents the predictable part in the original node feature matrix, represents the predictable part in the reconstructed node feature matrix output by the decoder; respectively represent the original feature vector and the reconstructed feature vector of the vth node; u and v represent two nodes in the graph.

[0033] represents the reconstruction loss of the adjacency matrix, which is calculated using the cross-entropy loss; W M represents the predictable part in the original adjacency matrix; represents the predictable part in the reconstructed adjacency matrix output by the decoder; represents the original connection probability and the reconstructed connection probability between nodes v and u;

[0034] The KL divergence is calculated as follows:

[0035]

[0036] In the formula, KL[q φ (z|G,Y)|p(z|Y)] represents the Kullback-Leibler divergence between the posterior distribution and the prior distribution in the latent space; s represents the dimension of the latent space. represents the variance of the jth dimension of node i in the latent space; represents the square of the mean of the jth dimension of node i in the latent space.

[0037] As a preferred technical solution of the present application: the interchange design target index in step S2.2 includes traffic capacity, ramp length, and node number.

[0038] As a preferred technical solution of the present application: the graph attribute predictor in step S3 performs attribute accuracy evaluation on the generated graph structure reconstruction result, the graph attribute predictor includes a graph classification module and a node classification module composed of a multi-layer graph neural network graph convolution layer, a pooling layer and a linear layer with a feature matrix G = {W, F} as input; after being processed by a softmax activation function, each item of the generated interflow interchange structure design target index is evaluated whether it falls into the corresponding interval range.

[0039] As a preferred technical solution of the present application: the graph similarity calculation module in step S3 evaluates the comprehensive performance of the generated graph structure reconstruction result based on similarity; the similarity calculation formula is:

[0040]

[0041] In the formula, denotes the similarity score of graphs G1 and G2, and denote the embedding vectors of graphs G1 and G2, respectively, and σ 2 is a hyperparameter;

[0042] After obtaining the similarity score, it is judged whether graphs G1 and G2 are the same according to the set threshold.

[0043] Advantages: Compared with the prior art, the present application has the following advantages:

[0044] 1. Constructing an interchange scheme library oriented to design tasks. The present application proposes a graph structure representation method oriented to interchange design tasks, which can uniformly represent the structure and parameter characteristics of different types of interflow interchanges, thereby constructing an interchange graph database covering structural diversity and design target diversity, providing support for data-driven automatic design and model training;

[0045] 2. Developing a graph-based interchange deep generation model. The present application first introduces a graph neural network structure into the interflow interchange generation task, introduces a conditional vector to control the generation direction in the model, and considers both topological constraints and design targets to realize directional control of attributes such as traffic capacity and ramp length. Compared with traditional CVAE models, it performs better in accuracy, uniqueness and effectiveness, and is suitable for intelligent design of complex road networks;

[0046] 3. Realizing a high-quality target-oriented interchange intelligent generation process. Compared with the traditional manual design process, the present application can automatically generate a large number of high-quality interflow interchange structures that meet the target constraints in a very short time, especially when facing unknown design targets, it still maintains a high generation accuracy, structural uniqueness and novelty. The generated interchanges are superior to the original training samples in terms of throughput, length and other indicators, and have significant engineering practical value and generalization ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 is a flow chart of a generative artificial intelligence-driven interchange design method provided according to an embodiment of the present invention;

[0048] Figure 2 is a flow chart of a method for graph representation of interchange topology according to an embodiment of the present invention;

[0049] Figure 3 is a schematic diagram of a graph enhancement strategy provided according to an embodiment of the present invention;

[0050] Figure 4 3 is a schematic diagram of an interchange structure generation model based on a graph conditional variational autoencoder provided according to an embodiment of the present invention. DETAILED DESCRIPTION

[0051] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention.

[0052] The embodiment of the present invention provides a generative artificial intelligence driven interchange design method for the interchange structure in the road network, referring to Figure 1 , perform the following steps S1 to S3 to complete the design of a new interchange structure that meets the preset goals:

[0053] Step S1: The interchange structure is converted into a topological graph for representation. Based on the parameters affecting the plane design and the connection relationship between road sections, an adjacency matrix and a node feature matrix are constructed to form a feature matrix. After graph enhancement, a design-oriented graph structure dataset is constructed.

