Traffic flow prediction method of double-layer multi-scale dynamic graph convolutional network

Abstract

The invention discloses a traffic flow prediction method for a double-layer multi-scale dynamic graph convolutional network, and relates to the technical field of traffic, and the method comprises the following steps: constructing a double-layer structure of a traffic network, which comprises a node layer composed of traffic sensor nodes and a region layer formed by clustering nodes in the node layer, node layer traffic flow data and area layer traffic flow data are extracted; respectively mapping the node layer traffic flow data and the region layer traffic flow data to a potential space to obtain a node layer initial hidden state and a region layer initial hidden state; inputting the initial hidden state of the node layer and the initial hidden state of the region layer into a sequence containing at least one space-time module for processing so as to capture a dynamic space-time dependency relationship; the method has the effect of providing more reliable decision support for traffic management and travel planning.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of transportation, more particularly, it relates to a traffic flow prediction method based on a double-layer multi-scale dynamic graph convolution network. BACKGROUND

[0002] It is of great significance to grasp the spatio-temporal dynamic characteristics of traffic flow in the field of traffic flow prediction for traffic management and intelligent travel. Traffic flow has complex dynamic spatio-temporal characteristics, and the correlation of traffic between road segments and between regions changes over time, and there is also multi-scale time dependence, such as short-term minute-level fluctuations and long-term day and week-level trends. Early methods, such as traditional time series models (such as ARIMA), can only capture linear dependence at a single time scale and are difficult to handle the nonlinear and multi-scale characteristics of traffic flow. Existing models lack effective modeling of the multi-level structure of the traffic network. Some models only focus on the traffic changes of a single sensor node, ignoring the macro traffic patterns at the regional level; some consider regions, but fail to effectively integrate the features of the node layer and the regional layer, resulting in the inability to fully utilize multi-level traffic flow information. For example, some graph convolution-based models only model the space at the node layer, and cannot simultaneously mine the macro spatial correlation between regions. The static graph structure assumption is difficult to adapt to the dynamic changes of traffic flow, such as the diffusion of traffic congestion, which can cause the spatial dependence relationship between nodes to change in real time, and traditional static graph convolution models cannot capture this dynamic change in time. The importance of the features of the node layer and the regional layer in traffic flow changes dynamically over time and space, and traditional fusion methods cannot dynamically adjust the weights of the features of each layer according to the actual situation, resulting in insufficient expressive power of the fused features and affecting the prediction accuracy. SUMMARY

[0003] In view of the deficiencies in the prior art, the purpose of the present application is to provide a traffic flow prediction method based on a double-layer multi-scale dynamic graph convolution network.

[0004] To achieve the above-mentioned purpose, the present application provides the following technical solutions: A traffic flow prediction method based on a double-layer multi-scale dynamic graph convolution network, the method comprising the following steps: Constructing a double-layer structure of a traffic network, including a node layer composed of traffic sensor nodes, and a regional layer formed by clustering nodes in the node layer, and extracting node layer traffic flow data and regional layer traffic flow data; Mapping the node layer traffic flow data and the regional layer traffic flow data to latent space respectively to obtain node layer initial hidden states and regional layer initial hidden states; inputting the node layer initial hidden state and the area layer initial hidden state into a sequence comprising at least one space-time module for processing to capture dynamic space-time dependency; wherein each space-time module comprises a parameter independent processing unit and a parameter shared processing unit for respectively transforming the hidden state of the node layer and the hidden state of the area layer in the spatial dimension, the time dimension and the feature dimension; based on an attention mechanism, dynamically fusing the node layer hidden state and the area layer hidden state processed by the space-time module sequence to obtain a fusion feature; processing the fusion feature to output a prediction result of future traffic flow.

[0005] Preferably, the area layer traffic flow data is obtained by the following method: using a K-means clustering algorithm to cluster traffic sensor nodes; calculating the average value of all node traffic flow features in each cluster to obtain the traffic flow feature of the corresponding area.

