Traffic flow prediction method based on regional graph structure

Abstract

The invention provides a traffic flow prediction method based on a regional graph structure, and the method comprises the steps: constructing a regional graph from an original graph structure through a region segmentation technology, integrating key region information through a graph convolutional neural network, and carrying out the traffic flow prediction through the introduction of a memory network, recognition and utilization of a potential mode between regions.

Description

Technical Field

[0001] The present invention belongs to the technical fields of traffic flow prediction, graph neural network, etc., and specifically relates to a traffic flow prediction method based on a regional graph structure. Background Art

[0002] Traffic flow forecasting is a complex task that relies on in-depth analysis of historical traffic data and takes into account a variety of external factors, such as weather changes, holiday arrangements, and spatial road network structure, to predict traffic flow in a specific period of time in the future. Compared with other time series forecasting problems, traffic flow forecasting faces greater challenges, not only because it involves processing large-scale high-dimensional data, but also because it must deal with a variety of dynamic and emergency situations including traffic accidents and peak hours.

[0003] In recent years, Graph Neural Network (GNN) has made significant progress in the field of deep learning and has become an ideal tool for solving traffic prediction problems. The core advantage of GNN is its ability to capture complex spatial dependencies, which are usually represented by non-Euclidean graph structures. For example, the road network itself is a graph structure with road intersections as nodes and road connections as edges. Using graphs as input, GNN-based models have shown performance that surpasses traditional methods in tasks such as traffic flow and speed prediction.

[0004] However, the performance of GNN is highly dependent on the accuracy of the graph structure. In the real world, the graph structure may not be available or there is a lot of noise, which requires graph structure learning technology to overcome. Graph structure learning learns and generates graph structures from real data and mines the potential graph topology, thereby effectively improving the performance of the model. Despite this, the application of graph structure learning in spatiotemporal data mining, especially in the field of traffic prediction, is still relatively rare.

[0005] In fact, the traffic network has obvious hierarchical characteristics. It includes not only the basic road network micro-layer, but also the macro-level network, which represents popular traffic areas in the form of nodes, such as the city center, central business district or other key areas (see Appendix Figure 1 ). The characteristics and information of these areas or communities play a vital role in transportation planning and macro-characteristic analysis.

[0006] Urban traffic zones have different functions, and traffic flow tends to shift between these zones. For example, during the morning rush hour, traffic flow from residential areas tends to shift toward commercial areas. However, this shift pattern may be different at other times of the day, such as lunch time or after get off work. Summary of the invention

[0007] In view of the defects and shortcomings of the existing technology, the present invention considers that the transmission mode of traffic flow between regions is crucial to understanding traffic flow. Regions with similar functions or similar population densities often show similar spatiotemporal patterns, while regions with different functions show different patterns. These spatiotemporal patterns that are prevalent in cities provide valuable auxiliary information for learning the spatiotemporal representation of the original graph of global dependencies.

[0008] A traffic flow prediction method based on a regional graph structure is proposed. Based on the regional graph structure, it aims to improve the accuracy of spatiotemporal prediction tasks. Compared with traditional time series prediction techniques, this method specifically targets the complexity of traffic flow prediction, including processing high-dimensional data and dynamically changing factors (such as traffic accidents and peak hours). The regional graph is constructed from the original graph structure through regional segmentation technology, and the key regional information is integrated using a graph convolutional neural network to reveal the complex interactive relationship between regions. In addition, through the introduction of memory networks, the potential patterns between regions can be identified and utilized, thereby deepening the understanding of the spatial structure of the transportation network and significantly improving the accuracy of prediction.

[0009] The technical solution specifically adopted by the present invention to solve the technical problem is:

[0010] A traffic flow prediction method based on regional graph structure: construct a regional graph from the original graph structure through regional segmentation technology, integrate key regional information using graph convolutional neural network, and identify and utilize potential patterns between regions for traffic flow prediction by introducing memory network.