[0054] The specific steps of step S1 are as follows:

[0055] Step S1.1: Refer to Figure 2 , transform the original interchange structure into a graph G = {V, E}, where V = {v1, v2, v3 ... v n} represents a node set, where the nodes in the set are the road sections in the original interchange structure, and n represents the number of nodes; represents a set of edges, u and v represent two nodes in the graph, and (u, v) is included in the set E as an edge. The edges in the set represent the connection relationship between adjacent road sections in the original interchange structure.

[0056] Step S1.2: For each node, select multiple parameters that affect the plane design to form the node feature matrix of the graph where d represents the number of features; in one embodiment, the selected parameters affecting the design of the intersection include: road type, design speed, curvature, length, direction angle, number of lanes, acceleration section length, and deceleration section length, d = 8; meanwhile, the connection relationship between adjacent road segments in the original interchange structure is constructed as an adjacency matrix, and undirected edges are constructed according to the connection relationship between the road segments to form an adjacency matrix W ∈ {0, 1} n×n ;

[0057] Step S1.3: referring to Figure 3 , in order to further enrich the data distribution, a graph enhancement method is used to expand the data distribution to construct a graph structure data set of the interchange structure; the graph enhancement method used includes a node-level disturbance or a graph-level pruning method. The node-level disturbance disturbs the node features by swapping the ramp position, rotating the left-turn ramp counterclockwise, and keeping the right-turn ramp position unchanged, changes the position of the node while retaining the topological information to obtain different context information. The graph-level pruning method obtains an incomplete interchange form under a partial flow condition by deleting individual flow ramps, i.e., a new subgraph. The attributes corresponding to the enhanced graph data will also change, and the labels need to be recalculated.

[0058] Step S2: constructing an interchange structure generation model based on a graph conditional variational autoencoder (GraphCVAE), referring to Figure 4 , including an encoder, a decoder, and a conditional vector embedding module, wherein the encoder extracts the representation of the graph structure and the node features based on a multi-layer graph neural network (GraphSAGE), embeds the interchange design target indicators into the model as a conditional vector, the decoder reconstructs the adjacency matrix and the node feature matrix of the graph based on a multi-layer perceptron (MLP), outputs the graph structure reconstruction result as the generated interchange structure, and trains the interchange structure generation model;

[0059] The specific steps of step S2 are as follows:

[0060] Step S2.1: constructing an encoder, the encoder is used to extract the hidden representation of each node from the input graph, encode the local topological structure and attribute feature information thereof; for the input feature matrix G = {W, F}, a multi-layer graph neural network (GraphSAGE) is used to encode the graph, the embedding representation of each layer of nodes is obtained by aggregating the neighbor node information and the previous layer state of itself, and each layer uses the following formula for node aggregation and update:

[0061]

[0062] wherein, wherein represents the hidden state of the kth layer node v1, σ represents an activation function, W kdenotes the learnable weight parameter of the k-th layer, N(v1) denotes the neighbor node set of node v1, v1 denotes the current target node in the multi-layer graph neural network, u1 denotes the neighbor node in the multi-layer graph neural network, AGG k denotes the aggregation function, which samples and aggregates the feature information from the neighbor nodes of the previous layer, and the commonly used aggregation functions include Mean, LSTM and Pooling aggregator;

[0063] The encoder calculates the mean of the latent space through linear transformation and the variance as follows:

[0064]

[0065]

[0066] wherein, denotes the hidden state of the k-th layer node v1, K is the number of layers of the encoder, W μ and b μ denote the weight matrix and bias term for outputting the mean value, respectively; LINEAR μ (·) denotes the calculation of the mean value through a linear layer; denotes the diagonal elements of the covariance matrix , forming a variance vector; denotes the calculation of the variance through a linear layer;

[0067] The latent variable z is sampled from the standard normal distribution and is passed to the decoder as one of the inputs for the conditional generation process;

[0068] Step S2.2: Based on the design target indicators of the interchange, including the traffic capacity, the slope length, and the number of nodes, a conditional vector Y is constructed, which is concatenated with the latent variable Z as the input of the decoder; the decoder recovers the node embedding representation through a multi-layer perception, which is used to reconstruct the adjacency matrix and the node feature matrix of the graph; the decoder generates the adjacency matrix and the node feature matrix of the graph through the following mapping:

