A traffic flow prediction method based on bidirectional heterogeneous dynamic graph convolutional network

Abstract

The present invention relates to a traffic flow prediction method based on a bidirectional heterogeneous dynamic graph convolutional network, which belongs to the field of transportation, including: constructing static heterogeneity factors and dynamic heterogeneity factors, and constructing a bidirectional heterogeneity model according to the static heterogeneity factors and the dynamic heterogeneity factors; a dynamic graph construction module uses the spatiotemporal embedding generated by the spatiotemporal embedding generation module and the connectivity filtering generated by the bidirectional heterogeneity model to construct a dynamic graph, and generates a Laplacian matrix for graph convolution according to the dynamic graph; graph convolution is performed through the Laplacian matrix, the characteristics of nodes in time and space are captured according to the gated cyclic graph convolution module, and multi-step prediction is performed using residual connections. The present invention constructs a bidirectional heterogeneity model, which not only considers the interdependence in space, but also the dynamic changes in time, and realizes the comprehensive capture and accurate prediction of traffic flow changes.

Description

Technical Field

[0001] The present invention relates to the field of transportation, and in particular to a traffic flow prediction method based on a bidirectional heterogeneous dynamic graph convolutional network. Background Art

[0002] Traffic conditions are closely related to our daily lives. Accurately predicting traffic conditions at future time nodes can bring convenience to our lives. Accurate traffic flow prediction information provides decision support for users in more and more scenarios, such as logistics vehicle scheduling, online car-hailing supply and demand scheduling, road fault troubleshooting, etc. Therefore, traffic flow prediction has gradually become an important task.

[0003] Traffic flow data has a large number of characteristics: First, time correlation, the current time of traffic flow is closely related to the current time's relative position within a day. Common examples are morning rush hour and evening rush hour. Second, spatial correlation, the traffic flows of multiple nodes in the same area are closely correlated. A common example is area-scale traffic jams. Many studies focus on modeling these two characteristics.

[0004] In recent years, deep learning technology has made great progress in traffic flow prediction. Recent research has focused on spatiotemporal graph modeling. Spatiotemporal models can model traffic flow data in spatial and temporal dimensions, extract spatiotemporal characteristics of traffic flow, and achieve good prediction results. There are currently a variety of traffic flow prediction models, such as Graph Multi-Attention Network (GMAN), Spatio-Temporal Graph Neural Controlled Differential Equation (STG-NCDE), Time Graph Neural Network (MTGNN), Decomposition Dynamic Graph Convolutional Recurrent Network (DDGCRN). These models use different methods to model the spatiotemporal relationship of traffic flow, integrate it into the spatiotemporal graph, and use graph convolutional neural networks for forward propagation.

[0005] However, most of the above existing technologies focus on static spatial dependencies and use distance as the weight of the adjacency matrix. However, the spatial dependencies in the actual road network change dynamically over time. These models use predefined adjacency matrices to express fixed spatial dependencies, ignoring the dynamic changes of spatial dependencies, and therefore it is difficult to capture sudden changes caused by events such as traffic accidents. In addition, the existing technology models overemphasize the periodic characteristics of traffic flow when designing, such as daily traffic peaks and troughs, and traffic changes on different days of the week; although this design method can better capture the periodic laws of traffic flow, due to the model's excessive reliance on periodicity, it lacks adaptability and flexibility when facing non-periodic changes, and it is difficult to quickly and accurately adjust the prediction to reflect the actual traffic conditions. Summary of the invention

[0006] The purpose of the present invention is to overcome the shortcomings of the prior art and provide a traffic flow prediction method based on a bidirectional heterogeneous dynamic graph convolutional network to solve the shortcomings of the prior art.

[0007] The purpose of the present invention is achieved through the following technical solution: a traffic flow prediction method based on a bidirectional heterogeneous dynamic graph convolutional network, the prediction method comprising:

[0008] S1. Construct static heterogeneity factors and dynamic heterogeneity factors, and construct a two-way heterogeneity model based on the static heterogeneity factors and dynamic heterogeneity factors;

[0009] S2, the dynamic graph construction module uses the spatiotemporal embedding generated by the spatiotemporal embedding generation module and the connectivity filtering generated by the bidirectional heterogeneity model to construct a dynamic graph, and generates a Laplacian matrix for graph convolution based on the dynamic graph;

[0010] S3. Perform graph convolution through the Laplacian matrix, capture the spatiotemporal features of nodes according to the gated recurrent graph convolution module, and use residual connections for multi-step prediction.

