Urban road network traffic jam state prediction method and system based on deep learning

Abstract

The invention provides an urban road network traffic congestion state prediction method and system based on deep learning, and relates to the technical field of intelligent traffic control, and the method comprises the steps: obtaining congestion state image data of an urban road network, and carrying out the preprocessing to extract structured road congestion information; constructing a road network topological graph based on the road congestion information, and performing structure sensing state coding on the road network topological graph to generate a structure sensing state vector fusing local congestion features and global topological position information; and inputting the structure perception state vector into an improved deep learning prediction model, performing feature deepening and enhancement through a graph neural network which dynamically adjusts a graph structure and aggregates node information, and finally outputting road network congestion state prediction results of a plurality of time scales in the future. According to the method, accurate prediction of the urban road network congestion state can be realized, and reliable technical support is provided for intelligent traffic management, resource scheduling optimization, travel path planning and the like.

Description

Technical Field

[0001] This invention belongs to the field of intelligent traffic control technology, specifically relating to a method and system for predicting urban road network traffic congestion based on deep learning. Background Technology

[0002] With the continuous acceleration of urbanization, urban road traffic demand has exploded, and traffic congestion has become a core bottleneck restricting urban operational efficiency, affecting residents' travel experience, and hindering sustainable development. Traditional traffic congestion monitoring relies on manual patrols, fixed road sensors, or dedicated navigation equipment, which have shortcomings such as limited coverage, high deployment and maintenance costs, and lagging data updates, making it difficult to meet the needs of real-time monitoring and prediction of the entire urban road network.

[0003] In recent years, online map service providers (such as Gaode, Baidu, and Tencent Maps) have leveraged massive amounts of mobile terminal user data to provide large-scale, wide-coverage, and frequently updated urban road network congestion image data, offering new data support for congestion analysis and prediction. However, this type of congestion data is presented in high-resolution image form, characterized by large data volume, high dimensionality, complex road hierarchy, and uneven information density. Directly using it for modeling can easily lead to problems such as the curse of dimensionality, redundant information interference, and the submersion of key features, urgently requiring breakthroughs in efficient feature extraction and structured representation technologies.

[0004] Currently, machine learning and deep learning technologies have made some progress in the application of traffic prediction. Traditional models such as KNN, SVM, and ANN, as well as deep learning models such as LSTM, CNN, and ST-GCN, have shown potential in temporal dependency modeling and spatial feature extraction. However, existing technologies still have significant shortcomings:

[0005] 1. Insufficient utilization of multi-level road network information: Most methods only focus on extracting congestion features of main roads such as arterial roads, and lack global unified quantification and fusion modeling of congestion status of multiple levels of roads such as arterial roads, secondary roads, and branch roads;

[0006] 2. Insufficient capture of road network topology information: Traditional models are unable to deeply explore the spatial connections and topological relationships between roads, and cannot effectively integrate local and global structural features, resulting in inaccurate modeling of congestion propagation paths and evolution patterns;

[0007] 3. Severe homogenization of node features: Existing graph models are prone to node feature convergence during feature aggregation, making it difficult to highlight congestion features of key road segments and affecting prediction accuracy;

[0008] 4. Limited remote information interaction: Traditional models rely heavily on the updates of information from neighboring nodes, which makes it difficult to achieve effective information interaction between nodes that are spatially distant but closely related to congestion, and makes it difficult to capture the long-distance dependence of congestion propagation.

[0009] 5. Poor dynamic adaptability of topology: The road network topology changes dynamically with factors such as congestion propagation and temporary traffic control. Existing models lack an adaptive adjustment mechanism and cannot dynamically optimize weight allocation according to differences in road levels.

[0010] Therefore, there is an urgent need for a high-precision prediction technology that can deeply integrate road network topology, effectively extract multi-level congestion features, and adapt to dynamically changing scenarios. Summary of the Invention

[0011] The core objective of this invention is to address the problems of insufficient topological information capture, homogenized node features, limited remote information interaction, and poor dynamic adaptability in existing urban road network traffic congestion prediction technologies. This invention provides a high-precision, robust, and adaptable congestion prediction method and system. Through innovative key technologies such as topological feature encoding, node feature enhancement, remote information interaction, and dynamic topology updates, this invention achieves accurate prediction of urban road network congestion, providing reliable technical support for intelligent traffic management, resource scheduling optimization, and travel route planning.

[0012] To achieve the above objectives, the following technical solution is adopted:

[0013] According to a first aspect of the present invention, a method for predicting urban road network traffic congestion based on deep learning is provided, comprising the following steps:

[0014] Data acquisition and preprocessing steps: Acquire image data of urban road network congestion status and perform preprocessing to extract structured road congestion information;

[0015] Topology feature extraction steps: Construct a road network topology map based on the road congestion information, and generate a structure-aware state vector that integrates local congestion features and global topology location information by performing structure-aware state encoding on the road network topology map;

[0016] Congestion state prediction steps: The structure-aware state vector is input into the improved deep learning prediction model. The graph neural network that dynamically adjusts the graph structure and aggregates node information is used to deepen and enhance the features. Finally, the prediction results of road network congestion state at multiple future time scales are output.

[0017] Furthermore, the topology feature extraction step includes road network topology map construction and optimization, specifically including:

[0018] Construct a road network topology model based on road geometry and intersection connections;

[0019] The road attribute information in the congestion status image is mapped to the road network topology map using a coordinate matching algorithm;

[0020] A spatial partitioning algorithm is used to adaptively reduce the dimensionality of the initial topology graph. Key nodes and edges are selected based on node density and edge importance indices to form a simplified topology graph.

[0021] Furthermore, the topological feature extraction step also includes topological feature vector construction, specifically including:

[0022] Multi-dimensional topological features are extracted from the simplified topology graph, including node congestion, neighborhood congestion gradient, and spatial correlation.

[0023] The node congestion level is obtained by calculating the average congestion level of adjacent road segments of the node;

[0024] The neighborhood congestion gradient is obtained by calculating the maximum absolute difference between the congestion level of the adjacent road segments of the node and the node's congestion degree;

[0025] The spatial correlation degree is obtained by weighting the average congestion level of the node's neighborhood by road level.

