Dual-end graph convolutional traffic flow prediction method with graph learning

Abstract

A double-end graph convolution traffic flow prediction method with graph learning solves the problem of how to improve the traffic flow prediction accuracy and belongs to the technical field of traffic flow prediction.The method comprises the following steps: S1, acquiring real-time traffic flow data; S2, inputting the acquired real-time traffic flow data into a prediction model to obtain predicted traffic flow; the prediction model comprises a feature extraction layer, a prediction layer, L dynamic graph learning layers, L gated time convolution layers and L double-end graph convolution layers; the method does not use a predefined static graph structure, but learns the dynamic and periodic characteristics of the traffic data itself through the dynamic graph learning layer; the gated time convolution layer and the double-end graph convolution layer are respectively used for capturing the time correlation and spatial correlation of the traffic sequence itself, so that the prediction result is as close to the real value as possible.

Description

Technical Field

[0001] This invention relates to a bi-terminal graph convolutional traffic prediction method with graph learning, belonging to the field of traffic flow prediction technology. Background Technology

[0002] The growth rate of vehicles in modern cities far exceeds the mileage of newly constructed roads, leading to a series of problems such as road congestion and environmental pollution, causing significant inconvenience to people's lives. The best solution to this problem is to develop Intelligent Traffic Systems (ITS), which utilize traffic guidance technology to improve the efficiency of the road network. This involves suggesting optimal driving routes for vehicles based on the current and future traffic conditions of the road network, thereby ensuring a more even distribution of traffic flow across the network and maximizing the functionality of each road.

[0003] A key variable reflecting the state of the road network is traffic flow, which is the number of vehicles passing through a specific road section within a certain time period. An excellent traffic guidance system needs to make guidance suggestions based on road traffic flow in the short term (5-15 minutes). However, due to the nonlinearity and noise interference of short-term traffic flow data, its patterns are difficult to grasp, making short-term traffic flow prediction a persistent challenge.

[0004] Traffic flow prediction has always been an important research topic in spatiotemporal data mining and intelligent transportation systems, and a research hotspot in the global transportation field. Early prediction models mainly included historical averages, linear regression, and time series analysis, but their prediction accuracy was low and their adaptability was limited. In recent years, models based on traffic simulation, chaos theory, neural networks, and Support Vector Machines (SVMs) have been widely studied. Machine learning methods, due to their strong theoretical framework and good prediction performance, are increasingly becoming popular reference models. Existing methods model the changes and hidden correlations between nodes. They typically capture dynamic relationships by sharing adjacency matrices, which introduces high temporal and spatial complexity, severely limiting the ability to solve large graphs and leading to inaccurate traffic flow predictions. Summary of the Invention

[0005] To address the issue of improving the accuracy of traffic flow prediction, this invention provides a bi-terminal graph convolutional traffic flow prediction method with graph learning.

[0006] The present invention provides a bi-terminal graph convolutional traffic flow prediction method with graph learning, comprising:

[0007] S1. Obtain real-time traffic flow data;

[0008] S2. Input the acquired real-time traffic flow data into the prediction model to obtain the predicted traffic flow;

[0009] The prediction model includes a feature extraction layer, a prediction layer, L dynamic graph learning layers, L gated temporal convolutional layers, and L bi-terminal graph convolutional layers;

[0010] The extraction layer is used to extract features from real-time traffic flow data, which serve as the initial feature H. 0 ;

[0011] A block consists of a dynamic graph learning layer, a gated temporal convolutional layer, and a dual-ended graph convolutional layer. L blocks are concatenated together. The output of the gated temporal convolutional layer in each block and the output of the Lth block are simultaneously input into the prediction layer. The prediction layer outputs the traffic flow for the next time step.