[0011] The static heterogeneity factors constructed in S1 include:

[0012] According to the collected traffic flow at time t, according to the formula Calculate the static heterogeneity factor, where represents the static heterogeneity factor of nodes i and j at the current time t, and are the traffic flows of node i and node j at the current time t respectively.

[0013] The dynamic heterogeneity factors constructed in S1 include:

[0014] According to the collected traffic flow at time t and time t-1, according to the formula Calculate the flow difference between adjacent time steps to obtain the traffic flow change at one time step interval, where: is the change in traffic flow at each node at time t, which is determined by the traffic flow X of each node at time t. (t) and the traffic flow X at time t-1 (t-1) Subtracting them, N is the total number of traffic nodes;

[0015] According to the traffic flow changes between nodes i and j, according to the formula The dynamic heterogeneity factor is calculated, where represents the dynamic heterogeneity factor of nodes i and j at the current time t, and They respectively represent the changes in the traffic flow of node i and node j at the current time t compared with the previous time t-1.

[0016] The method of constructing a bidirectional heterogeneity model according to the static heterogeneity factor and the dynamic heterogeneity factor includes:

[0017] The bidirectional heterogeneity factor is calculated based on the calculated static heterogeneity factor and dynamic heterogeneity factor. The probability that it satisfies the discrete distribution is ,when , indicating that the edges of the space-time graph tend to remember the connectivity status of the previous time node. , indicating that the edges of the space-time graph tend to forget the connectivity status of the previous time node and delete the connection. When , it means that the edges of the space-time graph tend to forget the connectivity status of the previous time node and establish new connections;

[0018] By formula Design a memory mechanism for spatiotemporal graph edges, where represents the connectivity between node i and node j at time t, Represents the indicator function, N×N Composition of the connectivity filter C at the current moment t .

[0019] The steps of S2 include the following:

[0020] Extract the relative time of day of traffic flow from the timestamp of traffic flow data , and the relative day of the week , and perform element-by-element multiplication with the spatial embedding E to obtain the spatiotemporal embedding ;

[0021] Use multi-layer perceptron to extract traffic flow features F t , and then the traffic flow feature F t and spatiotemporal embedding Do element-by-element multiplication to get the dynamic graph embedding ;

[0022] Dynamic graph building module uses spatiotemporal embedding of dynamic graphs Transpose Perform matrix multiplication to obtain the dynamic graph at time t , and then with the connectivity filter C t Perform element-by-element multiplication to obtain the filtered dynamic graph, and generate the Laplacian matrix for graph convolution based on the filtered dynamic graph: , D t is the degree matrix of the node at time t, and ReLU is the activation function.

[0023] The steps of S3 specifically include the following:

[0024] According to the generated Laplacian matrix, graph convolution is performed. The convolution process is: , is the weight of the graph convolution, b is the bias of the graph convolution, and Z t Represents the output of the current time step t after the graph convolution network, I N is the N×N identity matrix;

[0025] The gated recurrent graph convolution module contains multiple gated recurrent graph convolution units, which are expressed by the formula Calculate the hidden state of the gated recurrent graph convolution unit, and store the context information between time steps through the hidden state, so that the model can capture the characteristics of nodes in time and space, where σ is the sigmoid activation function, θ is the generated dynamic graph convolution, and x t and H t are the input and output of the model respectively. is a join operation. t is the reset gate, u t is the update gate, W r ,W u ,W c is the learnable weight of the gated recurrent graph convolutional unit, b r ,b u ,b c is a learnable bias;

[0026] The first The output of the gated recurrent graph convolutional unit of the layer After the convergence layer, it becomes The predicted value output of the layer ,in, It means The linear layer of the layer, f here represents the forward propagation;

[0027] At the same time, another set of linear layers is used to obtain the output for back propagation , the input of the current layer and the output of back propagation Take the residual to get the input of the next layer, that is ,in It means The linear layer of the layer, B here represents back propagation;

[0028] Accumulating the prediction output of each layer, the final traffic flow prediction result is , n represents the total number of residual blocks.