[0006] Preferably, based on an attention mechanism, dynamically fusing the node layer hidden state and the area layer hidden state processed by the space-time module sequence to obtain a fusion feature, specifically comprising the following steps: constructing a matrix describing the relationship between node i and area j , the elements of which are defined as formula (1); ; (1) based on an attention mechanism, constructing a dynamic transformation matrix , the calculation process of which is as formula (2) and formula (3); ; (2) (3) wherein, represents the hidden feature of the node, represents the hidden feature of the area, represents the convolution operation of the node layer hidden feature with a 1x1 convolution kernel, represents the convolution operation of the area layer hidden feature with a 1x1 convolution kernel, σ represents a sigmoid activation function, * represents matrix multiplication, and represents element-wise multiplication; the fusion feature is calculated by formula (4) ; ; (4) wherein, represents a connection operation.

[0007] Preferably, the node layer hidden state and the area layer hidden state The processing steps include: Spatial, temporal, and feature dimension transformations are performed using independent parameters at the l-th layer; Use parameters shared across layers to perform spatial, temporal, and feature dimension transformations; The hidden states obtained by independent parameter processing and parameter sharing processing are fused using equations (5) and (6); (5) (6) in, These represent the hidden states of the node layer after spatial, temporal, and feature dimension transformations, respectively, under parameter-independent and parameter-shared modes. These represent the hidden states of the region layer after spatial, temporal, and feature dimension transformations under parameter-independent and parameter-shared modes, respectively. The residuals are obtained by performing residual connection and batch normalization using equations (7) and (8); (7) (8) Where ReLU represents the activation function and BN represents the batch normalization operation.

[0008] Preferably, the method further includes the following steps: Spatial dimension modeling is achieved through spatial dynamic graph convolution, which is constructed based on Tucker decomposition; for the node layer, the spatial dynamic graph convolution operation in parameter sharing mode is Equation (9). (9) in, This represents the convolution operation. This represents the spatial dynamics of the node layer at time t.

[0009] Preferably, the method further includes the following steps: Temporal modeling is achieved by combining a multi-scale temporal convolution module and a temporal dynamic graph convolution module; the multi-scale temporal convolution module processes the input... The processing is as shown in equations (10), (11), and (12). (10) (11) (12); , , , respectively represent convolution operations with kernel size of 1x2, 1x3, 1x5 and 1x1, represents a normalization operation; For the node layer, the time dynamic graph convolution operation in the parameter sharing mode is formula (13); ; (13) wherein, represents a time dynamic graph at time t.

[0010] Further comprising the following steps: The transformation of feature dimension is independent of the graph convolution operation; for the node layer, the feature transformation in the parameter sharing mode is formula (14); = (14) ; (14) wherein, represents a feature transformation convolution operation.

[0011] Preferably, further comprising the following steps: The node layer fusion features output by each layer of the space-time module The jump connection summation is summed up, and the convolution layer processing is as formula (15); ; (15) wherein, is the final prediction result.

[0012] Compared with the prior art, the present application has the following beneficial effects: The application can extract micro node traffic flow data and macro regional traffic flow data at the same time by constructing a double-layer structure of the traffic network, i.e., a node layer containing traffic sensor nodes and a regional layer formed by node clustering. Such a structure enables the model to capture traffic flow characteristics from two different levels, focusing on both local flow changes of individual nodes and overall flow trends at the regional level. The traffic flow data of the node layer and the regional layer are mapped to the latent space to obtain the initial hidden state, which realizes effective abstraction and representation of the original data, making the data more suitable for subsequent model processing and helping to mine potential spatiotemporal feature patterns in the data. The initial hidden state of the node layer and the regional layer is processed by a spatiotemporal module sequence containing parameter independent processing units and parameter shared processing units, which can accurately capture dynamic spatiotemporal dependencies. The parameter independent processing units can perform unique transformations in the spatial, temporal and feature dimensions according to the characteristics of the node layer and the regional layer, such as focusing on the spatial correlation of individual nodes with their adjacent nodes and the temporal changes of the node layer, and focusing on the macro spatiotemporal patterns between regions of the regional layer. The parameter shared processing units strengthen the common relationship between the node layer and the regional layer, enabling the model to grasp the associated features between the two layers at the same time, thereby more comprehensively and accurately depicting the dynamic spatiotemporal characteristics of traffic flow, such as effectively capturing the commonalities and differences in the flow changes of different regions and their respective nodes during the morning rush hour. Dynamic fusion of the node layer and regional layer hidden states processed by the spatiotemporal module based on the attention mechanism enables the model to focus on more important information when fusing features. For example, during the traffic flow peak period, the flow changes of certain nodes or regions are more critical to the overall prediction, and the attention mechanism will give them higher weights, thereby improving the quality of the fused features and laying a foundation for accurate prediction. Processing the fused features and outputting the prediction results integrates the advantages of the double-layer structure, dynamic spatiotemporal modeling and attention fusion, making the model more accurate and robust in the traffic flow prediction task and better able to cope with the complex dynamic changes of traffic flow, providing more reliable decision support for traffic management and travel planning. BRIEF DESCRIPTION OF DRAWINGS