[0026] Furthermore, the structure-aware state encoding in the topology feature extraction step specifically includes:

[0027] Multi-layer graph convolution operation is performed on the simplified topology graph to capture the higher-order dependencies of nodes and output the local neighborhood congestion features of the nodes.

[0028] The structural position embedding vector is calculated based on the shortest path distance between nodes, representing the topological position of the node in the global road network;

[0029] The local neighborhood congestion features of the nodes output by graph convolution are fused with the structure location embedding vector to generate a structure-aware state vector that simultaneously contains local congestion features and global topological location information.

[0030] Furthermore, in the congestion state prediction step, feature enhancement and amplification are performed by a graph neural network that dynamically adjusts the graph structure and aggregates node information, including dynamic graph convolution feature extraction, personalized enhancement of node features, remote node information interaction, and dynamic updating of the topology graph structure.

[0031] The dynamic graph convolution feature extraction specifically includes:

[0032] The structure-aware state vector is enhanced by using a dynamic graph convolutional network. The adjacency matrix is ​​adaptively adjusted to capture dynamic topological changes. This includes: dynamically adjusting the adjacency matrix based on node feature similarity; fusing the similarity matrix of node features after linear transformation with the original adjacency matrix to generate an adaptive adjacency matrix; normalizing the adaptive adjacency matrix; and aggregating neighborhood information through graph convolution operations to achieve dynamic enhancement of node features.

[0033] Furthermore, in the process of personalized enhancement of node features, a score-driven mechanism is introduced to learn the contribution score of each node to global congestion prediction, and the node features are personalized and weighted based on the corrected contribution scores to strengthen the expression of key road segment features, specifically:

[0034] Learn the contribution score of each node to the global congestion prediction;

[0035] The node features are sorted and corrected based on the contribution scores;

[0036] The modified scores are used to personalize the weighting of node features to enhance the feature representation of key nodes.

[0037] Furthermore, the enhanced remote information interaction specifically includes:

[0038] Calculate the feature similarity between all nodes and construct a global similarity matrix;

[0039] Based on the global similarity matrix, information of nodes that are spatially distant but have similar features is aggregated, and the original features are preserved through residual connections to capture long-distance dependencies in congestion propagation and update node features.

[0040] Furthermore, the dynamic updating of the topology graph structure specifically includes:

[0041] Based on the model's prediction error feedback and real-time congestion status changes, the importance weight of each road segment in the road network is calculated.

[0042] Set dynamically adjustable retention and addition thresholds, and dynamically update the edge set of the topology graph according to the importance weights, retaining important edges and adding new key connections to adapt to the dynamic changes in the road network topology.

[0043] Furthermore, the congestion prediction step employs multi-scale prediction output, specifically including:

[0044] The processed node features are input into a bidirectional LSTM temporal prediction network to capture both the forward and reverse temporal dependencies of congestion states.

[0045] Based on the fusion features of bidirectional LSTM output, the predicted values ​​of road network congestion status at three time scales—5 minutes, 15 minutes, and 30 minutes—are output through a fully connected layer.

[0046] According to a second aspect of the present invention, a deep learning-based urban road network traffic congestion state prediction system is also provided for implementing the method described in the first aspect, the system comprising:

[0047] Data acquisition and preprocessing module: used to acquire image data of urban road network congestion status and perform preprocessing to extract structured road congestion information;

[0048] Topology feature extraction module: used to construct a road network topology map based on the road congestion information, and generate a structure-aware state vector that integrates local congestion features and global topology location information by performing structure-aware state encoding on the road network topology map;

[0049] Congestion state prediction module: It is used to input the structure-aware state vector into the improved deep learning prediction model, and to deepen and enhance the features through a graph neural network that dynamically adjusts the graph structure and aggregates node information, and finally outputs the road network congestion state prediction results for multiple future time scales.

[0050] Compared with the prior art, the present invention achieves the following beneficial effects:

[0051] 1. This invention proposes a structure-aware state encoding module to solve the problem of insufficient topological information capture in traditional models. By fusing multi-layer graph convolutional encoding with structural position embedding, it simultaneously captures local congestion features of nodes and global topological position information, generating highly expressive state vectors and realizing in-depth mining of topological information.

[0052] 2. This invention proposes a score-driven node ranking enhancer, which quantifies the contribution score of nodes to global congestion prediction, performs personalized weighted enhancement of node features, alleviates the problem of homogenization of node features, and strengthens the expression of congestion features of key road sections;

[0053] 3. This invention proposes a topological affinity fusion module, which constructs inter-node associations based on cosine similarity, breaking through the limitation of traditional models that only rely on neighboring nodes, and realizing information interaction between nodes that are spatially far apart but closely related to congestion, effectively capturing the long-distance dependence of congestion propagation;

[0054] 4. This invention proposes a strategy-adaptive graph update technique, which dynamically adjusts the topology graph structure based on prediction error feedback and changes in congestion status, and can automatically optimize weight allocation according to different road levels, thereby improving the model's adaptability to dynamic changes in the road network and its prediction accuracy.

[0055] It should be understood that the description in the Summary of the Invention is not intended to limit the key or essential features of the embodiments of the present invention, nor is it intended to restrict the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0056] The above and other features, advantages, and aspects of the various embodiments of the present invention will become more apparent from the accompanying drawings and the following detailed description. The drawings are provided for a better understanding of the invention and are not intended to limit the invention. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein:

[0057] Figure 1 This is a schematic diagram of the main steps of the deep learning-based urban road network traffic congestion prediction method according to Embodiment 1 of the present invention;

[0058] Figure 2 This is a schematic diagram of the overall process of the deep learning-based urban road network traffic congestion prediction method according to Embodiment 1 of the present invention.

[0059] Figure 3 This is a schematic diagram of the structure-aware state encoding process in Embodiment 1 of the present invention;

[0060] Figure 4 This is a flowchart illustrating the congestion state prediction steps in Embodiment 1 of the present invention.

[0061] Figure 5 This is a schematic diagram of the system architecture of the deep learning-based urban road network traffic congestion prediction system according to Embodiment 2 of the present invention.