[0029] The present invention has the following advantages: a traffic flow prediction method based on a bidirectional heterogeneous dynamic graph convolutional network, a set of new indicators and calculation methods for measuring static traffic heterogeneity factors and dynamic heterogeneity factors are developed, and the flow differences between different traffic nodes at the same time point are captured, thereby quantifying spatial correlation; a bidirectional heterogeneity model is constructed, which not only considers the spatial interdependence, but also the dynamic changes in time, thereby achieving comprehensive capture and accurate prediction of traffic flow changes. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 It is a schematic diagram of the model structure of the present invention;

[0031] Figure 2 Schematic diagram of the structure of the spatiotemporal embedding generation module;

[0032] Figure 3 A schematic diagram of the structure of the dynamic graph building module;

[0033] Figure 4 It is a schematic diagram of the structure of the gated recurrent graph convolution module;

[0034] Figure 5 Schematic diagram of the structure of the gated recurrent graph convolution unit. DETAILED DESCRIPTION

[0035] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the detailed description of the embodiments of the present application provided below in conjunction with the drawings is not intended to limit the scope of protection of the application claimed for protection, but merely represents the selected embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without making creative work belong to the scope of protection of the present application. The present invention is further described below in conjunction with the drawings.

[0036] The present invention relates to a traffic flow prediction method based on a bidirectional heterogeneous dynamic graph convolutional network, so as to solve the problem that the existing traffic flow prediction model is difficult to accurately identify sudden traffic flow pattern mutations, resulting in poor prediction performance when non-periodic events occur; the present invention proposes an innovative bidirectional heterogeneous model, which specifically designs a bidirectional heterogeneous model for detecting and quantifying abnormal changes in traffic flow. The model can accurately identify mutation points in traffic flow by generating abnormal flow filtering, which are often caused by traffic accidents, road construction or other emergencies. In this way, the model can dynamically correct the constructed spatiotemporal graph in real time, ensuring that the graph convolutional network can more accurately consider the impact of these sporadic events on the entire road network during the forward propagation process. This correction mechanism significantly improves the sensitivity of the model to traffic flow mutations, enabling it to more accurately predict traffic conditions when non-periodic events occur, thereby providing more powerful support for traffic management and planning. The introduction of the bidirectional heterogeneous model not only enhances the model's ability to capture non-periodic changes, but also improves the reliability of the prediction results.

[0037] like Figure 1 As shown, specifically including the following:

[0038] Step 1: Construction of static heterogeneity factor: This step aims to quantify the static spatial correlation between nodes in the traffic network at a given time point t. Based on the collected traffic flow at time t, the static heterogeneity factor is calculated according to formula (1).

[0039] (1),

[0040] in represents the static heterogeneity factor of nodes i and j at the current time t, and are the traffic flows of node i and node j at the current time t. This formula uses the monotonically decreasing characteristic of the exponential function to convert the traffic difference between nodes into a weight between 0 and 1 to reflect the static spatial dependency between nodes. The smaller the traffic difference between nodes, the higher the similarity of traffic status and the more likely it is to establish connectivity, and vice versa.

[0041] Step 2: Dynamic heterogeneity factor construction: This step aims to quantify the changes in traffic flow between nodes in the traffic network between given time points t and t-1. Based on the collected traffic flow at time t and time t-1, the dynamic heterogeneity factor is calculated according to the following steps.

[0042] First, the traffic flow difference between adjacent time steps is calculated according to formula (2) to obtain the traffic flow change at an interval of one time step.

[0043] (2),

[0044] in is the change in traffic flow at each node at time t, which is determined by the traffic flow X of each node at time t. (t) and the traffic flow X at time t-1 (t-1) Subtract the traffic flow of each node. Get the traffic flow change of each node , N is the total number of traffic nodes.

[0045] Then, according to the traffic flow changes at nodes i and j, the dynamic heterogeneity factor is calculated according to formula (3).