[0069]

[0070] wherein, denotes the adjacency matrix of the generated graph, i.e., the connection probability matrix, denotes the node feature matrix of the generated graph; LINEAR W (·) denotes the linear transformation layer of the decoder, which is used to output the adjacency matrix; LINEAR F (·) denotes the linear transformation layer of the decoder, which is used to output the adjacency matrix and the node feature matrix; H K ​denotes the hidden state matrix obtained by the multi-layer perception at the Kth layer;

[0071] Step S2.3: The model for the grade separation structure is trained by maximizing the Evidence Lower Bound (ELBO), which is the objective function of the variational autoencoder and is used to maximize the likelihood probability of the data in the latent space; the model training process adopts a variational inference strategy, and the loss function is composed of a reconstruction loss and a KL divergence term, and the loss function is designed as follows:

[0072]

[0073] In the formula, denotes the loss function, and ELBO denotes the evidence lower bound; the first term on the right side of the equation is the reconstruction loss, and the generated graph has a high degree of similarity with the input graph G. The second term on the right side is defined as the KL divergence loss, which regularizes the latent space and measures the difference between the latent distribution and the prior distribution; denotes the expectation under the conditional probability distribution q φ (z|G,Y); p θ ((G|z,Y)) denotes the probability distribution of the generation model given the latent variable z and the conditional vector Y; p(z|Y) denotes the prior distribution in the latent space, which is set as a standard normal distribution; β denotes the weight coefficient of the KL divergence term, which is used to adjust the weight of the loss term and balance the reconstruction loss and the regularization of the latent space;

[0074] wherein the reconstruction loss is further split into a feature reconstruction term and a connection matrix reconstruction term, which are defined as follows:

[0075]

[0076] In the formula, denotes the reconstruction loss of the node feature matrix, which calculates the mean square error between the predicted matrix and the original matrix, F M denotes the predictable part (after applying the mask) in the original node feature matrix, denotes the predictable part in the reconstructed node feature matrix output by the decoder; denotes the original feature vector and the reconstructed feature vector of the vth node, respectively; u and v denote two nodes in the graph.

[0077] denotes the reconstruction loss of the adjacency matrix, which is calculated by using a cross-entropy loss; W M denotes the predictable part (after applying the mask) in the original adjacency matrix; denotes the predictable part in the reconstructed adjacency matrix output by the decoder; denotes the original connection probability and the reconstructed connection probability between nodes v and u.

[0078] The KL divergence is calculated as follows:

[0079]

[0080] where KL[q φ (z|G,Y)|p(z|Y)] represents the Kullback-Leibler divergence of the latent space posterior distribution and the prior distribution; s represents the latent space dimension; represents the variance of the jth dimension of node i in the latent space; represents the square of the mean of the jth dimension of node i in the latent space.

[0081] Through the above training mechanism, the model can effectively capture the interchange topological structure characteristics and learn an efficient and controllable graph generation strategy under the condition of a given design target, and finally realize the ability of "target-driven structure generation".

[0082] Step S3: using the trained interchange structure generation model, generating a new interflow interchange structure according to the interchange design target index, and combining the graph attribute predictor and the graph similarity calculation module to evaluate the quality of the generated interflow interchange structure.

[0083] When the interchange structure generation model is trained, the decoder part will be used as a generator, combined with the input target condition vector, to drive the generation of the interflow interchange structure. The target condition vector and the hidden variable obtained by sampling are spliced and sent to the decoder, and the adjacency matrix and node attribute matrix of the generated graph are output, so as to construct a new interchange scheme.

[0084] In order to verify the quality of the generated interflow interchange structure, the present application proposes two types of evaluation index systems:

[0085] Attribute accuracy evaluation: whether the reconstructed result of the generated graph structure meets the input conditions is judged by the graph attribute predictor, and the indicators include ramp traffic capacity, ramp number and other key attributes. A regression and classification network is constructed based on a multi-layer graph neural network (GraphSAGE) for accurate regression / classification of multiple structure attributes;

[0086] The graph attribute predictor evaluates the attribute accuracy of the reconstructed result of the generated graph structure, and the graph attribute predictor includes a graph classification module and a node classification module composed of a multi-layer graph neural network graph convolution layer, a pooling layer and a linear layer with a feature matrix G={W,F} as input; after being processed by a softmax activation function, the traffic capacity, throughput and ramp length of the generated interflow interchange structure are respectively evaluated whether they fall within the corresponding interval range.

