Urban traffic flow prediction method and system

Abstract

The invention provides an urban traffic flow prediction method and system, and relates to the technical field of intelligent traffic systems, and the method comprises the steps: receiving traffic data, and carrying out the linear transformation processing through a full connection layer; the processed traffic data are input into multiple layers of time-space modules of the same structure for feature extraction, each layer of time-space module is composed of a time convolution module and a mixed time-varying graph module, and time features output by the time convolution module and spatial features output by the mixed time-varying graph module are spliced to form output of the time-space modules; parameterized learning is carried out on the mixed time-varying graph module; performing jump connection on the outputs of all the space-time modules to generate fused space-time features; and processing the fused spatial-temporal characteristics through a ReLU activation function, and transmitting a processing result to a full connection layer to obtain a final prediction result. According to the method, the urban traffic condition and the change trend can be accurately judged, and effective decision support is provided for urban planning and traffic management.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent transportation systems, and in particular, to an urban traffic flow prediction method and system. Background Art

[0002] With the accelerating urbanization process and the sharp increase in the number of automobiles, the construction of infrastructure and intelligent transportation systems has become a common concern in cities around the world. Therefore, it is crucial to build a modern urban intelligent transportation system. As one of the important tasks of intelligent transportation, traffic flow prediction has received extensive attention and become a research hotspot in this field. However, traffic data has complex dynamic spatio-temporal relationships, and achieving accurate prediction is a challenging task. Traffic flow prediction is a classic time series prediction problem, whose purpose is to use historical traffic data to predict future traffic flow and realize the judgment of future traffic trends. In the process of modeling traffic flow, it is necessary to fully consider its temporal correlation, spatial correlation, and spatio-temporal correlation.

[0003] In the prior art, there are statistical analysis models, machine learning models, and deep learning models. Statistical analysis models include methods such as autoregressive integrated moving average ARIMA and vector autoregression VAR. Statistical analysis models have simple structures and convenient calculations, but they need to assume that the traffic state is static and cannot process complex non-linear data. Machine learning models include methods such as support vector regression SVR, support vector machine SVM, and K-nearest neighbor KNN. Although machine learning models can extract non-linear features in traffic data, they rely on manual extraction of traffic features, which will undoubtedly lead to complex and time-consuming modeling. Deep learning models have strong spatio-temporal feature learning capabilities and can better handle complex non-linear relationships. Among them, convolutional neural network CNN and graph convolutional network GCN are widely used in urban traffic flow prediction. How to mine its effective spatial relationship is the key issue in the research of traffic flow prediction methods in recent years. Some research scholars model traffic roads as a grid and use convolutional neural network CNN to learn the spatial interaction between different grids to capture spatio-temporal correlation. However, the structure of large-scale traffic networks in the real world is a non-Euclidean space structure with topology and irregularity, and its spatial structure cannot be fully characterized by CNN. To overcome the limitations of CNN in spatial structure, based on graph convolutional network GCN processes non-Euclidean structured data by aggregating the feature information of adjacent nodes. GCN applies convolutional operations to spatio-temporal graph data with non-Euclidean structures and has powerful spatio-temporal data representation capabilities, which are widely used in urban traffic flow prediction.

[0004] The above methods have contributed to the improvement of traffic flow prediction performance from different perspectives. However, in the process of implementing the present invention, the applicant found that most of the existing models rely on adaptive graphs or predefined graph structures, making it difficult to fully capture complex deep dynamic spatio-temporal features. In addition, traffic flow exhibits complex spatio-temporal correlation characteristics of both local and global aspects in the spatio-temporal dimension, while existing methods hardly consider simultaneously and evenly mining local and global information, which limits the improvement of the model prediction performance and results in insufficient accuracy of existing traffic flow prediction methods. Therefore, how to improve the accuracy of traffic flow prediction has become a technical problem to be urgently solved. Summary of the Invention

[0005] The present invention aims to solve at least one of the technical problems existing in the prior art or related technologies, and discloses an urban traffic flow prediction method and system, which improves the prediction accuracy of traffic flow prediction tasks.