[0012] In the l-th block, l = {1, 2, ..., L}, the output H of the (l-1)-th block is... l-1 Simultaneously serving as the input to the l-th gated temporal convolutional layer and the l-th dynamic graph learning layer, the input in the first block is the initial feature H. 0 The l-th gated temporal convolutional layer is used to capture the temporal correlation features of traffic flow changes over time for each road segment, and to obtain the output hidden state of the gated temporal convolutional layer. And hide the output state The input is fed into the l-th bipolar graph convolutional layer; simultaneously, the dynamic graph learning layer is used to obtain the adjacency matrix A by learning the continuous graph and the alternate graph. l The input is then fed into the l-th binary delimited graph convolutional layer; the l-th binary delimited graph convolutional layer is used to determine the hidden state based on the output. and adjacency matrix A l Capture the spatial correlation characteristics of traffic flow changes over time between different road segments to obtain H l .

[0013] Preferably, the dynamic graph learning layer is used to obtain the adjacency matrix A by learning the continuous graph and the alternate graph. l include:

[0014]

[0015]

[0016]

[0017] A l =ReLU(tanh(βA) 12 ))

[0018] in, and Let represent two learnable nodes embedded in the l-th block, d be the size of the embedded learnable nodes, and N be the number of nodes. and This corresponds to the linear layer parameter, A. 12 is the cosine similarity matrix, ReLU(·) represents the activation function; β is a hyperparameter used to control the saturation rate of the activation function.

[0019] Preferably, the gated temporal convolutional layer includes a dilated convolutional layer and a dilated inception layer:

[0020] The output H of the (l-1)th block l-1 Simultaneously input to the dilated convolutional layer and the dilated inception layer

[0021] The output of the dilated convolutional layer is The output of the void inception layer is Output hidden state for:

[0022]

[0023] ⊙ is the element-wise multiplication operator, and σ(·) represents the sigmoid function.

[0024] As a preferred option, the output of the dilated convolutional layer is:

[0025]

[0026] in, This indicates the kernel size of the l-th layer, and ★ represents dilated convolution operation.

[0027] As a preferred embodiment, the perforated inception layer includes a 1×1 convolutional layer, four different perforated convolutional layers, four ConvGRU layers, and one splicing layer;

[0028] The output H of the (l-1)th block l-1 Input into a 1×1 convolutional layer, and obtain Will The outputs are passed to four different dilated convolutional layers to obtain the corresponding outputs. Will The inputs are fed into four ConvGRU layers, and the four ConvGRU layers output from each layer respectively. splicing layer obtained concat(·) means concatenation.

[0029] As a preferred option, The inputs are fed into four ConvGRU layers, and the four ConvGRU layers output from each layer respectively. The methods include:

[0030]

[0031]

[0032]

[0033]

[0034] k = 1, 2, 3, 4, ⊙ is the element-wise multiplication operator, and σ(·) represents the sigmoid function. This represents the learnable parameters, and * indicates the convolution operation.

[0035] Preferably, the dual-ended graph convolutional layer includes an inflow graph convolutional layer and an outflow graph convolutional layer;

[0036] The output of the inflow graph convolutional layer How to obtain:

[0037] Obtain the adjacency matrix A l degree matrix elements on the diagonal Represents the adjacency matrix A l Features of the i-th row and j-th column;

[0038] Obtain matrix For matrix Update rowsum(·) represents the sum of the values ​​in each row;

[0039] get The subscripts m = 1, ..., M, where M represents the step number, and γ is a hyperparameter used to control the proportion of original features retained. This represents a feature filter implemented using a 1×1 convolution; This represents the adjacency matrix at the m-th step of the inflow end. Higher-order features;

[0040] For each step obtained Obtain by splicing

[0041] Output of the outflow graph convolutional layer How to obtain:

[0042] Obtain the feature information of the m-th step in the outflow end.