[0011] Furthermore, firstly, through the regional graph structure generation module, starting from the original adjacency matrix, a regional graph reflecting the regional characteristics of the transportation network is constructed;

[0012] Then, the original graph and the region graph are subjected to the same space-time coding operation by using the space-time coding module to embed the space-time information of the nodes of the original graph and the nodes of the newly generated region graph, and then the nodes are fused by the dynamic interaction module;

[0013] Then, the memory pattern learning module is used to memorize the pattern library based on the meta-knowledge. Through the key-value memory structure, the association between the input data and the meta-knowledge items in the memory library is learned to extract the pattern reflecting the periodic flow changes to enhance the representation of the hidden state.

[0014] Finally, the learned adaptive road network structure is fed back to the decoder of the traffic prediction module; the decoder completes the prediction of traffic flow by decoding sequences of uniform size.

[0015] Furthermore, the regional graph generation module first calculates the modularity of each node in the adjacency matrix set of the road network structure:

[0016]

[0017] Among them, ∑in represents the weight inside community c, ∑tot represents the weight of the edge connected to the points inside community c, including the edges inside the community and the edges outside the community; m is the sum of the weights of all links in the network;

[0018] By maximizing the modularity, the original image is segmented into regions to obtain a new traffic area map G. region , and the values ​​of all features in all regions T, with timestamps denoted as X region =(X1, X2, ..., X k ) T ∈R K×T×c ; The traffic flow of the area is represented by the average value of the traffic flow of the road sections.

[0019] Furthermore, the spatiotemporal encoding module first reshapes the features of all time steps into a vector of length T×n×c, where T is the number of time steps, n is the number of nodes, and c is the number of input features; then a graph convolutional network is used to utilize the adjacency matrix A∈R n×n Determine how nodes and their neighbors are clustered to capture the spatial correlation between different locations and obtain the embedding vector H∈R T×n×d ;The output vector H captures the temporal dependency through a recurrent neural network;

[0020] We then use stackable graph convolutional recurrent units to simultaneously capture spatiotemporal dependencies.

[0021] Furthermore, the same spatiotemporal encoding operation is performed on both the original image and the region image to obtain the hidden representations H of the two images after embedding spatiotemporal information. t and

[0022] The fusion through the dynamic interaction module is specifically: The information is transferred to the original graph H through the dynamic transfer matrix t Then, the two are spliced ​​together to obtain the fusion feature H′ with regional information t :

[0023]

[0024] Furthermore, the memory pattern learning module performs the fusion of the features H′ t First, initialize a memory pattern library consisting of m memory items Φ of length e, denoted as θ m ,Will That is, H′ t The i-th node vector in is projected onto the query vector in the memory

[0025]

[0026] Among them, W1 and b1 are learnable parameters; Calculate the matching weight a for each memory item in the memory pattern library, that is, the attention score, and combine each memory item based on the attention score to obtain the hidden traffic pattern representative of node i

[0027]

[0028] Use M t ∈R n×e Hidden representation H′ of enhanced coding t , expressed as:

[0029]

[0030] Enter the decoder for traffic prediction.

[0031] Furthermore, the memory items in the memory bank are updated by using two constraints:

[0032]

[0033] where Φ[p] and Φ[n] represent the first two similar memory items.

[0034] Furthermore, in the traffic prediction module, the decoder uses the same graph convolutional recurrent network as the spatiotemporal encoding module, firstly Provide adaptive graph structure A apt , initialize a node embedding representation dictionary E for all nodes A ∈R n×(e+d) ; The similarity between new nodes is expressed by E A E A T express:

[0035] A apt =Softmax(ReLU(E A E A T ))

[0036] During the training process, E A By automatically updating, we learn the hidden details between different traffic sequences and calculate the adaptive adjacency matrix A for graph convolution. apt ;

[0037] and A aptThe sequence prediction value converted to the future time step is obtained through the graph convolutional recurrent network During the training process, the goal is to minimize the error between the actual traffic flow on the road and the predicted value; Represent the actual traffic flow and predicted traffic flow respectively; the total loss function is:

[0038]

[0039] And, a traffic flow prediction system based on a regional graph structure, comprising:

[0040] The regional graph structure generation module is used to construct a regional graph reflecting the regional characteristics of the transportation network based on the original adjacency matrix;

[0041] A spatiotemporal coding module, used for performing the same spatiotemporal coding operation on the original graph and the region graph to embed spatiotemporal information into the nodes of the original graph and the nodes of the newly generated region graph, and then fusing them through a dynamic interaction module;

[0042] A memory pattern learning module is used for learning the association between input data and meta-knowledge items in the memory library based on the fusion result of the dynamic interaction module through a key-value memory structure, so as to extract patterns reflecting periodic traffic changes, so as to enhance the representation of hidden states;

[0043] The traffic prediction module is used to receive the adaptive road network structure learned by the memory pattern learning module and feed it back to the decoder; the decoder completes the prediction of traffic flow by decoding sequences of uniform size.

[0044] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps of the traffic flow prediction method based on the regional graph structure as described above are implemented.

[0045] A non-transitory computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the traffic flow prediction method based on the regional graph structure as described above.

[0046] Compared with the prior art, the present invention and its preferred solution not only deeply consider the basic road network structure, but also further integrate the significant hierarchical structural characteristics of the traffic network. Through the graph convolutional recurrent neural network, the present invention can accurately capture the spatiotemporal correlation between regional and basic level traffic flow nodes, and realize the comprehensive grasp of the multi-level and multi-scale spatiotemporal information of traffic flow.

[0047] The present invention fully considers the spatial characteristics and the problem that the number of nodes in the same area may be inconsistent. A community discovery method that fully utilizes the spatial characteristics of the graph structure is adopted to segment the original graph into regions through the strategy of maximizing modularity, forming a traffic area with clear functional division, thereby further highlighting the spatial characteristics of the traffic network.

[0048] In addition, the present invention introduces a memory network-based method to perceive traffic pattern information. By capturing representative spatiotemporal patterns in the city, it provides auxiliary information for spatiotemporal representation learning and significantly improves the accuracy and reliability of predictions.

[0049] The core design points of the present invention and its preferred solution that are different from the prior art include at least:

[0050] 1. An advanced method is proposed to capture the information in the regional graph structure. The method segments the original graph into regions by maximizing the modularity strategy, thus forming traffic regions with clear functional divisions. Furthermore, a graph convolutional recurrent neural network is used to perceive the spatiotemporal correlation between the original nodes and the newly formed regional nodes. Through a carefully designed dynamic interaction module, these two features are effectively fused to achieve a comprehensive capture of traffic flow changes.

[0051] 2. In order to better understand and predict the dynamic changes of traffic flow, a memory network-based method is proposed to perceive traffic pattern information. The method first maps the encoded hidden features to the query vector in the memory bank. Then, through the key-value pair query mechanism, the memory items in the memory bank are matched with the current traffic status to form a representation that can represent the hidden traffic pattern. This representation is combined with the feature representation to predict the traffic flow in the future time step, thereby improving the accuracy and reliability of the prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0053] Figure 1 It is a schematic diagram of the urban area structure.

[0054] Figure 2 It is a flow chart of the overall model of an embodiment of the present invention.

[0055] Figure 3 It is a schematic diagram of generating a regional map structure according to an embodiment of the present invention.

[0056] Figure 4 It is a diagram of the space-time coding structure of an embodiment of the present invention.

[0057] Figure 5 It is a schematic diagram of memory mode learning according to an embodiment of the present invention. DETAILED DESCRIPTION

[0058] In order to make the features and advantages of the present invention more clearly understood, the following embodiments are specifically described in detail as follows:

[0059] It should be noted that the following detailed descriptions are illustrative and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used in this specification have the same meanings as those commonly understood by those skilled in the art to which the present application belongs.