[0046] (3),

[0047] in is the dynamic heterogeneity factor of nodes i and j at the current time t, and They represent the changes in the traffic flow of node i and node j at the current time t compared with the previous time t-1. By calculating the dynamic heterogeneity factor ,The model not only focuses on the current traffic differences between nodes, but also the ,changing trends of these differences over time.,This allows the model to adapt to the fluctuation of traffic ,flow, and provides a dynamic perspective for prediction.

[0048] Step 3: Construction of bidirectional heterogeneity model: Using the static heterogeneity factor and dynamic heterogeneity factor calculated previously, the bidirectional heterogeneity factor can be calculated The probability of . It satisfies the discrete distribution:

[0049] (4),

[0050] when , indicating that the edges of the space-time graph tend to remember the connectivity status of the previous time node. , indicating that the edges of the space-time graph tend to forget the connectivity status of the previous time node and delete the connection. , it means that the edges of the space-time graph tend to forget the connectivity status of the previous time node and establish new connections.

[0051] The indicator function is used in formula (5) to design the memory mechanism of the spatiotemporal graph edges.

[0052] (5),

[0053] in represents the connectivity between node i and node j at time t, Represents the indicator function, N×N Composition of the connectivity filter C at the current moment t This step achieves dynamic adjustment of the connection status between nodes, simulating the complex interactions in the transportation network that change over time and events.

[0054] Step 4: Figure 2 As shown, the spatiotemporal embedding is generated according to the spatiotemporal embedding generation module: the relative time of the traffic flow in a day is extracted from the timestamp of the traffic flow data , and the relative day of the week , Indicates the frequency of sensor sampling data within a day, represents the days of the week, and D represents the dimension of the period embedding. Perform element-by-element multiplication to get spatiotemporal embedding .

[0055] (6),

[0056] This makes it possible to incorporate time factors into the model and enhance the understanding of traffic flow change patterns, especially in the analysis of traffic characteristics in different time periods and dates.

[0057] Step 5: Generate a Laplacian matrix suitable for graph convolution and first use a multi-layer perceptron to extract traffic flow features F t .

[0058] (7),

[0059] Then the traffic flow feature F t and spatiotemporal embedding Do element-by-element multiplication to get the dynamic graph embedding .

[0060] (8),

[0061] in It is embedded in space and time The spatiotemporal embedding of each time step extracted from .

[0062] like Figure 3 As shown, the dynamic graph building module uses dynamic graph embedding Transpose Perform matrix multiplication to obtain the dynamic graph at time t Then add the connectivity filter C obtained in step 3 t Do element-by-element multiplication to get the filtered dynamic graph. is the activation function. The Laplacian matrix for graph convolution is generated according to the filtered dynamic graph, see formula (9).

[0063] (9),

[0064] Where D t is the degree matrix of the node at time t, and ReLU is the activation function, which takes each element x .

[0065] Step 6: Use the Laplacian matrix generated in step 5 to perform graph convolution. The process of graph convolution is shown in formula (10).

[0066] (10),

[0067] in is the weight of the graph convolution, is the bias of the graph convolution, C represents the number of traffic flow feature dimensions, F represents the output dimension of the graph convolution network, and Z t Represents the output of the current time step t after the graph convolution network, I N is the N×N identity matrix.

[0068] Step 7: Figure 4 and Figure 5 As shown, the gated recurrent graph convolution module contains multiple gated recurrent graph convolution units. The present invention uses a gated recurrent graph convolution unit (DGCRU) to capture time dependencies. The hidden state in the DGCRU is used to store contextual information between time steps. The hidden state retains the information of the previous step in the sequence and passes it to the subsequent time steps, so that the model can capture and utilize long-term dependencies in the sequence. The hidden state of DGCRU can be calculated by the following formula:

[0069] (11),

[0070] Where σ is the sigmoid activation function, θ is the generated dynamic graph convolution, and x t and H t are the input and output of the model respectively. is a join operation. t is the reset gate, u t is the update gate, W r ,W u ,W c is the learnable weight of the gated recurrent graph convolutional unit, b r ,b u ,b c It is a learnable bias. With the help of gated recurrent graph convolution units, the spatiotemporal features of nodes can be captured dynamically.