[0062] Figure 6 This is a schematic diagram of the module of the deep learning-based urban road network traffic congestion prediction system according to Embodiment 2 of the present invention. Detailed Implementation

[0063] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0064] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0065] This invention provides a method and system for predicting urban road network traffic congestion based on deep learning. By performing structured representation on high-resolution congestion image data provided by online map service providers, and deeply integrating the road network topology to construct efficient compressed features, combined with an improved deep learning model, it achieves accurate prediction of short-term congestion status of urban road networks. It balances prediction accuracy, computational efficiency, and dynamic adaptability, and is applicable to large-scale, multi-level road scenarios.

[0066] Example 1

[0067] like Figure 1 and Figure 2 As shown, this embodiment of the invention provides a method 100 for predicting urban road network traffic congestion based on deep learning, including the following steps:

[0068] S1: Data Acquisition and Preprocessing Steps: Acquire image data of urban road network congestion status and perform preprocessing to extract structured road congestion information;

[0069] Step A1 is used to collect and preprocess urban road network congestion data. Specifically:

[0070] We collect image data on urban road network congestion status released by online map service providers, covering all levels of roads including main roads, secondary roads, and branch roads. The data includes key information such as road geometry, congestion level (represented by color coding to indicate smooth traffic, light congestion, heavy congestion, etc.), road direction, and road grade, ensuring full coverage and real-time data.

[0071] The collected raw image data is preprocessed, including image denoising (using Gaussian filtering to remove salt-and-pepper noise), road region segmentation (extracting road pixel regions based on a semantic segmentation model), and congestion level quantization (mapping color codes to continuous values ​​from 0 to 5, where 0 represents smooth traffic and 5 represents severe congestion), laying the foundation for subsequent topology mapping.

[0072] S2: Topology feature extraction step: Construct a road network topology map based on the road congestion information, and generate a structure-aware state vector that integrates local congestion features and global topology location information by performing structure-aware state encoding on the road network topology map;

[0073] Step S2 is used to implement feature extraction based on the road network topology. Taking the topological relationships of the urban road network as the core, it achieves structured compression and feature reconstruction of high-resolution congestion images. Further, the topological feature extraction steps include road network topology map construction and optimization, including: constructing a road network topology map model based on road geometric locations and intersection connections; mapping road attribute information from congestion state images to the road network topology map using a coordinate matching algorithm; adaptively reducing the dimensionality of the initial topology map using a spatial partitioning algorithm, selecting key nodes and edges based on node density and edge importance indicators to form a simplified topology map; extracting multi-dimensional topological features from the simplified topology map, including node congestion degree, neighborhood congestion gradient, and spatial correlation. Node congestion degree is obtained by calculating the average congestion level of adjacent road segments; neighborhood congestion gradient is obtained by calculating the maximum absolute difference between the congestion level of adjacent road segments and the node congestion degree; spatial correlation is obtained by the average neighborhood congestion level of the node weighted by road level. Specifically, it includes the following steps:

[0074] Step S2.1: Construction of road network topology map

[0075] Based on the actual geometric location of roads and the connection relationships of intersections, a road network topology model is constructed. ,in Represents a set of nodes (nodes correspond to road intersections or key road segment endpoints). Represents the set of edges (each edge corresponds to a road segment). It represents the node feature matrix (including the congestion level of the intersection where the node is located, road level weight, etc.). Represents the adjacency matrix ( Represents a node and Direct connection, (Indicates no direct connection). Specifically: A city road network topology model is used to structurally represent the nodes, road segments, node characteristics, and connection relationships of a road network.

[0076] : A set of nodes, each node This corresponds to road intersections or key road segment endpoints in the urban road network (such as the start and end points of road segments, the two ends of bridges, etc.).

[0077] : Set of edges, each edge Corresponding to actual road segments in the urban road network, connecting two adjacent nodes. and .

[0078] Node feature matrix For the number of nodes, For the feature dimension, each element in the matrix Corresponding node The feature vector contains key attributes such as the real-time congestion level of the intersection where the node is located, the level weight of the corresponding road, and the traffic capacity of the intersection.

[0079] Adjacency matrix The number of nodes, matrix elements Represents a node and Directly connected via road sections. This indicates that there is no direct road connection between the two nodes.

[0080] Step S2.2: Image-Topology Mapping

[0081] The attribute information of each road region in the congestion image is mapped to a topological map using a coordinate matching algorithm. Specifically, for any road pixel region in the image... (corresponding road section) Extract its congestion level quantification value. Road direction angle Road grade coefficient (Set the main road to 1.0, the secondary road to 0.7, and the branch road to 0.4), and update the edge feature matrix. .

[0082] Traffic congestion status image and road segment The corresponding road pixel areas are precisely matched with the edges of the topological graph using a coordinate matching algorithm.

[0083] Road section The congestion level quantification value is obtained by color coding of road areas in the image (e.g., smooth traffic corresponds to 0, light congestion corresponds to 1, moderate congestion corresponds to 2, heavy congestion corresponds to 3, and extreme congestion corresponds to 4).

[0084] Road section The road direction angle, that is, the angle between the road's direction of extension and true north (range). ), used to characterize the directional features of a road.

[0085] Road section The road grade coefficient is set according to the road function grade (1.0 for main roads, 0.7 for secondary roads, and 0.4 for local roads) to distinguish the importance weight of roads of different grades.

[0086] Edge feature matrix, storing each road segment The core attributes, such as the congestion level quantification value, road direction angle, and road level coefficient, provide a foundation for subsequent feature extraction.

[0087] Step S2.3: Dimensionality reduction optimization of the topology graph

[0088] The quadtree algorithm is used to adaptively reduce the dimensionality of the initial topology graph: the urban road network space is divided into several quadtree grids, and the node density and edge importance in each grid are calculated. Road grade coefficient Congestion level quantification value), retaining the top importance By identifying the edges and corresponding nodes, redundant nodes and edges are removed to form a simplified topology graph. ,in This significantly reduces the computational complexity of subsequent calculations.

[0089] Node density: The ratio of the number of nodes in a single quadtree grid to the grid area, used to measure the density of the road network within the grid.

[0090] Importance of the side: road segment The importance quantification indicator is calculated as the "road grade coefficient". Congestion Level Quantification Value "" is used to filter out key road sections that are critical for congestion prediction.

[0091] The simplified topological graph after dimensionality reduction. The set of nodes to be retained. The set of edges to be retained. The simplified node feature matrix, This is the simplified adjacency matrix.