[0006] The first aspect of the present invention discloses an urban traffic flow prediction method, including: Input data preprocessing: receiving traffic data, performing linear transformation processing through a fully connected layer to generate processed traffic data; Multi-layer spatio-temporal module processing: inputting the processed traffic data into multi-layer spatio-temporal modules with the same structure for feature extraction. Each spatio-temporal module consists of a temporal convolutional module and a hybrid time-varying graph module. The temporal features output by the temporal convolutional module and the spatial features output by the hybrid time-varying graph module are concatenated to form the output of the spatio-temporal module; The temporal convolutional module is a Gated TCN model, which is used to capture the dependencies in the time dimension as temporal features; The hybrid time-varying graph module includes a static graph model and a dynamic hypergraph model. The static graph model is a graph neural network that captures long-term stable spatial information, and the dynamic hypergraph model is a dynamic hypergraph neural network that captures short-term spatial correlations and high-order spatial correlations. The adjacency matrix of the static graph and the adjacency matrix of the dynamic hypergraph are fused as spatial features; Parameterized graph learning module processing: performing parameterization processing on the graph adjacency matrices of the static graph model and the dynamic hypergraph model to extract the unidirectional relationships between nodes, so that the static graph model and the dynamic hypergraph model can adaptively learn and adjust the graph structure during the training process; Fusing spatio-temporal features: performing a skip connection operation on the outputs of all spatio-temporal modules to fuse the spatio-temporal features extracted by the multi-layer spatio-temporal modules to generate fused spatio-temporal features; Outputting prediction results: processing the fused spatio-temporal features through a ReLU activation function and transmitting the processing results to a fully connected layer to obtain the final prediction results.

[0007] According to the urban traffic flow prediction method disclosed by the present invention, preferably, the traffic data is a traffic flow map.

[0008] According to the urban traffic flow prediction method disclosed by the present invention, preferably, the calculation process of the Gated TCN model includes:

[0009] Capture temporal dynamics using dilated causal convolutional networks, which are defined as:

[0010]

[0011] where \(x\) represents a one-dimensional time series, \(k\) represents the size of the convolutional kernel, \(s\) represents the index of the convolutional kernel, \(t\) represents the time step, \(d\) represents the dilation factor, \(f(t)\) represents the value of the convolutional kernel at time step \(t\), and \(f(s)\) represents the value of the convolutional kernel at time offset \(s\).

[0012] The process of capturing dependencies in the time dimension is defined as:

[0013] .

[0014] where \(\tanh\) represents the hyperbolic tangent activation function, \(\sigma\) represents the sigmoid activation function, \(h\) represents the output feature, \(b\) represents the bias parameter of \(\tanh\), \(g(t)\) represents the learnable convolutional filter parameter, and \(c\) represents the bias parameter of \(\sigma\).

[0015] According to the urban traffic flow prediction method disclosed in the present invention, preferably, the method for constructing a static graph model includes: constructing a geospatial graph adjacency matrix based on the Euclidean distance ; supplementing additional hidden points of interest based on the input; the static graph structure is defined as:

[0016]

[0017] where \(\Delta A\) represents the learnable graph adjacency matrix.

[0018] The geospatial graph adjacency matrix is obtained based on the Euclidean distance between node pairs and is defined as follows:

[0019]

[0020] where represents the element in the \(i\)-th row and \(j\)-th column of the geospatial adjacency matrix, represents the Euclidean distance between node \(i\) and node \(j\), represents the threshold.

[0021] According to the urban traffic flow prediction method disclosed in the present invention, preferably, the method for constructing a dynamic hypergraph model includes: for the historical traffic state feature \(X\), measuring the similarity between the time series of \(N\) nodes based on the Euclidean distance of vectors; performing K-NN clustering on each node, regarding one cluster as a hyperedge, and connecting multiple nodes in the cluster by this hyperedge, and then obtaining a normalized hypergraph adjacency matrix according to the constructed hypergraph structure, which is defined as follows:

[0022]

[0023]

[0024] 。

[0025] Among them, represents the degree of node , represents the weight of hyperedge , represents the degree matrix of hypergraph nodes, which is a diagonal matrix, and H represents the node-hyperedge incidence matrix of the hypergraph, represents the normalized hypergraph adjacency matrix, represents the degree of hyperedge , and h(v, e) represents the incidence function value between node and hyperedge , D ε represents the matrix of hypergraph edges, and W represents the hyperedge weight.

[0026] According to the urban traffic flow prediction method disclosed in the present invention, the parameterized graph learning module is used to optimize and update the graph structure, provide a better graph structure for the multi-layer spatio-temporal module, generate a new fused relationship matrix A2 as the input of the spatio-temporal module, and the multi-layer spatio-temporal module is based on this graph structure to perform deeper feature extraction and prediction. The calculation process of the parameterized graph learning module specifically includes:

[0027] First, generate the coarse affinity matrix A1:

[0028]

[0029] Among them, M1 and M2 represent learnable parameter matrices;

[0030] represents diagonalizing; represents a learnable parameter;

[0031] Since is a skew-symmetric matrix, the activation function ReLU sets half of the diagonal positions and other positions to zero to enhance sparsity;

[0032] Subsequently, fuse the old relationship and the new spatial dependence relationship:

[0033]

[0034]

[0035] Among them, denotes: the fusion weight; A2 denotes: the new relationship matrix after fusion;

[0036] denotes: the non-linear activation function;

[0037] denotes: extract fusion and features, denotes the old relationship matrix of the coarse affinity matrix A1.