[0013] Fig. 1 An overall structural schematic diagram in the embodiment; Fig. 2 A structural schematic diagram of STBlock in the embodiment. DETAILED DESCRIPTION

[0014] To make the above objectives, features and advantages of the application more obvious and easy to understand, the specific embodiments of the application will be described in detail below with reference to the accompanying drawings.

[0015] In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present application. However, it will be apparent to one skilled in the art that the present application can be practiced without the specific details set forth in this description. In other instances, well-known methods, procedures, components, and circuits have not been described in detail so as not to unnecessarily obscure aspects of the present application.

[0016] Second, the "one embodiment" or "an embodiment" described herein as including a particular implementation as covered within the scope of a claim can also include features, structures or characteristics of a different embodiment that is not covered by the claim. The residing of a feature, structure, or characteristic under one embodiment within the scope of a claim does not preclude instantiating the feature, structure, or characteristic under a different embodiment within the scope of claims.

[0017] Referring to Figs. 1-2 as shown.

[0018] Embodiments further illustrate a traffic flow prediction method of a double-layer multi-scale dynamic graph convolutional network proposed by the present application.

[0019] A traffic flow prediction method of a double-layer multi-scale dynamic graph convolutional network, the method comprising the following steps: constructing a double-layer structure of a traffic network, including a node layer composed of traffic sensor nodes, and a region layer formed by clustering nodes in the node layer, and extracting node layer traffic flow data and region layer traffic flow data; mapping the node layer traffic flow data and the region layer traffic flow data to latent space respectively to obtain node layer initial hidden states and region layer initial hidden states; inputting the node layer initial hidden states and the region layer initial hidden states into a sequence containing at least one space-time module for processing to capture dynamic space-time dependencies; wherein each space-time module includes a parameter independent processing unit and a parameter shared processing unit for respectively transforming the hidden states of the node layer and the region layer in the spatial dimension, the time dimension and the feature dimension; dynamically fusing the node layer hidden states and the region layer hidden states processed by the space-time module sequence based on an attention mechanism to obtain fusion features; processing the fusion features to output a prediction result of future traffic flow.

[0020] First, a double-layer structure of the traffic network is constructed, one layer is a node layer composed of traffic sensor nodes, such as traffic flow sensors at different positions on urban roads; the other layer is a region layer formed by clustering nodes in the node layer, for example, a plurality of adjacent traffic sensor nodes in the city can be clustered into a region, such as a plurality of traffic nodes in the city center can be clustered into a region layer unit. Through such a double-layer structure, node layer traffic flow data and region layer traffic flow data are extracted respectively, so as to obtain micro node flow information and macro region flow information.

[0021] The node layer traffic flow data and the region layer traffic flow data are respectively mapped to a potential space to obtain an initial hidden state of the node layer and an initial hidden state of the region layer. The actual traffic flow data is converted into an internal representation form that can be understood and processed by the model, preparing for subsequent processing.

[0022] The node layer initial hidden state and the region layer initial hidden state are input into a sequence containing at least one space-time module for processing, so as to capture dynamic space-time dependency. Each space-time module contains a parameter independent processing unit and a parameter shared processing unit. The parameter independent processing unit can perform unique transformation on the hidden state in the spatial dimension, the time dimension and the feature dimension according to the characteristics of the node layer and the region layer, such as the spatial dimension transformation of the node layer can adjust the relationship between the single node and its directly adjacent nodes, and the spatial dimension transformation of the region layer processes the relationship between regions; while the parameter shared processing unit shares some parameters between the node layer and the region layer, and strengthens the common relationship between them, such as in the time dimension, the node layer and the region layer may follow a common time variation mode, and the parameter sharing is used to capture this commonality. Through such processing, the dynamic dependency of traffic flow in space and time can be effectively mined, such as the correlation of the flow changes of adjacent nodes and adjacent regions in a certain time period.