[0071] Step 8: Use residual connections for multi-step prediction. Residual connections are a technique that allows the network to directly learn the residual (i.e., difference) between the input and output during training. The core concept of residual connections is to simplify the network's learning ability by adding the output of backpropagation, making it easier to learn the identity mapping. The output of the gated recurrent graph convolutional unit of the layer After the pooling layer (linear layer), it is transformed into The predicted value output of the layer . Where T is the length of the traffic flow prediction time window.

[0072] (12),

[0073] in It means The linear layer of the layer, f here represents the forward propagation. At the same time, another set of linear layers is used to obtain the output for the reverse propagation. .

[0074] (13),

[0075] in It means The linear layer of the layer, B here represents back propagation.

[0076] The input of the current layer and the output of back propagation Take the residual to get the input of the next layer.

[0077] (14),

[0078] The prediction output of each layer is accumulated to obtain the final traffic flow prediction result, where n represents the total number of residual blocks.

[0079] (15),

[0080] This achieves accurate prediction of the future state of traffic flow. At the same time, the residual link helps to avoid the gradient vanishing problem in deep network training and improves the prediction performance of the model.

[0081] The present invention introduces dynamic heterogeneity factors , as shown in formula (3), can quantify the changes in traffic flow between nodes in the traffic network between given time points t and t-1. This design enables the model to respond to sudden changes in traffic flow in real time. By calculating the bidirectional heterogeneity of traffic flow at paired time steps, the present invention can generate a corresponding dynamic adjacency matrix for each time node, thereby accurately capturing the dynamic changes in traffic flow.

[0082] The present invention provides a dynamic adjustment mechanism through a bidirectional heterogeneity model, as shown in formulas (4) and (5), which can update the connection status between nodes in real time according to the flow changes in the current and previous time steps. Directly use a single time step to extract the dynamic graph for graph convolution. This model uses the abnormal flow filter C generated by the bidirectional heterogeneity model t The extracted dynamic graph is filtered, and then the filtered dynamic graph is convolved, see formula (9). This mechanism enables the model to quickly identify and adapt to sudden changes in traffic flow when faced with emergencies. For example, when a significant flow change is detected, the model forgets the old edge weights and re-establishes the weights based on the current traffic conditions. This dynamic weight adjustment strategy improves the model's sensitivity to sudden changes in traffic flow, enabling the model to provide more accurate predictions when non-periodic events occur.

[0083] The above is only a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the form disclosed herein, and should not be regarded as excluding other embodiments, but can be used for various other combinations, modifications and improvements, and can be modified within the scope of the concept described herein through the above teachings or the technology or knowledge of the relevant field. The changes and modifications made by those skilled in the art do not deviate from the spirit and scope of the present invention, and should be within the scope of protection of the claims attached to the present invention.