[0092] The simplified topology graph has a node count constraint to ensure that the number of nodes after dimensionality reduction does not exceed the number of nodes in the original graph. This significantly reduces the computational complexity of subsequent calculations.

[0093] Step S2.4: Construction of topological feature vectors

[0094] Simplified topology graph Extracting multi-dimensional topological features, including:

[0095] Node congestion level: ( For nodes The degree, For nodes (the set of neighboring nodes).

[0096] Neighborhood congestion gradient: (Measures the degree of difference in congestion status within a node's neighborhood).

[0097] Spatial correlation: (Weighted congestion correlation index that integrates road levels).

[0098] Ultimately, node topology feature vectors are formed. .

[0099] :node The degree of node congestion, For nodes The degree (i.e., the number of edges directly connected to the node). For nodes The set of neighboring nodes, this index represents the node's... The average congestion level of all adjacent road segments reflects the overall congestion level around the node.

[0100] :node The neighborhood congestion gradient is calculated by the node. The maximum absolute difference between the congestion level of adjacent road segments and the congestion level of a node measures the degree of difference in congestion status within the node's neighborhood (the larger the difference, the more uneven the congestion distribution within the neighborhood).

[0101] :node Spatial correlation, through nodes The congestion levels of adjacent road segments are weighted by road level and then averaged. The road level weights are integrated to reflect the congestion correlation strength between the node and its neighbors (higher-level roads have higher weights and have a greater impact on the correlation).

[0102] :node The multi-dimensional topological feature vector integrates three core features: node congestion degree, neighborhood congestion gradient, and spatial correlation degree, comprehensively representing the congestion-related attributes of nodes.

[0103] Step S2.5: Structure-Aware State Encoding

[0104] A structure-aware state encoding module is introduced to fuse local features and global topological information to generate a highly expressive state vector. Further, the structure-aware state encoding process includes: performing multi-layer graph convolution operations on the simplified topological graph to capture higher-order dependencies between nodes and outputting the local neighborhood congestion features of the nodes; calculating a structural location embedding vector based on the shortest path distance between nodes to represent the topological position of the nodes in the global road network; and fusing the local neighborhood congestion features of the nodes output by the graph convolution with the structural location embedding vector to generate a structure-aware state vector that simultaneously contains local congestion features and global topological location information. Figure 3 As shown, the specific steps are as follows:

[0105] Step S2.5.1: Multi-layer graph convolution operation

[0106] By capturing higher-order dependencies of nodes through multi-layer graph convolution, the... The formula for calculating layer graph convolution is:

[0107]

[0108] : No. The output node hidden representation of the layer graph convolution is integrated with the first layer. Layer features and neighborhood node information allow for a deeper capture of higher-order dependencies between nodes. Activation function: used to introduce nonlinear feature transformations to enhance the model's ability to fit complex congestion relationships. Add a self-loop adjacency matrix. To simplify the topology diagram The adjacency matrix, For an identity matrix (diagonal elements are 1, and the rest are 0), adding self-loops can ensure that nodes retain their own features during convolution and avoid loss of their own information. The corresponding degree matrix is ​​a diagonal matrix, with diagonal elements... ( To simplify the number of nodes in the topology graph, the node is represented as... The connection degree after adding self-loops is used to normalize the adjacency matrix to stabilize the training process. Degree matrix The inverse square root, and Combining these methods normalizes the adjacency matrix, avoiding feature scale imbalance caused by differences in node degree. : No. Hidden representation of layer nodes, To simplify the number of nodes in the topology graph, The hidden layer dimension is represented by each row vector in the matrix, which corresponds to a low-dimensional feature representation of a node. : No. The learnable weight matrix of the layer graph convolution is used to adjust the weights of the first layer graph convolution. The hidden features of the layer are linearly transformed to learn the association patterns between node features and neighborhood features. : No. The bias term of layer graph convolution is used to adjust the feature offset after linear transformation, thereby improving the flexibility of model fitting.

[0109] Step S2.5.2: Structural Position Embedding

[0110] Calculate the node structure location embedding based on the shortest path distance, for any node Calculate its relationship with all other nodes in the topology graph. Shortest path length Through the embedding layer Mapped to a low-dimensional vector (Assuming the embedding dimension is 16), the formula is:

[0111]

[0112] :node The structural position embedding vector, The embedding dimension (set to 16) controls the dimension of the location embedding vector, encoding the global topological location information of the node into a low-dimensional dense vector, thus reducing computational complexity while preserving the topological location information.

[0113] Embedding : Embedding layer mapping function, used to map the shortest path length vector of nodes. Convert to low-dimensional embedding vector The learnable parameters of the embedding layer (including weight matrix and bias term) are learned adaptively through training to learn the feature representation of the topological location.

[0114] :node With other nodes in the topology graph The shortest path length, which is the minimum number of road segments connecting two nodes, is used to characterize the relative positional relationship of nodes in the global road network.

[0115] :node The shortest path length vector, containing nodes With all other nodes in the simplified topology graph The shortest path length comprehensively describes the global topological position of the node.

[0116] Step S2.5.3: State Vector Fusion

[0117] The final output of multi-layer graph convolution ( (The number of convolutional layers in the graph is set to 3) and the embedding of the structural location. The vectors are concatenated to generate a structure-aware state vector.

[0118]

[0119] :node The structure-aware state vector is obtained through a concatenation operation. It integrates the final features of multi-layer graph convolution with structural location embedding, and includes both local node congestion features and global topological location information. The final output of multi-layer graph convolution. Let be the number of convolutional layers in the graph (set to 3). For nodes In the The hidden features of the layer represent the local congestion characteristics of a node after the aggregation of information from multiple neighborhoods. :node The structural position embedding vector represents the topological positional relationship of a node in the global road network. Structure-aware state vector dimensional space, For the hidden layer dimension, As the embedding dimension, this vector dimension is the sum of the two, achieving deep fusion of local features and global location information.

[0120] This vector contains both local congestion features of nodes and global topological location information, effectively solving the problem of insufficient topological information capture in traditional models.

[0121] S3: Congestion state prediction step: Input the structure-aware state vector into the improved deep learning prediction model, and deepen and enhance the features through a graph neural network that dynamically adjusts the graph structure and aggregates node information, and finally outputs the road network congestion state prediction results for multiple future time scales.