[0087] Model performance evaluation: Based on large-scale generation of new graphs, the graph similarity calculation module is used to measure the difference between the generated graph and the training set. Specifically, the kernel function based on node embedding is used to calculate the graph structure similarity, and the evaluation indexes such as "effectiveness, uniqueness, novelty" are introduced to measure the generation ability, which comprehensively describes the generalization and diversity of the generation model;

[0088] The graph similarity calculation module evaluates the comprehensive performance of the generated graph structure reconstruction result based on similarity. The similarity calculation formula is:

[0089]

[0090] In the formula, G1 and G2 represent the similarity score of the graph G1 and G2, and respectively represent the embedding vector of the graph G1 and G2, σ 2 is a hyperparameter for controlling the function shape;

[0091] After obtaining the similarity score, it is judged whether the graph G1 and G2 are the same according to the set threshold.

[0092] The embodiments of the application are described in detail above in combination with the drawings, but the application is not limited to the above embodiments, and various changes can be made within the knowledge of those skilled in the art without departing from the purpose of the application.

Claims

1. A generative artificial intelligence-driven interchange design method, characterized in that: For the interchange structure in the road network, execute the following steps S1 to S3 to complete the design of a new interchange structure that meets the preset objectives: Step S1: The interchange structure is converted into a topological graph for representation. Based on the parameters affecting the plane design and the connection relationship between road sections, an adjacency matrix and a node feature matrix are constructed to form a feature matrix. After graph enhancement, a design-oriented graph structure dataset is constructed. Step S2: Constructing an interchange structure generation model based on a graph conditional variational autoencoder, comprising an encoder, a decoder, and a conditional vector embedding module, wherein the encoder extracts representations of graph structure and node features based on a multi-layer graph neural network, embeds interchange design target indicators into the model as conditional vectors, and the decoder reconstructs the graph adjacency matrix and node feature matrix based on a multi-layer perceptron, outputs the graph structure reconstruction result as the generated interchange structure, and trains the interchange structure generation model; Step S3: Using the trained interchange structure generation model, a new interchange structure is generated according to the interchange design target indicators. The quality of the generated interchange structure is evaluated by combining the graph attribute predictor and the graph similarity calculation module.

2. A generative AI-driven interchange design method according to claim 1, characterized in that: The specific steps of step S1 are as follows: Step S1.1: Convert the original interchange structure into a graph G = {V, E+, where V = {v1, v2, v3 ... v n + represents a node set, where the nodes in the set are the road sections in the original interchange structure, and n represents the number of nodes; Represents a set of edges, u and v represent two nodes in the graph; the edges in the set represent the connection relationship between adjacent road sections in the original interchange structure; Step S1.2: For each node, select multiple parameters that affect the plane design to form the node feature matrix of the graph Where d represents the number of features; at the same time, the connection relationship between adjacent sections in the original interchange structure is constructed as an adjacency matrix, and undirected edges are constructed according to the connectivity relationship between each section to form an adjacency matrix W∈{0,1+ n×n ; Step S1.3: Use graph enhancement methods to expand the data distribution and construct a graph structure dataset of the interchange structure.

3. The generative artificial intelligence-driven interchange design method according to claim 2, characterized in that: The parameters affecting the plane design selected in step S1.2 include: road type, design speed, curvature, length, direction angle, number of lanes, acceleration section length and deceleration section length.

4. The generative artificial intelligence-driven interchange design method according to claim 2, characterized in that: The graph enhancement methods used in step S1.3 include node-level perturbation or graph-level pruning methods.