[0043] right Obtain by splicing

[0044] This represents a feature filter implemented using a 1×1 convolution;

[0045] Output of a double-ended graph convolutional layer

[0046] Preferably, the prediction layer includes two standard convolutional layers, the first of which aggregates L+1 inputs to obtain... The second standard convolutional layer will Transform dimensions to S1 acquires real-time traffic flow data X∈R H×N×C H represents the number of time steps in the input traffic flow data sequence, N represents the number of nodes, and C represents the number of features.

[0047] The beneficial effect of this invention is that it does not use a predefined static graph structure, but instead learns the dynamic and periodic characteristics inherent in traffic data through a dynamic graph learning layer. Gated temporal convolutional layers and bi-terminal graph convolutional layers are used to capture the temporal and spatial correlations inherent in traffic sequences, respectively, making the prediction results as close as possible to the true values. Attached Figure Description

[0048] Figure 1 This is a schematic diagram illustrating the principle of the present invention;

[0049] Figure 2 This is a schematic diagram of the void inception layer. Detailed Implementation

[0050] 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.

[0051] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0052] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but this is not intended to limit the scope of the invention.

[0053] This embodiment of a bi-terminal graph convolutional traffic flow prediction method with graph learning includes:

[0054] Step 1: Obtain real-time traffic flow data;

[0055] Step 2: Input the acquired real-time traffic flow data into the prediction model to obtain the predicted traffic flow;

[0056] The prediction model includes a feature extraction layer, a prediction layer, L dynamic graph learning layers, L gated temporal convolutional layers, and L bi-terminal graph convolutional layers;

[0057] The extraction layer is used to extract features from real-time traffic flow data, which serve as the initial features H. 0 ;

[0058] A block consists of a dynamic graph learning layer, a gated temporal convolutional layer, and a dual-ended graph convolutional layer. L blocks are concatenated together. The output of the gated temporal convolutional layer in each block and the output of the Lth block are simultaneously input into the prediction layer. The prediction layer outputs the traffic flow for the next time step.

[0059] In the l-th block, l = {1, 2, ..., L}, the output H of the (l-1)-th block is... l-1 Simultaneously serving as the input to the l-th gated temporal convolutional layer and the l-th dynamic graph learning layer, the input in the first block is the initial feature H. 0 The l-th gated temporal convolutional layer is used to capture the temporal correlation features of traffic flow changes over time for each road segment, and to obtain the output hidden state of the gated temporal convolutional layer. And hide the output state The input is fed into the l-th bipolar graph convolutional layer; simultaneously, the dynamic graph learning layer is used to obtain the adjacency matrix A by learning the continuous graph and the alternate graph. l The input is then fed into the l-th biplot convolutional layer; the l-th biplot convolutional layer is used to determine the hidden state based on the output. and adjacency matrix A l Capture the spatial correlation characteristics of traffic flow changes over time between different road segments to obtain H l .

[0060] The overall framework of the prediction model in this implementation is as follows: Figure 1 As shown, there are three key components: a dynamic graph learning layer, a gated temporal convolutional layer, and a double-ended graph convolutional layer. Unlike previous prediction methods, this implementation does not use a predefined static graph structure. Instead, it learns the dynamic and periodic characteristics of the traffic data itself through a dynamic graph learning layer. The gated temporal convolutional layer and the double-ended graph convolutional layer are used to capture the temporal and spatial correlations inherent in the traffic sequences, respectively.

[0061] The goal of the dynamic graph learning layer is to construct a graph adjacency matrix to model the changes and hidden correlations between nodes. Existing methods typically capture dynamic relationships by sharing the adjacency matrix, which introduces high time and space complexity, severely limiting the model's ability to solve large graphs. This implementation's dynamic graph learning layer addresses the learning problem of dynamic graphs by learning from both continuous and alternate graphs. and Let represent the learnable node embeddings embedded in the l-th block, where l = {1, 2, ..., L} and d is the size of the embedding. Then the l-th dynamic graph learning layer can be represented as:

[0062]

[0063]

[0064]

[0065] A l =ReLU(tanh(βA) 12 ))