[0122] Step S3 is used to implement congestion prediction based on an improved deep learning model. For example... Figure 4 As shown, a deep learning prediction model integrating topological adaptation and dynamic updating is constructed, using structure-aware state vectors. The input is a graph neural network, which predicts the road network congestion status for the next 5, 15, and 30 minutes. The feature enhancement and processing are achieved through a graph neural network that dynamically adjusts the graph structure and aggregates node information. This includes dynamic graph convolutional feature extraction, personalized node feature enhancement, enhanced remote information interaction, and dynamic updating of the topology graph structure. Specifically, the steps are as follows:

[0123] Step S3.1: Dynamic Graph Convolution Feature Extraction

[0124] Specifically, the state graph convolution feature extraction includes: deepening the features of the structure-aware state vector through a Dynamic Graph Convolutional Network (DGCN); and capturing dynamic topological changes by adaptively adjusting the adjacency matrix, including: dynamically adjusting the adjacency matrix based on node feature similarity; fusing the similarity matrix after linear transformation of node features with the original adjacency matrix to generate an adaptive adjacency matrix; normalizing the adaptive adjacency matrix; and aggregating neighborhood information through graph convolution operations to achieve dynamic deepening of node features. The details are as follows:

[0125] Dynamic Graph Convolutional Network (DGCN) is used to enhance the features of the structure-aware state vector. It adaptively adjusts the adjacency matrix to capture dynamic topological changes. The formula for calculating layer dynamic graph convolution is:

[0126]

[0127] : No. The adaptive adjacency matrix of the layer dynamic graph convolution dynamically adjusts the connection relationship by fusing node features, accurately capturing the dynamic changes in road network topology (such as changes in the road segment association strength caused by congestion propagation). Activation function, output range is Between these, the similarity of node features is used to convert the similarity of nodes into the weight values ​​of the adjacency matrix, representing the dynamic association strength between nodes. : No. Layer-by-layer batch normalized node feature matrix To perform batch normalization, the model training stability is improved by standardizing the feature distribution, thus avoiding training oscillations caused by differences in feature scale. The learnable parameter matrix of the dynamic adjacency matrix. The dimension of the structure-aware state vector For the hidden layer dimension, (for the embedded dimension) This is the intermediate dimension, used to map node features to the intermediate space to calculate similarity. The node feature similarity calculation process involves obtaining a feature similarity matrix between nodes through matrix multiplication, which reflects the degree of correlation at the node feature level. Element-wise product (Hadamard product) is used to combine the feature similarity matrix with the original simplified adjacency matrix. Element-by-element multiplication preserves the physical connections of the road network while incorporating dynamic association information of node characteristics. The identity matrix is ​​used to add self-loops to ensure that nodes retain their own characteristics and to avoid losing node information when dynamically adjusting the adjacency matrix. : No. The output node feature matrix of the layer dynamic graph convolution integrates dynamic adjacency relationships with the features of the previous layer to achieve more accurate feature deepening. : Activation function, introducing nonlinear transformation to fit complex congestion feature correlations. The corresponding degree matrix is ​​a diagonal matrix, with diagonal elements... ( To simplify the number of nodes in the topology graph, it is used to normalize the dynamic adjacency matrix and stabilize the training process. Degree matrix The inverse square root, and Combining these methods normalizes the dynamic adjacency matrix, avoiding feature scale imbalance caused by differences in node association strength. : No. The input node feature matrix of the layer (which can be a structure-aware state vector or the output of the previous layer's dynamic graph convolution). : No. The learnable weight matrix of the layer dynamic graph convolution is used to perform linear transformations on the input features and learn the feature aggregation pattern under dynamic topology.

[0128] Step S3.2: Node Personalization Enhancement

[0129] Furthermore, a score-driven mechanism is introduced during the personalized enhancement of node features to learn the contribution score of each node to global congestion prediction, and personalized weighted enhancement of node features is performed based on the corrected contribution score to strengthen the feature expression of key road segments. The specific process is as follows: learn the contribution score of each node to global congestion prediction; sort and correct the node features based on the contribution score; and use the corrected score to perform personalized weighted enhancement of node features to strengthen the feature expression of key nodes.

[0130] A score-driven node ranking enhancer is introduced to alleviate the problem of homogeneous node features and strengthen the expression of congestion features in key road segments. Specific steps include:

[0131] Step S3.2.1: Learning Node Contribution Scores

[0132] The contribution scores of nodes to global congestion prediction are learned through fully connected layers:

[0133]

[0134] Node contribution score vectors are used to learn the importance of nodes in global congestion prediction through fully connected layers. For fully connected layer operations, The node feature matrix is ​​the output of the dynamic graph convolution. These are the learnable parameters (including the weight matrix and bias terms) of the fully connected layer. : The dimensional space of the contribution score vector To simplify the number of nodes in the topology graph, each element in the vector The contribution score corresponding to a node. :node The contribution score is a factor that indicates the greater the impact of the node (corresponding to the intersection or key road segment endpoint) on the prediction of global road network congestion status. It is used to further enhance the feature representation of key nodes.

[0135] Step S3.2.2: Sorting and Correcting Score

[0136] Sort the contribution scores in non-decreasing order to obtain the sorted scores. and corresponding index And the fraction distribution is corrected by a fraction modulation control factor:

[0137]

[0138] : Score sorting function, contributing a score vector to each node. Perform a non-decreasing sort and output the sorted score vector. and the index vector of the corresponding original node ( Each element in the table represents the original node number corresponding to the sorted score. The sorted node contribution score vectors are arranged in ascending order of scores, which facilitates subsequent batch correction of score distribution. : Sort index vector, records The original node position corresponding to each score is used for subsequent sorting and recovery of feature vectors. Fractional modulation control factor For fully connected layer operations, Sorted score vector The mean, The learnable parameters (weights + biases) of the fully connected layer are learned to adaptively adjust the score range, avoiding interference from extremely high / low scores on feature weighting. The average contribution score after sorting reflects the overall level of contribution scores of nodes in the current population and provides a benchmark for score correction. The learnable parameters of the fractional modulation control factor are optimized through training to make the fractional correction more suitable for the congestion prediction task. The corrected sorted score vector is obtained by subtracting the score modulation control factor. Adjust the score distribution, reduce the gap between extreme values, and improve the rationality of score weighting.