[0038] The second aspect of the present invention discloses an urban traffic flow prediction system, including: a memory for storing program instructions; a processor for calling the program instructions stored in the memory to implement the urban traffic flow prediction method of any of the above technical solutions.

[0039] Compared with the prior art, the beneficial effects of the present invention at least include: the dynamic hypergraph based on hybrid time-varying proposed by the present invention can simultaneously learn the global stable spatial association and local dynamic spatial association between different nodes in the road network, improving the prediction accuracy. The hybrid time-varying graph module consists of a dynamic hypergraph capable of capturing complex high-order correlations and a static graph capable of describing long-term stable spatial information. The dynamic hypergraph utilizes the dynamic correlations at the edges of the traffic flow graph, which helps to characterize the complex multi-node interaction relationships in the traffic system; secondly, inspired by the expectation maximization algorithm, the parameters of the prediction network module and the graph learning module are alternately trained and optimized; then, the long sequence is effectively processed by dilated causal convolution, and increasing the layer depth allows an exponentially larger receptive field. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 Shows a schematic flowchart of an urban traffic flow prediction method according to an embodiment of the present invention.

[0041] Figure 2 Shows a schematic block diagram of an urban traffic flow prediction system according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0042] In order to more clearly understand the above objects, features and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Many specific details are set forth in the following description in order to fully understand the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the present invention is not limited by the limitations of the specific embodiments disclosed below.

[0043] As Figure 1 shown, according to an embodiment of the invention, an urban traffic flow prediction method is disclosed, including:

[0044] Step 1, Input data preprocessing: Receive traffic data, perform linear transformation processing through a fully connected layer, and generate processed traffic data;

[0045] Step 2, Multi-layer spatio-temporal module processing: The processed traffic data is input into multi-layer spatio-temporal modules with the same structure for feature extraction. Each spatio-temporal module consists of a temporal convolutional module and a hybrid time-varying graph module. The temporal features output by the temporal convolutional module and the spatial features output by the hybrid time-varying graph module are concatenated to form the output of the spatio-temporal module; The temporal convolutional module is a Gated TCN model, which is used to capture dependencies in the time dimension as temporal features; The hybrid time-varying graph module includes a static graph model and a dynamic hypergraph model. The static graph model is a graph neural network that captures long-term stable spatial information, and the dynamic hypergraph model is a dynamic hypergraph neural network that captures short-term spatial correlations and high-order spatial correlations. The adjacency matrix of the static graph and the adjacency matrix of the dynamic hypergraph are fused as spatial features;

[0046] Step 3, Parameterized graph learning module processing: Parametrize the graph adjacency matrices of the static graph model and the dynamic hypergraph model to extract the unidirectional relationships between nodes, so that the static graph model and the dynamic hypergraph model can adaptively learn and adjust the graph structure during the training process;

[0047] Step 4, Fuse spatio-temporal features: Perform a skip connection operation on the outputs of all spatio-temporal modules to fuse the spatio-temporal features extracted by the multi-layer spatio-temporal modules and generate fused spatio-temporal features;

[0048] Step 5, Output prediction results: Process the fused spatio-temporal features through the ReLU activation function and transmit the processing results to the fully connected layer to obtain the final prediction results.

[0049] According to the above embodiments, further, Step 2 specifically includes:

[0050] Static graph construction: The static graph model is used to describe long-term stable spatial information. First, construct a geographical spatial graph adjacency matrix based on the Euclidean distance , and then, based on the input, supplement additional hidden points of interest (POIs) to avoid the cumbersome manual annotation process. The long-term static graph structure is defined as:

[0051] The geographical spatial graph adjacency matrix is generally obtained based on the Euclidean distance between node pairs and can be defined as:

[0052] .