[0066] in, and This represents the embedding of two learnable nodes in the l-th block, where d is the size of the embedded learnable nodes, and N represents the number of nodes. and This corresponds to the linear layer parameter, A. 12 It is the cosine similarity matrix, ReLU(·) denotes the activation function; β is a hyperparameter used to control the saturation rate of the activation function, and A l Represents the adjacency matrix;

[0067] To maintain the sparsity of the adjacency matrix, in A l The above uses a top-τ masking scheme, retaining only the τ nearest neighbors of each node. The masked adjacency matrix A... l It will be fed into the double-ended graph convolutional layer.

[0068] The gated temporal convolutional layer captures the temporal correlation of traffic flow changes over time for each road segment. It considers the diverse temporal characteristics of each node in the graph over time. This implementation designs a dilated convolutional layer and a dilated inception layer to extract the diverse temporal correlations, and then uses a gating structure to integrate the two modules, i.e., using the tanh and sigmoid functions to map the results of the two components and adjust their outputs. In the l-th gated temporal convolutional layer, this implementation uses the output H of the previous block... l-1 The data are fed into the dilated convolutional layer and the dilated inception layer, respectively, to obtain the following outputs:

[0069] The output H of the (l-1)th block l-1 Simultaneously input to both the dilated convolutional layer and the dilated inception layer, the output of the dilated convolutional layer is... The output of the void inception layer is Output hidden state for:

[0070]

[0071] ⊙ is the element-wise multiplication operator, and σ(·) represents the sigmoid function.

[0072] In this embodiment, a tanh activation function follows the dilated convolutional layer to perform the filtering task, and the output of the dilated convolutional layer is:

[0073]

[0074] in, This indicates the kernel size of the l-th layer, and ★ represents dilated convolution operation.

[0075] The size of the convolutional kernel is a key factor affecting model performance. Considering the short-term nature of traffic data, its periodic patterns typically exhibit 7, 12, and 60 cycles. Therefore, this implementation uses a 1×7 filter with a dilatancy coefficient of q to learn temporal correlations. As the dilatancy coefficient increases exponentially, the receptive field of the model in this implementation expands continuously through stacked convolutional layers.

[0076] The hollow inception layer in this implementation can simultaneously learn short-term and long-term periodic patterns in traffic data. Then, a sigmoid activation function is used to adjust the information flow passed from the filter to the next module. For example... Figure 2 As shown, the hole inclusion layer includes a 1×1 convolutional layer, four different hole convolutional layers, four ConvGRU layers, and one splicing layer;

[0077] The output H of the (l-1)th block l-1 Input to a 1×1 convolutional layer, and obtain Will The outputs are passed to four different dilated convolutional layers to obtain the corresponding outputs. Will The inputs are fed into four ConvGRU layers, and the four ConvGRU layers output from each layer respectively. splicing layer obtained concat(·) means concatenation.

[0078] To capture both short-term and long-term temporal patterns in the sequence, this implementation employs four different dilated convolutional layers, with corresponding kernel sizes of 1×2, 1×3, 1×6, and 1×7, and a dilation coefficient of q. The input is H. l -1 First, a 1×1 convolutional layer is used to obtain the variables. Then, respectively The output is passed to four dilated convolutional layers, which then produce the corresponding outputs. Next, the outputs of these dilated convolutional layers are fed into a ConvGRU layer to obtain the corresponding hidden states.

[0079]

[0080]

[0081]

[0082]

[0083] k = 1, 2, 3, 4, ⊙ is the element-wise multiplication operator, and σ(·) represents the sigmoid function. This represents the learnable parameters, and * indicates the convolution operation. The outputs of all ConvGRU layers are concatenated as follows:

[0084]

[0085] The bi-terminal graph convolutional layer captures the spatial correlation of traffic flow changes over time between different road segments. It considers two node-level features for each node in the graph: inflow and outflow information. Accordingly, this implementation designs two dynamic graph convolutions to capture these features. The bi-terminal graph convolutional layer includes an inflow end-graph convolutional layer and an outflow end-graph convolutional layer, and a gate structure is used to integrate the two parts.