[0060] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0061] 1. Model Overview:

[0062] First, the model of the embodiment of the present invention constructs a regional graph reflecting the characteristics of the traffic network area from the original adjacency matrix through the regional graph structure generation module. This step is the basis of the model and provides a structured graph representation for subsequent analysis. Subsequently, the model uses the spatiotemporal encoding module to deeply embed spatiotemporal information into the nodes of the original graph and the nodes of the newly generated regional graph. This process enables the model to capture and understand the complex changes in traffic flow at different time and spatial locations, providing rich contextual information for prediction. In order to identify periodic patterns in traffic flow, the model introduces a meta-knowledge memory pattern library. Through the key-value memory structure, the model learns the association between the input data and the meta-knowledge items in the memory library, thereby extracting patterns that reflect periodic traffic changes. These patterns are used to enhance the model's representation of hidden states and improve the accuracy of predictions. Finally, the model feeds back the learned adaptive road network structure to the decoder. The decoder uses this structure to complete the prediction of traffic flow by decoding sequences of uniform size. This adaptive mechanism enables the model to flexibly respond to different traffic conditions and improves the adaptability and accuracy of predictions. For details of the overall process of the model, see Figure 2 , which shows how the model can achieve accurate prediction of traffic flow through the collaboration of these modules.

[0063] 2. Area map generation module

[0064] The traffic distribution on roads with the same function or level will have a high degree of similarity. On this basis, the traffic road network is divided into regions to form a traffic regional network, such as Figure 3As shown in the figure, the regional graph generation module first calculates the modularity of each node in the adjacency matrix set of the road network structure:

[0065]

[0066] Among them, ∑in represents the weight inside community c, ∑tot represents the weight of the edge connected to the points inside community c, including the edges inside the community and the edges outside the community. m is the sum of the weights of all links in the network.

[0067] The original image is segmented by maximizing the modularity. From this construction, a new traffic area graph G can be obtained. region , and the values ​​of all features in all regions T timestamps can be expressed as X region =(X1, X2, ..., X k ) T ∈R K×T×c The traffic volume in the area is represented by the average value of the traffic volume of the road segments.

[0068] 3. Spatiotemporal Coding Module

[0069] First, reshape the features at all time steps into a vector of length T×n×c, where T is the number of time steps, n is the number of nodes, and c is the number of input features, such as flow, speed, etc. Graph Convolutional Networks (GCN) uses the adjacency matrix A∈R n×n To determine how nodes and their neighbors are clustered, thereby capturing the spatial correlation between different locations. The general operation of the graph convolutional network on the original features is:

[0070]

[0071] Where σ is an activation function, which is preferably set to Softmax in this embodiment. k It is a learnable parameter, and the embedding vector H∈R is obtained through graph convolution operation T×n×d The convolution operation captures spatial dependencies by applying a filter to the input.

[0072] The output H captures the time dependency through a recurrent neural network, and the formula at time t can be expressed as:

[0073]

[0074] Where u represents the update gate and r represents the reset gate. These two gate units can help selectively retain historical data information. C represents the candidate state in the cell. θ u ,θ r ,θC Represents the gate parameter. The overall structure of the decoder is basically the same as that of the encoder. Figure 4 .

[0075] Next, we use a stackable graph convolutional recurrent unit to capture the spatiotemporal dependencies simultaneously. We perform the same spatiotemporal encoding operation on both the original graph and the region graph, and obtain the hidden representations H of the two graphs after embedding the spatiotemporal information. t and

[0076] 4. Memory mode learning module

[0077] In the real world, traffic flow in traffic networks exhibits periodic characteristics, and cities have many representative spatiotemporal patterns (the peak hours of commuting traffic in living areas are usually similar, while those in central business districts are different), which can provide auxiliary information for learning the spatiotemporal representation of the original graph as a global dependency.