Claims

1. A traffic flow prediction method based on a bidirectional heterogeneous dynamic graph convolutional network, characterized by: The prediction method comprises: S1. Construct static heterogeneity factors and dynamic heterogeneity factors, and construct a two-way heterogeneity model based on the static heterogeneity factors and dynamic heterogeneity factors; S2, the dynamic graph construction module uses the spatiotemporal embedding generated by the spatiotemporal embedding generation module to construct a dynamic graph, and uses the bidirectional heterogeneity model to construct a connectivity filter, uses the connectivity filter to orient the dynamic graph neural network, and constructs an improved Laplacian matrix; S3, use the Laplacian matrix improved by the bidirectional heterogeneity model for graph convolution, capture the spatiotemporal context of the node according to the gated recurrent graph convolution module, and use residual connections for multi-step prediction; The static heterogeneity factors constructed in S1 include: According to the collected traffic flow at time t, according to the formula The static heterogeneity factor is calculated to characterize the static heterogeneity of the node, where represents the static heterogeneity factor of nodes i and j at the current time t, and are the traffic flows of nodes i and j at the current time t; The dynamic heterogeneity factors constructed in S1 include: According to the collected traffic flow at time t and time t-1, according to the formula ΔX (t) =X (t) -X (t-1) Calculate the flow difference between adjacent time steps to obtain the traffic flow change at one time step interval, where ΔX (t) is the change in traffic flow at each node at time t, which is determined by the traffic flow X of each node at time t. (t) and the traffic flow X at time t-1 (t-1) Subtracting them, N is the total number of traffic nodes; According to the traffic flow changes at nodes i and j, according to the formula The dynamic heterogeneity factor is calculated to characterize the dynamic changes of traffic flow, where: represents the dynamic heterogeneity factor of nodes i and j at the current time t, and They represent the changes of the traffic flow of node i and node j at the current time t compared with the previous time t-1; The method of constructing a bidirectional heterogeneity model according to the static heterogeneity factor and the dynamic heterogeneity factor includes: The bidirectional heterogeneity factor is calculated based on the calculated static heterogeneity factor and dynamic heterogeneity factor. The probability that it satisfies the discrete distribution is ,when , indicating that the edges of the space-time graph tend to remember the connectivity state of the nodes at the previous time. , indicating that the edges of the spatiotemporal graph tend to forget the connectivity status of the previous time node and delete the connection. When , it means that the edges of the space-time graph tend to forget the connectivity status of the previous time node and establish new connections; By formula Design a memory mechanism for spatiotemporal graph edges, where represents the connectivity between node i and node j at time t, Represents the indicator function, N×N Composition of the connectivity filter C at the current moment t , so that the propagation neurons in the graph convolutional neural network can adjust the connectivity according to the traffic conditions.

2. The traffic flow prediction method based on bidirectional heterogeneous dynamic graph convolutional network according to claim 1 is characterized by: The steps of S2 include the following: Extract the relative time of day of traffic flow from the timestamp of traffic flow data , and the relative day of the week , and perform element-by-element multiplication with the spatial embedding E to obtain the spatiotemporal embedding ; Use multi-layer perceptron to extract traffic flow features F t , and then the traffic flow feature F t and spatiotemporal embedding Do element-by-element multiplication to get the dynamic graph embedding ; Dynamic graph building module uses spatiotemporal embedding of dynamic graphs Transpose Perform matrix multiplication to obtain the dynamic graph at time t , and then with the connectivity filter C t Perform element-by-element multiplication to orient the propagation nerves of the graph convolutional neural network, thereby giving the graph convolutional network the ability to perceive abnormalities. The Laplacian matrix for graph convolution is generated based on the filtered dynamic graph: , D t is the degree matrix of the node at time t, and ReLU is the activation function.

3. The traffic flow prediction method based on bidirectional heterogeneous dynamic graph convolutional network according to claim 2 is characterized by: The steps of S3 specifically include the following: Use the improved Laplacian matrix constructed in steps S1 and S2 to perform graph convolution. The convolution process is as follows: , is the weight of the graph convolution, b is the bias of the graph convolution, and Z t Represents the output of the current time step t after the graph convolution network, I N is the N×N identity matrix; The gated recurrent graph convolution module contains multiple gated recurrent graph convolution units, which use directed graph convolution for forward propagation. Calculate the hidden state of the gated recurrent graph convolution unit, and store the context information between time steps through the hidden state, so that the model can capture the characteristics of nodes in time and space, where σ is the sigmoid activation function, θ is the generated dynamic graph convolution, and x t and H t are the input and output of the model respectively. is the join operation, r t is the reset gate, u t is the update gate, W r ,W u ,W c is the learnable weight of the gated recurrent graph convolutional unit, b r ,b u ,b c is a learnable bias; The first The output of the gated recurrent graph convolutional unit of the layer After the convergence layer, it becomes The predicted value output of the layer ,in, It means The linear layer of the layer, f here represents the forward propagation; at the same time, another set of linear layers is used to obtain the output for reverse propagation , the input of the current layer and the output of back propagation Take the residual to get the input of the next layer, that is ,in It means The linear layer of the layer, B here represents back propagation; this enables the normal training of the bidirectional heterogeneous dynamic graph convolutional network to avoid the problem of gradient disappearance during the neural network training process. Finally, the prediction output of each layer is accumulated to obtain the final traffic flow prediction result: , n represents the total number of residual blocks.

Images