[0086] Taking capturing node-level inflow features as an example, given input A and A' correspond to the outputs of the l-th gated temporal convolutional layer and the l-th dynamic graph learning layer, respectively. First, this implementation uses M-step propagation to derive the output features Z. in .

[0087] Then, the features captured in each step are filtered to obtain the output of the corresponding inflow graph convolutional layer. How to obtain:

[0088] Obtain the adjacency matrix A l degree matrix elements on the diagonal Represents the adjacency matrix A l Features of the i-th row and j-th column;

[0089] Obtain matrix For matrix Update rowsum(·) represents the sum of the values ​​in each row;

[0090] get The subscripts m = 1, ..., M, where M represents the step number, and γ is a hyperparameter used to control the proportion of original features retained. This represents a feature filter implemented using a 1×1 convolution; This represents the adjacency matrix at the m-th step of the inflow end. Higher-order features;

[0091] For each step obtained Obtain by splicing

[0092] Output of the outflow graph convolutional layer How to obtain:

[0093] Obtain the feature information of the m-th step in the outflow end.

[0094] right Obtain by splicing

[0095] This represents a feature filter implemented using a 1×1 convolution;

[0096] Output of a double-ended graph convolutional layer

[0097] In this embodiment, the prediction layer includes two standard convolutional layers. The first standard convolutional layer aggregates L+1 inputs to obtain the result. The second standard convolutional layer will Transform dimensions to S1 acquires real-time traffic flow data X∈R H×N×C H represents the number of time steps in the input traffic flow data sequence, N represents the number of nodes, and C represents the number of features. Multi-step prediction is performed using the extracted diverse spatiotemporal features. The goal of this implementation is to make the prediction result as close as possible to the true value; here, the L2 loss function is chosen as the objective function for the prediction task.

[0098]

[0099] in, It is the predicted value, while y H+i It is the actual value.

[0100] While the invention has been described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the invention. Therefore, it should be understood that many modifications can be made to the exemplary embodiments, and other arrangements can be designed without departing from the spirit and scope of the invention as defined by the appended claims. It should be understood that different dependent claims and features described herein can be combined in ways different from those described in the original claims. It is also understood that features described in conjunction with individual embodiments can be used in other described embodiments.

Claims

1. A bi-terminal graph convolutional traffic flow prediction method with graph learning, characterized in that, The method includes: S1. Obtain real-time traffic flow data; S2. Input the acquired real-time traffic flow data into the prediction model to obtain the predicted traffic flow; The prediction model includes a feature extraction layer, a prediction layer, L dynamic graph learning layers, L gated temporal convolutional layers, and L bi-terminal graph convolutional layers; The extraction layer is used to extract features from real-time traffic flow data, which serve as initial features. ; A block consists of a dynamic graph learning layer, a gated temporal convolutional layer, and a paired-terminal graph convolutional layer. The blocks are concatenated together, and the output of the gated convolutional layer in each block is summed with the output of the first block. The outputs of each block are simultaneously input into the prediction layer, and the prediction layer outputs the traffic flow for the next time step. No. In each block, , by the Output of each block At the same time as the first The first gated temporal convolutional layer and the first The input to each dynamic graph learning layer, with the initial features being the input in the first block. , No. A gated temporal convolutional layer is used to capture the temporal correlation features of traffic flow changes over time for each road segment, and to obtain the output hidden state of the gated temporal convolutional layer. And hide the output. Input to the A double-ended graph convolutional layer; simultaneously, a dynamic graph learning layer is used to obtain the adjacency matrix by learning the continuous graph and the backup graph. , and input to the first The first dual-ended graph convolutional layer; A pairwise graph convolutional layer is used to determine the output hidden state. and adjacency matrix By capturing the spatial correlation characteristics of traffic flow changes over time between different road segments, we can obtain... . A two-ended graph convolutional layer consists of an inflow graph convolutional layer and an outflow graph convolutional layer; The output of the inflow graph convolutional layer How to obtain: Obtain the adjacency matrix degree matrix , The elements on the diagonal are , Representing the adjacency matrix Middle Line 1 Column characteristics; Obtain matrix : For the matrix Update , This represents the sum of the values ​​in each row; get : subscript , Indicates the number of steps. It is a hyperparameter used to control the proportion of original features retained. express Feature filter implemented using convolution; This represents the adjacency matrix at the m-th step of the inflow end. Higher-order features; For each step obtained Obtain by splicing ; ; Output of the outflow graph convolutional layer How to obtain: Obtain the feature information of the m-th step in the outflow end. , ; right Obtain by splicing ; , express Feature filter implemented using convolution; Output of a double-ended graph convolutional layer .