[0078] First, through the dynamic interaction module fusion, the regional map The information is transferred to the original graph H through the dynamic transfer matrix t Then, the two are spliced ​​together to obtain the fusion feature H′ with regional information t :

[0079]

[0080] For the fused feature H′ t The present invention further explores the traffic pattern information it contains and initializes a memory pattern library (see Figure 5 ), the memory bank consists of m memory items Φ of length e, denoted by θ m ,Will That is, H′ t The i-th node vector in is projected onto the query vector in the memory

[0081]

[0082] Among them, W1 and b1 are learnable parameters. Calculate the matching weight a for each memory item in the memory pattern library, that is, the attention score, and combine each memory item based on the attention score to obtain the hidden traffic pattern representative of node i

[0083]

[0084] Use M t ∈R n×e To enhance the encoded hidden representation H′ t , expressed as:

[0085]

[0086] Enter the decoder for traffic prediction. In addition, since the sensor signals in different areas are still entangled, in order to enhance its ability to distinguish various situations on different roads over a longer period of time, the present invention updates the memory items in the memory network library by using two constraints:

[0087]

[0088] Where Φ[p] and Φ[n] represent the first two similar memory items. Through these constraints, we can strive to ensure that the memory network library can store a wider variety of patterns with distinctive features.

[0089] 5. Traffic prediction module

[0090] The decoder uses the same graph convolutional recurrent network as the encoder. First, Provide adaptive graph structure A apt The predefined adjacency matrix A is not used because it will be affected by factors such as geographical distance, cannot contain complete spatial dependencies, and is not directly related to specific prediction tasks. The data-driven adaptive graph can capture hidden information. First, a node embedding representation dictionary E is initialized for all nodes. A ∈R n×(e+d) The similarity between new nodes can be expressed as E A E A T express:

[0091] A apt =Softmax(ReLU(E A E A T )), (11)

[0092] During the training process, E A It can be automatically updated to learn the hidden details between different traffic sequences and calculate the adaptive adjacency matrix A for graph convolution apt .

[0093] and A apt The sequence prediction value converted to the future time step is obtained through the graph convolutional recurrent network During the training process, the goal is to minimize the error between the actual traffic flow on the road and the predicted value. Represent the actual traffic flow and predicted traffic flow respectively. The total loss function is:

[0094]

[0095] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0096] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0097] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0098] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0099] It should be noted that, unless otherwise defined, the technical terms or scientific terms used in the present invention should be understood by people with ordinary skills in the field to which the present invention belongs. The "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprise" and similar words mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0100] The above is only a preferred embodiment of the present invention, and does not limit the present invention in other forms. Any technician familiar with the profession may use the above disclosed technical content to change or modify it into an equivalent embodiment with equivalent changes. However, any simple modification, equivalent change and modification made to the above embodiment according to the technical essence of the present invention without departing from the technical solution of the present invention still belongs to the protection scope of the technical solution of the present invention.

[0101] The present invention is not limited to the above-mentioned optimal implementation mode. Anyone can derive other various forms of traffic flow prediction methods based on regional graph structures under the inspiration of the present invention. All equal changes and modifications made within the scope of the invention application of the present invention should fall within the scope of the present invention.

Claims

1. A traffic flow prediction method based on a regional graph structure, characterized in that: The regional map is constructed from the original graph structure through region segmentation technology, and the key regional information is integrated using graph convolutional neural network. By introducing memory network, the potential patterns between regions are identified and utilized for traffic flow prediction.

2. The traffic flow prediction method based on the regional graph structure according to claim 1 is characterized by: Firstly, the regional graph structure generation module is used to construct a regional graph reflecting the regional characteristics of the transportation network based on the original adjacency matrix. Then, the original graph and the region graph are subjected to the same space-time coding operation by using the space-time coding module to embed the space-time information of the nodes of the original graph and the nodes of the newly generated region graph, and then the nodes are fused by the dynamic interaction module; Then, the memory pattern learning module is used to memorize the pattern library based on meta-knowledge; Through the key-value memory structure, the association between input data and meta-knowledge items in the memory library is learned to extract patterns that reflect periodic traffic changes to enhance the representation of hidden states; Finally, the learned adaptive road network structure is fed back to the decoder of the traffic prediction module; the decoder completes the prediction of traffic flow by decoding sequences of uniform size.