[0139] Step S3.2.3: Node Feature Weighting Enhancement

[0140] The node features are weighted based on the corrected scores, using the following formula:

[0141]

[0142] Then, using a reverse index Restore the weighted node features to the original topological order:

[0143]

[0144] The sorted node feature matrix, based on the index vector Feature matrix output from convolution of dynamic graph Perform row rearrangement so that the node order of the feature matrix matches the sorted score vector. Consistent. : The original node feature matrix output by dynamic graph convolution, which contains the high-dimensional features of each node after feature enhancement. The weighted node feature matrix. The element-wise product of the corrected sorted score vectors With the sorted feature matrix Element-wise multiplication enables personalized weighting of node features (the higher the contribution score, the greater the feature weight). : Inverse index vector, which records the original node positions corresponding to the sorted node features, and is used to restore the weighted feature matrix to the original topological order. : Restore the weighted feature matrix in its original order by inverting the index right Perform row rearrangement to ensure that the node order of the feature matrix matches the original topology. Figure 1 This facilitates subsequent model processing.

[0145] Step S3.3: Enhanced Remote Information Interaction

[0146] A topological affinity fusion module is introduced to overcome the limitations of traditional graph models that rely solely on neighboring nodes, enabling information exchange between nodes that are spatially distant but closely related to congestion. Further enhancements to long-distance information exchange include: calculating the feature similarity between all nodes to construct a global similarity matrix; based on this global similarity matrix, aggregating information from spatially distant but feature-similar nodes, and preserving the original features through residual connections to capture long-distance dependencies in congestion propagation and update node features. Specifically, the steps are as follows:

[0147] Step S3.3.1: Similarity Matrix Calculation

[0148] Calculate the cosine similarity between node features and construct a similarity matrix. :

[0149]

[0150] Node feature similarity matrix To simplify the number of nodes in the topology graph, matrix elements Represents a node With nodes The feature cosine similarity is used to measure the degree of correlation between congestion features between nodes. Cosine similarity calculation formula, where the numerator is the node With nodes The recovered feature vector , The dot product (measures the consistency of feature orientation), the denominator is the product of the L2 norms of the two feature vectors (normalization factor), and the output range is within... Between these values, the closer the value is to 1, the more similar the features of the two nodes are (the stronger the congestion correlation). :node The weighted feature vector after restoring the original order contains congestion feature information after personalized enhancement of nodes. :node The L2 norm of the recovered feature vector is used to normalize the dot product result, ensuring that the similarity value is within a uniform range.

[0151] Step S3.3.2: Remote Information Aggregation

[0152] Global node information is aggregated based on the similarity matrix, and node features are updated:

[0153]

[0154] The node feature matrix after remote information aggregation integrates the original features with global remote node association information, effectively capturing the long-distance dependence of congestion propagation (such as the indirect impact of main road congestion on remote branch roads). Node feature similarity matrix, matrix elements Represents a node With nodes The degree of correlation between congestion characteristics provides a weighting basis for remote information aggregation. The remote node information aggregation item uses matrix multiplication to sum the features of global nodes by similarity weight, so that the features of each node are integrated into the association information of all remote nodes (the higher the similarity, the greater the contribution weight of the remote node). The weighted feature matrix that restores the original order contains local congestion features after personalized enhancement of nodes, which are retained as residual terms.

[0155] This formula preserves the original features through residual connections while incorporating the association information of remote nodes, effectively capturing the long-distance dependency of congestion propagation.

[0156] Step S3.4: Dynamic update of topology graph structure

[0157] An adaptive graph update mechanism is introduced to dynamically adjust the topology graph structure based on prediction error feedback and changes in congestion status, ensuring the model adapts to the dynamics of the road network. Furthermore, the dynamic update of the topology graph structure includes: calculating the importance weights of each road segment in the road network based on the model's prediction error feedback and real-time congestion status changes; setting dynamically adjusted retention and addition thresholds; and dynamically updating the edge set of the topology graph according to the importance weights, retaining important edges and adding new key connections to adapt to the dynamic changes in the road network topology. Specifically, the steps are as follows:

[0158] Step S3.4.1: Calculation of Policy Behavior Deviation

[0159] Define policy network at any time Behavioral biases, used to measure the difference between current and historical policies:

[0160]

[0161] :time Policy behavior bias measures the degree of difference between current congestion prediction policies and historical policies, and is used to perceive the impact of changes in road network conditions on prediction strategies. :time Distribution of policy behaviors For a moment Model parameters, For a moment The road network status (including congestion levels of each node, road segment correlation strength, etc.) represents the distribution of the model's prediction decisions under the current status. :time Distribution of policy behaviors For a moment Model parameters, For a moment The road network status serves as a historical policy benchmark. :time The road network status includes comprehensive information such as the congestion level of all nodes, the traffic status of road segments, and topological relationships, providing input for policy decision-making. :time The learnable parameters of the model are dynamically updated with training iterations, which determine the model's decision-making mode for the current road network state.

[0162] Step S3.4.2: Edge Importance Assessment

[0163] Based on global rewards With side scheduling frequency Calculate the edges Importance:

[0164]

[0165] :time Section Importance weights, after normalization, range from The values ​​between these values ​​are used to determine the criticality of road segments in congestion prediction, providing a basis for dynamic updates to the topology map. :time Section The scheduling frequency, that is, the time when this road segment is scheduled. The number of times a road segment is identified as a critical path for congestion propagation by the model reflects its actual participation in the spread of congestion. :time The global reward is calculated based on the model's congestion prediction error (the smaller the prediction error, the larger the reward value), and is used to quantify the overall effectiveness of the current policy. : Global normalization factor, for all retained road segments ( To simplify the "scheduling frequency" of the edge set of the topology graph The sum of "global rewards" is used to normalize the importance weights of edges to a uniform range, making it easier to compare them horizontally. The road segments in the simplified topology map correspond to the actual roads in the urban road network, and their importance weight determines whether they are retained in the dynamic topology update.