[0053] Dynamic hypergraph construction: A hypergraph is constructed by mining relevant information in node attributes from the input, enabling the capture of short-term and high-order spatial correlations. For historical traffic state features , the similarity between time series of N nodes is measured based on the Euclidean distance of vectors, and then each node is subjected to K-NN clustering. A cluster can be regarded as a hyperedge, and multiple nodes in the cluster are connected by this hyperedge. Then, according to the constructed hypergraph structure, a normalized hypergraph adjacency matrix can be obtained, which can be defined as:

[0054]

[0055]

[0056]

[0057] Multi-graph structure fusion mechanism: The long-term stable graph adjacency matrix and the short-term time-varying hypergraph adjacency matrix are fused and then normalized to obtain an optimal graph adjacency matrix , and the formula is defined as:

[0058]

[0059] where Norm represents the normalization operation.

[0060] According to the above embodiments, further, in step 2, the temporal convolutional module uses a gated TCN (Gated TCN) to extract temporal dimension features. Different from using RNN-based methods in previous studies to model node temporal correlations, the present invention uses a dilated causal convolutional network to capture temporal dynamics. RNN-based methods process long sequences recursively, resulting in long training computation time and being prone to gradient vanishing or explosion problems. However, the dilated causal convolutional network can perform parallel computations and exponentially expand the receptive field by increasing the layer depth, enabling it to efficiently process long sequences. The dilated causal convolution can be defined as:

[0061]

[0062] Using Gated TCN to capture complex temporal dependencies, which can be defined as:

[0063] .

[0064] According to the above embodiments, further, step 3 specifically includes: The parametric graph learning module aims to extract the unidirectional relationships between nodes and alternately trains and optimizes the prediction network module. Alternate training means initializing the graph structure and parameters first, and then alternately performing parameter optimization and graph structure update. In the parameter optimization stage, the graph structure is fixed and other parameters are updated through backpropagation; in the graph structure update stage, the parameters are fixed and the graph structure is updated using the designed formula. Repeat this process until the model converges. The specific process includes:

[0065] First, generate a coarse affinity matrix:

[0066]

[0067] Among them, M1 and M2 represent: learnable parameter matrices, represents: a learnable parameter, represents diagonalizing Since is a skew-symmetric matrix, the activation function ReLU sets half of the diagonal positions and other positions to zero to enhance sparsity.

[0068] Subsequently, fuse the old relationships and new spatial dependencies, as shown in the formula:

[0069]

[0070]

[0071] Among them, represents: the fusion weight; A2 represents: the new relationship matrix after fusion; represents: a non-linear activation function; represents: extracting and fusing and features, represents the old relationship matrix of the coarse affinity matrix A1.

[0072] The parametric graph learning module ensures the sparsity of the generated matrix and does not limit the number of related nodes for each node. It should be noted that some nodes may have a strong correlation with other nodes, while some other nodes are relatively isolated. Therefore, the parametric graph learning module can more effectively generate the relationships between research nodes. During the iteration process, prior information and newly generated information can be continuously integrated, thereby significantly improving the convergence speed of the model, which is crucial for optimizing the model performance.

[0073] Such as Figure 2As shown, according to another embodiment of the present invention, a urban traffic flow prediction system 200 is also disclosed, including: a memory 201 for storing program instructions; a processor 202 for calling the program instructions stored in the memory to implement the urban traffic flow prediction method as in the above embodiment.

[0074] In summary, the present invention proposes a time-varying spatial network for traffic flow prediction. The time-varying spatial network realizes the information fusion of graph convolutional network and dynamic hypergraph, capturing long-term static and short-term dynamic time-varying spatial correlations from global and local perspectives respectively. The time-varying spatial correlation from the local perspective is constructed using a hypergraph that can capture complex high-order correlations. The present invention also uses a parameterized graph learning module to parameterize the adjacency matrix of the graph, enabling the model to adaptively learn and adjust the graph structure during training, while enhancing the sparsity of the generated matrix and improving the performance and generalization ability of the model. In addition, the present invention also constructs dilated causal convolution in the time series module to increase the layer depth and expand the receptive field. This framework can fully capture the spatial correlation and its long-time dependence between regions, thus achieving accurate traffic flow prediction.

[0075] All or part of the steps in the various methods of the above embodiments can be completed by a program controlling related hardware. This program can be stored in a readable storage medium, and the storage medium includes read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc memories, magnetic disk memories, tape memories, or any other readable medium capable of carrying or storing data.