[0023] The node layer hidden state and the region layer hidden state processed by the space-time module sequence are dynamically fused based on the attention mechanism to obtain fusion features. The attention mechanism can make the model pay attention to the features of the node layer or the region layer that are more important for the prediction result when fusing, for example, during the morning peak period, the flow change of some node layer has a greater impact on the overall region, and the attention mechanism will give it a higher weight.

[0024] The fusion features are processed to output the prediction result of future traffic flow, so as to realize accurate prediction of traffic flow and provide strong support for traffic management, travel planning, etc.

[0025] The region layer traffic flow data is obtained by the following method: The K-means clustering algorithm is used to cluster the traffic sensor nodes; The average value of the traffic flow features of all nodes in each cluster is calculated to obtain the traffic flow features of the corresponding area.

[0026] First, the K-means clustering algorithm is used to cluster the traffic sensor nodes. For example, in a city's traffic network, there are many traffic sensor nodes collecting traffic flow data. The K-means algorithm will divide these nodes into different clusters according to their location, traffic variation pattern, and other characteristics. The average value of the traffic flow features of all nodes in each cluster is calculated to obtain the traffic flow features of the corresponding area. For example, a cluster contains multiple traffic sensor nodes in the city center. The average value of the traffic flow data (such as hourly traffic flow and traffic peak feature) of these nodes represents the traffic flow features of the city center, thus forming the regional traffic flow data.

[0027] Based on the attention mechanism, the node layer hidden state and the regional layer hidden state processed by the space-time module sequence are dynamically fused to obtain the fusion features, including the following steps: A matrix describing the relationship between node i and region j is constructed , the elements of which are defined as formula (1); ; (1) Based on the attention mechanism, a dynamic transformation matrix is constructed , the calculation process of which is formula (2) and formula (3); ; (2) (3) Where, represents the hidden features of the node, represents the hidden features of the region, represents the convolution operation of the node layer hidden features with a 1x1 convolution kernel, represents the convolution operation of the region layer hidden features with a 1x1 convolution kernel, and σ represents the sigmoid activation function, * represents matrix multiplication, and ⊙ represents element-wise multiplication. The fusion features are calculated by formula (4) ; ; (4) Where, represents the connection operation.

[0028] A matrix describing the relationship between node i and region j is constructed , where the element is defined as 1 when i equals j, otherwise 0. For example, given nodes 1, 2 and regions 1, 2, the matrix has elements 1 at the positions corresponding to node 1 and region 1, node 2 and region 2, and 0 at the other positions, thus explicitly indicating the correspondence between nodes and regions.

[0029] Then, the dynamic transformation matrix is constructed based on the attention mechanism . The hidden features of nodes are first convolved with a 1x1 convolution kernel to obtain , and the hidden features of regions are convolved with a 1x1 convolution kernel to obtain . Then, the matrix multiplication, bias addition and sigmoid activation function are used to obtain . Then, the is subtracted by its average value and activated by the sigmoid function, and then element-wise multiplied with the matrix to obtain . For example, assuming that the node hidden features and the region hidden features are processed by convolution and other operations, the reflecting the importance between nodes and regions is obtained by calculation, and then the dynamic transformation matrix is obtained by processing, which can dynamically reflect the correlation strength between node and region features.

[0030] Finally, the fusion features are obtained by concatenation operation . The node hidden features Hᴾ are concatenated with the results of element-wise multiplication of and the region hidden features . For example, the features of the nodes and the region features adjusted by are spliced together to form fusion features containing effective information of nodes and regions, which provide comprehensive feature support for subsequent output of traffic flow prediction results.

[0031] The processing process of the node layer hidden state and the region layer hidden state includes: using the lth layer independent parameters to perform spatial, temporal and feature dimension transformation; using cross-layer shared parameters to perform spatial, temporal and feature dimension transformation; fusing each hidden state obtained by independent parameter processing and parameter sharing processing through equation (5) and equation (6); ; (5) ; (6) wherein, These represent the hidden states of the node layer after spatial, temporal, and feature dimension transformations, respectively, under parameter-independent and parameter-shared modes. These represent the hidden states of the region layer after spatial, temporal, and feature dimension transformations under parameter-independent and parameter-shared modes, respectively. The residuals are obtained by performing residual connection and batch normalization using equations (7) and (8); (7) (8) Where ReLU represents the activation function and BN represents the batch normalization operation.