[0166] Step S3.4.3: Dynamic adjustment of edge sets

[0167] Set dynamic threshold (Edge retention threshold) and (Add threshold to edge) Update edge set :

[0168]

[0169] :time The dynamically updated edge set includes the original road segments that are retained and the newly added potential key road segments, adapting to dynamic changes in the road network topology (such as changes in road segment associations caused by congestion propagation path switching and temporary traffic control). :time Simplify the current set of edges in the topology graph to provide a baseline for edge set updates. :time Section The importance weight is used to determine whether the original road segment should be retained. Edge retention threshold, initially set to 0.6, decays every 100 iterations. Used to screen existing road sections that meet the importance criteria. (If the threshold is retained), the dynamic decay of the threshold can adapt to the refined requirements of road segment selection in the later stage of training. Candidate edges (not in) The policy relevance score of potential road segments (potential connections in the congestion forecast) measures the degree of matching between potential road segments and current congestion forecast policies (such as temporarily opened diversion road segments and potential connections corresponding to new congestion propagation paths). Add a threshold to the edge, initially set to 0.4, which decreases every 100 iterations. Used to filter candidate edges that need to be added ( Then add), and work with the retention threshold to achieve dynamic optimization of the edge set. : Indicates road segment It is a moment Existing road segments in the edge set are included in the retention selection. : Indicates road segment It is not in time Candidate edges in the edge set are included in the new edge selection process.

[0170] Step S3.5: Multi-scale prediction output

[0171] Updated node features The input time-series prediction branch (using bidirectional LSTM) outputs predicted congestion states at different future time scales. Furthermore, a multi-scale prediction output is employed, including: inputting processed node features into the bidirectional LSTM time-series prediction network to capture both forward and backward temporal dependencies of congestion states; and, based on the fused features from the bidirectional LSTM output, outputting predicted road network congestion states at three time scales—5 minutes, 15 minutes, and 30 minutes—through a fully connected layer. The specific process is as follows:

[0172] 1. Bidirectional LSTM forward computation:

[0173]

[0174] The hidden state vector of the bidirectional LSTM forward propagation, based on the features of the current and historical nodes ( ) and the previous forward hidden state ( ) Calculate and capture the positive temporal dependence of congestion status (such as the propagation trend of congestion from upstream road segments to downstream road segments). The computation function for the forward branch of a bidirectional LSTM. The learnable weight parameters for the forward branch (including the weight matrix and bias terms for the input gate, forget gate, cell state, and output gate). The node feature matrix after remote information aggregation serves as the input feature of the bidirectional LSTM, containing local congestion features and global correlation information of nodes. :time The hidden state vector of the forward branch of the bidirectional LSTM conveys historical time sequence information.

[0175] 2. Bidirectional LSTM backward computation:

[0176]

[0177] The hidden state vector of bidirectional LSTM backpropagation, based on the features of the current and future nodes ( ) and the backward hidden state at the next moment ( ) calculate and capture the inverse temporal dependency of congestion status (such as the feedback effect of downstream road congestion on upstream road segment, such as traffic backflow). The computation function for the backward branch of a bidirectional LSTM. These are learnable weight parameters for the backward branch, which are optimized independently of the forward branch weights. :time The hidden state vector of the bidirectional LSTM backward branch conveys future temporal information, enabling bidirectional temporal context fusion.

[0178] 3. Multi-scale prediction fusion:

[0179]

[0180] The fused bidirectional LSTM hidden state vectors are concatenated. By integrating forward and backward hidden states and capturing both forward and reverse temporal dependencies of congestion states, the ability to express temporal features is improved. :future Minute-by-minute congestion level predictions It corresponds to three time scales: short-term, medium-term, and long-term, to meet the congestion prediction needs in different scenarios (such as real-time route planning and traffic resource scheduling). : Calculation of fully connected layers for multi-scale prediction For fully connected layer operations, For the corresponding time scale The learnable parameters (weight matrix + bias term) are used to adapt to the congestion evolution patterns at different time scales through training.

[0181] Step S4: Model parameter optimization

[0182] This invention employs a hyperparameter optimization method based on random search, combined with cross-validation and early stopping mechanisms, to automatically optimize the key hyperparameters of a deep learning prediction model. The specific steps are as follows:

[0183] Step S4.1: Define the hyperparameter search space

[0184] Identify the key hyperparameters that need optimization and their value ranges:

[0185] The hidden dimension h of the graph convolutional layer is searched within the set {32, 64, 128}.

[0186] LSTM layers : Search within the set {1, 2};

[0187] Learning rate α: Random sampling within the logarithmic space [1e-4, 1e-2];

[0188] Dropout rate Search within the set {0.1, 0.2, 0.3};

[0189] Step S4.2: Random Search and Cross-Validation

[0190] Set the total number of trials =50.

[0191] For each trial sy (sy = 1 to 50):

[0192] Randomly select a set of hyperparameter combinations from the above search space. .

[0193] use Configure the deep learning prediction model, and initialize and train the model on the training set.

[0194] The performance of this hyperparameter combination was evaluated using 5-fold cross-validation: Given the strong temporal dependence of traffic data, a forward chaining method was used to divide the training and validation sets, ensuring that the timestamps of the validation set always follow those of the training set, thus simulating real-world prediction scenarios. The training set was divided into 5 parts, with 4 parts used alternately as the training subset and 1 part as the validation subset. The mean squared error of the 5 validations was calculated as the performance metric. .

[0195] Step S4.3: Model Selection and Final Training

[0196] Performance metrics were selected from all 50 trials. Minimal hyperparameter combination .

[0197] use As the final configuration, the deep learning prediction model is retrained on the entire training set. An early stopping mechanism is enabled during training, using the mean squared error (MSE) of the validation set as a monitoring metric. When this metric stops decreasing for 10 consecutive epochs, training is automatically terminated, and the model parameters saved at the minimum point of this metric during the entire training process are restored.

[0198] Step S4.4: Model Evaluation

[0199] Evaluate the predictive performance of the final model on an independent test set to ensure its generalization ability.

[0200] Example 2

[0201] like Figure 5 and Figure 6 As shown, a deep learning-based urban road network traffic congestion state prediction system 200 is used to implement a deep learning-based urban road network traffic congestion state prediction method 100. The system 200 includes:

[0202] Data acquisition and preprocessing module 210: used to acquire image data of urban road network congestion status and perform preprocessing to extract structured road congestion information;

[0203] Topology feature extraction module 220: used to construct a road network topology map based on the road congestion information, and generate a structure-aware state vector that integrates local congestion features and global topology location information by performing structure-aware encoding on the road network topology map;

[0204] Congestion state prediction module 230: It is used to input the structure-aware state vector into the improved deep learning prediction model, and to deepen and enhance the features through a graph neural network that dynamically adjusts the graph structure and aggregates node information, and finally outputs the road network congestion state prediction results for multiple future time scales.