3. The traffic flow prediction method based on regional graph structure according to claim 2 is characterized by: The regional graph generation module first calculates the modularity of each node in the adjacency matrix set of the road network structure: Among them, ∑in represents the weight inside community c, ∑tot represents the weight of the edge connected to the points inside community c, including the edges inside the community and the edges outside the community; m is the sum of the weights of all links in the network; By maximizing the modularity, the original image is segmented into regions to obtain a new traffic area map G. region , and the values ​​of all features in all regions T, with timestamps denoted as X region =(X1,X2,…,X k ) T ∈R K×T×c ; The traffic flow of the area is represented by the average value of the traffic flow of the road sections.

4. The traffic flow prediction method based on regional graph structure according to claim 3 is characterized by: The spatiotemporal encoding module first reshapes the features of all time steps into a vector of length T×n×c, where T is the number of time steps, n is the number of nodes, and c is the number of input features; Then, a graph convolutional network is used to use the adjacency matrix A∈R n×n Determine how nodes and their neighbors are clustered to capture the spatial correlation between different locations and obtain the embedding vector H∈R T ×n×d ;The output vector H captures the temporal dependency through a recurrent neural network; We then use stackable graph convolutional recurrent units to simultaneously capture spatiotemporal dependencies.

5. The traffic flow prediction method based on regional graph structure according to claim 4 is characterized by: The same spatiotemporal encoding operation is performed on both the original image and the region image to obtain the hidden representations H of the two images after embedding spatiotemporal information. t and The fusion through the dynamic interaction module is specifically: The information is transferred to the original graph H through the dynamic transfer matrix t Then, the two are spliced ​​together to obtain the fusion feature H' with regional information t :

6. The traffic flow prediction method based on regional graph structure according to claim 5 is characterized by: The memory pattern learning module performs the fusion of the features H' t First, initialize a memory pattern library consisting of m memory items Φ of length e, denoted as θ m ,Will That is H' t The i-th node vector in is projected onto the query vector in the memory Where W1 and b1 are learnable parameters; Calculate the matching weight a for each memory item in the memory pattern library, that is, the attention score, and combine each memory item based on the attention score to obtain the hidden traffic pattern representative of node i Use M t ∈R n×e Hidden representation H' of enhanced coding t , expressed as: Enter the decoder for traffic prediction.

7. The traffic flow prediction method based on regional graph structure according to claim 6 is characterized by: The memory items in the memory bank are updated by using two constraints: where Φ[p] and Φ[n] represent the first two similar memory items.

8. The traffic flow prediction method based on regional graph structure according to claim 6 is characterized by: In the traffic prediction module, the decoder uses the same graph convolutional recurrent network as the spatiotemporal encoding module. First, Providing adaptive graph structure A apt , initialize a node embedding representation dictionary E for all nodes A ∈R n×(e+d) ; The similarity between new nodes is expressed by E A E A T express: TO apt =Softmax(ReLU(E A AND A T )) During the training process, E A By automatically updating, we learn the hidden details between different traffic sequences and calculate the adaptive adjacency matrix A for graph convolution. apt ; and A apt The sequence prediction value converted to the future time step is obtained through the graph convolutional recurrent network During the training process, the goal is to minimize the error between the actual traffic flow on the road and the predicted value; Represent the actual traffic flow and predicted traffic flow respectively; the total loss function is:

9. A traffic flow prediction system based on a regional graph structure, characterized in that: include: The regional graph structure generation module is used to construct a regional graph reflecting the regional characteristics of the transportation network based on the original adjacency matrix; A spatiotemporal coding module, used for performing the same spatiotemporal coding operation on the original graph and the region graph to embed spatiotemporal information into the nodes of the original graph and the nodes of the newly generated region graph, and then fusing them through a dynamic interaction module; A memory pattern learning module, used for memorizing the fusion results of the dynamic interaction module based on the meta-knowledge memory pattern library; Through the key-value memory structure, the association between input data and meta-knowledge items in the memory library is learned to extract patterns that reflect periodic traffic changes to enhance the representation of hidden states; The traffic prediction module is used to receive the adaptive road network structure learned by the memory pattern learning module and feed it back to the decoder; the decoder completes the prediction of traffic flow by decoding sequences of uniform size.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps of the traffic flow prediction method based on the regional graph structure as described in any one of claims 1-8 are implemented.

Images