5. The generative artificial intelligence-driven interchange design method according to claim 2, characterized in that: The specific steps of step S2 are as follows: Step S2.1: Construct an encoder. For the input feature matrix G = {W, F + , a multi-layer graph neural network is used to encode the graph. Each layer uses the following formula for node aggregation and update: Among them, represents the hidden state of the k-th layer node v1, σ represents the activation function, W k represents the learnable weight parameter of the kth layer, N(v1) represents the set of neighbor nodes of node v1, v1 represents the current target node in the multi-layer graph neural network, u1 represents the neighbor node in the multi-layer graph neural network, AGG k Represents an aggregate function; The encoder calculates the mean of the latent space through linear transformation and variance As follows: Where, represents the hidden state of the K-th layer node v1, K is the number of layers of the encoder, W μ 、b μ Represents the weight matrix and bias term of the output mean respectively; LINEAR μ (·) indicates that the mean is calculated through the linear layer; Represents the extracted covariance matrix The diagonal elements of form the variance vector; Indicates that the variance is calculated through the linear layer; From the standard normal distribution Hidden variables sampled from Step S2.2: Based on the interchange design target, construct the conditional vector Y and concatenate it with the latent variable Z as the input to the decoder. The decoder recovers the node embedding representation using a multi-layer perceptron and generates the graph adjacency matrix and node feature matrix using the following mapping: Where, represents the adjacency matrix of the generated graph, Represents the node feature matrix of the generated graph; LINEAR W (·) represents the linear transformation layer of the decoder, which is used to output the adjacency matrix; LINEAR F (·) represents the linear transformation layer of the decoder, which is used to output the adjacency matrix and node feature matrix; H K Represents the hidden state matrix obtained by the multilayer perceptron at the Kth layer; Step S2.3: Train the interchange structure generation model and design the loss function as follows: Where, represents the loss function, ELBO represents the lower bound of evidence; the first term on the right side of the equal sign is the reconstruction loss, and the second term on the right side is the KL divergence loss; Indicates that in the conditional probability distribution q φ expectation under (z|G,Y); p θ ((G|z,Y)) represents the probability distribution of the generative model given the latent variable z and the conditional vector Y; p(z|Y) represents the prior distribution of the latent space, which is set to the standard normal distribution; β represents the weight coefficient of the KL divergence term; Among them, the reconstruction loss is further split into feature reconstruction term and connection matrix reconstruction term, which are defined as follows: Where, represents the reconstruction loss of the node feature matrix, F M Represents the predictable part of the original node feature matrix, Represents the predictable part of the reconstructed node feature matrix output by the decoder; Represent the original eigenvector and reconstructed eigenvector of the vth node respectively; u and v represent two nodes in the graph; represents the reconstruction loss of the adjacency matrix, which is calculated using cross entropy loss; W M Represents the predictable part of the original adjacency matrix; represents the predictable part of the reconstructed adjacency matrix output by the decoder; Represents the original connection probability and reconstructed connection probability between nodes v and u; KL divergence is calculated as follows: In the formula, KL[q φ (z|G,Y)|p(z|Y)] represents the Kullback-Leibler divergence between the posterior distribution and the prior distribution of the latent space; s represents the dimension of the latent space; represents the variance of the j-th dimension of node i in the latent space; represents the square of the mean of the j-th dimension of node i in the latent space.

6. The generative artificial intelligence-driven interchange design method according to claim 5, characterized in that: The target indicators for interchange design described in step S2.2 include traffic capacity, ramp length, and number of nodes.

7. The generative artificial intelligence-driven interchange design method according to claim 1, characterized in that: The graph attribute predictor described in step S3 evaluates the attribute accuracy of the generated graph structure reconstruction result. The graph attribute predictor includes a graph classification module and a node classification module consisting of a multi-layer graph neural network graph convolution layer, a pooling layer and a linear layer with the feature matrix G = {W, F+ as input; after the output is processed by the softmax activation function, it evaluates whether the various interchange design target indicators of the generated interchange structure fall within the corresponding interval range.

8. The generative artificial intelligence-driven interchange design method according to claim 1, characterized in that: The graph similarity calculation module described in step S3 evaluates the comprehensive performance of the generated graph structure reconstruction result based on the similarity; the similarity calculation formula is: Where, represents the similarity score between graphs G1 and G2, and Denote the embedding vectors of graphs G1 and G2 respectively, σ 2 is a hyperparameter; After obtaining the similarity score, we determine whether graphs G1 and G2 are the same based on the set threshold.

Images