2. The bi-terminal graph convolutional traffic flow prediction method with graph learning according to claim 1, characterized in that, The dynamic graph learning layer is used to obtain the adjacency matrix by learning from continuous graphs and alternate graphs. include: in, and Indicates embedding in the In each block, there are two learnable nodes, where d is the size of the embedded learnable nodes and N represents the number of nodes. and These correspond to the linear layer parameters. yes Similarity matrix Indicates the activation function; It is a hyperparameter used to control the saturation rate of the activation function.

3. The bi-terminal graph convolutional traffic flow prediction method with graph learning according to claim 2, characterized in that, Gated temporal convolutional layers include dilated convolutional layers and dilated inception layers: No. Output of each block Simultaneously input to the dilated convolutional layer and the dilated inception layer The output of the dilated convolutional layer is The output of the void inception layer is Output hidden state for: It is the element-wise multiplication operator. express function.

4. The bi-terminal graph convolutional traffic flow prediction method with graph learning according to claim 3, characterized in that, The output of the dilated convolutional layer is: in, Indicates the first kernel size of the layer This is a dilated convolution operation.

5. The bi-terminal graph convolutional traffic flow prediction method with graph learning according to claim 4, characterized in that, The perforated inception layer consists of a 1×1 convolutional layer, four different perforated convolutional layers, four ConvGRU layers, and one splicing layer; No. Output of each block Input to a 1×1 convolutional layer, and obtain ,Will The outputs are passed to four different dilated convolutional layers to obtain the corresponding outputs. ,Will The inputs are fed into four ConvGRU layers, and the four ConvGRU layers output from each layer respectively. splicing layer obtained , Indicates splicing.

6. The bi-terminal graph convolutional traffic flow prediction method with graph learning according to claim 5, characterized in that, Will The inputs are fed into four ConvGRU layers, and the four ConvGRU layers output from each layer respectively. The methods include: , It is the element-wise multiplication operator. express function, , , , , , Represents the learnable parameters. This represents the convolution operation. .

7. The bi-terminal graph convolutional traffic flow prediction method with graph learning according to claim 1, characterized in that, The prediction layer includes two standard convolutional layers. The first standard convolutional layer aggregates L+1 inputs to obtain the result. The second standard convolutional layer will Transform dimensions to S1 real-time traffic flow data H represents the number of time steps in the input traffic flow data sequence, N represents the number of nodes, and C represents the number of features.

8. A computer-readable storage device storing a computer program, characterized in that, When the computer program is executed, it implements the bi-terminal graph convolutional traffic flow prediction method with graph learning as described in any one of claims 1 to 7.

9. A traffic flow prediction device with graph learning-based bi-terminal graph convolution, comprising a storage device, a processor, and a computer program stored in the storage device and executable on the processor, characterized in that, The processor executes the computer program to implement the bi-terminal graph convolutional traffic flow prediction method with graph learning as described in any one of claims 1 to 7.

Images