[0205] The deep learning-based urban road network traffic congestion prediction system provided in this embodiment of the invention can execute the deep learning-based urban road network traffic congestion prediction method provided in any of the above embodiments of the invention. It has the corresponding functions and beneficial effects of executing the deep learning-based urban road network traffic congestion prediction system method. For detailed process, please refer to the relevant operations in the aforementioned embodiment one.

[0206] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0207] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined in the embodiments of this application may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown in this application, but is to be accorded the widest scope consistent with the principles and novel features disclosed in the embodiments of this application.

Claims

1. A method for predicting urban road network traffic congestion based on deep learning, characterized in that, Includes the following steps: Data acquisition and preprocessing steps: Acquire image data of urban road network congestion status and perform preprocessing to extract structured road congestion information; Topology feature extraction steps: Construct a road network topology map based on the road congestion information, and generate a structure-aware state vector that integrates local congestion features and global topology location information by performing structure-aware state encoding on the road network topology map; Congestion state prediction steps: The structure-aware state vector is input into the improved deep learning prediction model. The graph neural network that dynamically adjusts the graph structure and aggregates node information is used to deepen and enhance the features. Finally, the prediction results of road network congestion state at multiple future time scales are output.

2. The method according to claim 1, characterized in that, The topology feature extraction step includes road network topology map construction and optimization, specifically including: Construct a road network topology model based on road geometry and intersection connections; The road attribute information in the congestion status image is mapped to the road network topology map using a coordinate matching algorithm; A spatial partitioning algorithm is used to adaptively reduce the dimensionality of the initial topology graph. Key nodes and edges are selected based on node density and edge importance indices to form a simplified topology graph.

3. The method according to claim 2, characterized in that, The topological feature extraction step further includes topological feature vector construction, specifically including: Multi-dimensional topological features are extracted from the simplified topology graph, including node congestion, neighborhood congestion gradient, and spatial correlation. The node congestion level is obtained by calculating the average congestion level of adjacent road segments of the node; The neighborhood congestion gradient is obtained by calculating the maximum absolute difference between the congestion level of the adjacent road segments of the node and the node's congestion degree; The spatial correlation degree is obtained by weighting the average congestion level of the node's neighborhood by road level.

4. The method according to claim 3, characterized in that, The structure-aware state encoding in the topology feature extraction step specifically includes: Multi-layer graph convolution operation is performed on the simplified topology graph to capture the higher-order dependencies of nodes and output the local neighborhood congestion features of the nodes. The structural position embedding vector is calculated based on the shortest path distance between nodes, representing the topological position of the node in the global road network; The local neighborhood congestion features of the nodes output by graph convolution are fused with the structure location embedding vector to generate a structure-aware state vector that simultaneously contains local congestion features and global topological location information.

5. The method according to claim 1, characterized in that, In the congestion state prediction step, feature enhancement and amplification are performed by a graph neural network that dynamically adjusts the graph structure and aggregates node information, including dynamic graph convolution feature extraction, personalized node feature enhancement, remote information interaction enhancement, and dynamic updating of the topology graph structure. The dynamic graph convolution feature extraction specifically includes: The structure-aware state vector is enhanced by using a dynamic graph convolutional network. The adjacency matrix is ​​adaptively adjusted to capture dynamic topological changes. This includes: dynamically adjusting the adjacency matrix based on node feature similarity; fusing the similarity matrix of node features after linear transformation with the original adjacency matrix to generate an adaptive adjacency matrix; normalizing the adaptive adjacency matrix; and aggregating neighborhood information through graph convolution operations to achieve dynamic enhancement of node features.

6. The method according to claim 5, characterized in that, The process of personalized enhancement of node features introduces a score-driven mechanism to learn the contribution score of each node to global congestion prediction, and performs personalized weighted enhancement of node features based on the corrected contribution scores to strengthen the representation of key road segment features, specifically: Learn the contribution score of each node to the global congestion prediction; The node features are sorted and corrected based on the contribution scores; The modified scores are used to personalize the weighting of node features to enhance the feature representation of key nodes.

7. The method according to claim 5, characterized in that, The enhanced remote information interaction specifically includes: Calculate the feature similarity between all nodes and construct a global similarity matrix; Based on the global similarity matrix, information of nodes that are spatially distant but have similar features is aggregated, and the original features are preserved through residual connections to capture long-distance dependencies in congestion propagation and update node features.

8. The method according to claim 5, characterized in that, The dynamic updating of the topology graph structure specifically includes: Based on the model's prediction error feedback and real-time congestion status changes, the importance weight of each road segment in the road network is calculated. Set dynamically adjustable retention and addition thresholds, and dynamically update the edge set of the topology graph according to the importance weights, retaining important edges and adding new key connections to adapt to the dynamic changes in the road network topology.

9. The method according to claim 1, characterized in that, The congestion prediction step employs multi-scale prediction output, specifically including: The processed node features are input into a bidirectional LSTM temporal prediction network to capture both the forward and reverse temporal dependencies of congestion states. Based on the fusion features of bidirectional LSTM output, the predicted values ​​of road network congestion status at three time scales—5 minutes, 15 minutes, and 30 minutes—are output through a fully connected layer.

10. A deep learning-based urban road network traffic congestion prediction system, used to implement the method as described in any one of claims 1 to 9, characterized in that, The system includes: Data acquisition and preprocessing module: used to acquire image data of urban road network congestion status and perform preprocessing to extract structured road congestion information; Topology feature extraction module: used to construct a road network topology map based on the road congestion information, and generate a structure-aware state vector that integrates local congestion features and global topology location information by performing structure-aware state encoding on the road network topology map; Congestion state prediction module: It is used to input the structure-aware state vector into the improved deep learning prediction model, and to deepen and enhance the features through a graph neural network that dynamically adjusts the graph structure and aggregates node information, and finally outputs the road network congestion state prediction results for multiple future time scales.

Images