[0076] The above are only the preferred embodiments of the present invention and are not used to limit the present invention. For those skilled in the art, the present invention can have various changes and modifications. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for predicting urban traffic flow, characterized in that, Including: Input data preprocessing: Receive traffic data, perform linear transformation processing through a fully connected layer, and generate processed traffic data; Multi-layer spatio-temporal module processing: The processed traffic data is input into spatio-temporal modules with the same structure in multiple layers for feature extraction. Each spatio-temporal module consists of a temporal convolutional module and a hybrid time-varying graph module. The temporal features output by the temporal convolutional module and the spatial features output by the hybrid time-varying graph module are concatenated to form the output of the spatio-temporal module. The temporal convolutional module is a Gated TCN model, which is used to capture dependencies in the time dimension as the temporal features. The hybrid time-varying graph module includes a static graph model and a dynamic hypergraph model. The static graph model is a graph neural network that captures long-term stable spatial information, and the dynamic hypergraph model is a dynamic hypergraph neural network that captures short-term spatial correlations and high-order spatial correlations. The adjacency matrix of the static graph and the adjacency matrix of the dynamic hypergraph are fused as the spatial features; Parameterized graph learning module processing: Parametrize the graph adjacency matrices of the static graph model and the dynamic hypergraph model to extract the unidirectional relationships between nodes, so that the static graph model and the dynamic hypergraph model can adaptively learn and adjust the graph structure during the training process; Fuse spatio-temporal features: Perform a skip connection operation on the outputs of all the spatio-temporal modules to fuse the spatio-temporal features extracted by the multi-layer spatio-temporal modules and generate fused spatio-temporal features; Output prediction result: Process the fused spatio-temporal features through a ReLU activation function and transmit the processing result to a fully connected layer to obtain the final prediction result.

2. The urban traffic flow prediction method according to claim 1, wherein The traffic data is a traffic flow map.

3. The urban traffic flow prediction method according to claim 1, characterized in that The calculation process of the Gated TCN model includes: Use a dilated causal convolutional network to capture time dynamics. The dilated causal convolutional network is defined as: where x represents: a one-dimensional time series; k represents: the size of the convolutional kernel; s represents: the index of the convolutional kernel; t represents: the time step; d represents: the dilation factor; f(t) represents: the value of the convolutional kernel at time step t; f(s) represents: the value of the convolutional kernel at the time offset s; The process of capturing dependencies in the time dimension is defined as: where tanh represents: the hyperbolic tangent activation function; σ represents: the sigmoid activation function; h represents: the output feature; b represents: the bias parameter of tanh; g(t) represents: the learnable convolutional filter parameter; c represents: the bias parameter of σ.

4. The urban traffic flow prediction method according to claim 1, characterized in that, The construction method of the static graph model includes: Constructing the Adjacency Matrix of Geospatial Graph Based on Euclidean Distance ; Static graph structure Defined as: where ΔA represents: the learnable graph adjacency matrix; The geospatial graph adjacency matrix is obtained based on the Euclidean distance between node pairs and is defined as follows: Among them, represents the element in the i-th row and j-th column of the geospatial adjacency matrix; Denote: the Euclidean distance between node i and node j; Indicates: Threshold value.

5. The urban traffic flow prediction method according to claim 1, wherein The construction method of the dynamic hypergraph model includes: For the historical traffic state feature X, measure the similarity between the time series of N nodes based on the Euclidean distance of vectors; Perform K-NN clustering on each node, and regard a cluster as a hyperedge. Multiple nodes in the cluster are connected by this hyperedge, Then, according to the constructed hypergraph structure, obtain the normalized hypergraph adjacency matrix, which is defined as follows: Among them, denotes: the degree of a node ; Denotes: hyperedge weight of; Denote: The degree matrix of a hypergraph node, which is a diagonal matrix; H represents: the node-hyperedge incidence matrix of the hypergraph Denote: the normalized hypergraph adjacency matrix; Denote: hyperedge Degree of; h(v, e) represents: the node and the hyperedge associated function value; D ε Denotes: the matrix of hypergraph edges; W represents: hyperedge weight.

6. The urban traffic flow prediction method according to claim 1, wherein The calculation process of the parameterized graph learning module specifically includes: First, generate a coarse affinity matrix A1: Among them, M1 and M2 represent: learnable parameter matrices; Indicates: Diagonalize; Indicates: Learnable parameters; Since is a skew-symmetric matrix, the ReLU activation function sets the diagonal positions and half of the other positions to zero to enhance sparsity; Subsequently, fuse the old relationships and new spatial dependency relationships: Among them, denotes: fusion weight; A2 denotes: the new fused relationship matrix; Denote: non-linear activation function; Indicates: extraction and fusion and characteristics of represents the old relational matrix of the coarse affinity matrix A1.

7. A urban traffic flow prediction system, characterized in that, Including: A memory for storing program instructions; A processor for calling the program instructions stored in the memory to implement the urban traffic flow prediction method according to any one of claims 1 to 6.

Images