[0032] First, for the hidden states at the node and region levels, we perform parameter-independent and cross-layer-shared dimensional transformations, respectively. Parameter-independent processing uses parameters unique to layer l to perform targeted transformations on the hidden states of the node and region levels in spatial, temporal, and feature dimensions. For example, for the node layer in urban transportation, independent parameters from different layers capture the unique spatial adjacency relationships, temporal patterns, and feature expressions of that layer; independent parameters can uncover unique macroscopic spatiotemporal features between regions. Cross-layer-shared parameters, on the other hand, use a shared set of parameters between the node and region layers to perform spatial, temporal, and feature dimension transformations. The aim is to strengthen the commonalities between the two layers. For example, nodes and regions may both follow a certain periodic temporal dependency, and sharing parameters can simultaneously capture this commonality.

[0033] The hidden states obtained from parameter-independent processing and parameter-sharing processing are fused using equations (5) and (6). Equation (5) represents the multiple hidden states of the node layer after spatial, temporal, and feature dimension transformations under parameter-independent and parameter-sharing modes. ,pass The operations are fused; Equation (6) is the corresponding hidden state of the region layer (such as...). )pass Operation fusion. This step is like integrating the "feature fragments" from the node layer and region layer under different processing modes to form a more comprehensive intermediate representation.

[0034] Features are optimized through residual connections and batch normalization using equations (7) and (8). Equation (7) adds the original hidden state of the node layer to the fused hidden state after ReLU activation, and then performs batch normalization. This alleviates the gradient vanishing problem and makes the feature distribution more stable. For example, the traffic features of the node layer at a certain moment can retain the original information and incorporate newly learned features after residual connection. Equation (8) ensures that the features of the region layer can also be optimized while retaining the original information. Through this series of operations, the hidden states of the node layer and the region layer are fully processed, providing a more robust and expressive input for subsequent feature fusion and traffic prediction.

[0035] It also includes the following steps: Spatial dimension modeling is achieved through spatial dynamic graph convolution, which is constructed based on Tucker decomposition; for the node layer, the spatial dynamic graph convolution operation in parameter sharing mode is Equation (9). (9) in, This represents the convolution operation. This represents the spatial dynamics of the node layer at time t.

[0036] First, spatial dimension modeling is achieved through spatial dynamic graph convolution, which is constructed based on Tucker decomposition. Tucker decomposition can dynamically characterize the structure of a graph. For example, in urban traffic scenarios, the spatial relationships between road nodes change at different times due to traffic flow variations. Tucker decomposition can capture these dynamic spatial dependencies, thereby constructing a dynamic graph that reflects the real-time spatial structure.

[0037] For the spatial dynamic graph convolution operation of the node layer in parameter-sharing mode, as shown in equation (9), through Convolution operations extract the hidden states of node layers. Spatial dynamic graph of the node layer at time t Combined. Taking the urban morning rush hour as an example, the spatial correlation between traffic nodes in a certain area at time t (such as the degree of traffic flow impact at adjacent intersections) is determined by... To indicate, The operation utilizes this dynamic graph. By performing convolution, dynamic dependency features between nodes can be mined in the spatial dimension, enabling the model to accurately capture changes in the spatial relationships of the node layer at different times, and providing accurate spatial dimension information support for subsequent traffic flow prediction.

[0038] It also includes the following steps: Temporal modeling is achieved by combining a multi-scale temporal convolution module and a temporal dynamic graph convolution module; the multi-scale temporal convolution module processes the input... the processing of formula (10), formula (11) and formula (12); ; (10) ; (11) ; (12) , , , respectively represent the convolution operation with the convolution kernel size of 1x2, 1x3, 1x5 and 1x1, represents the normalization operation; For the node layer, the time dynamic graph convolution operation in the parameter sharing mode is formula (13); ; (13) wherein, represents the time dynamic graph of the node layer at time t.

[0039] The modeling of the time dimension is realized by combining the multi-scale time convolution module and the time dynamic graph convolution module. The multi-scale time convolution module aims to capture the dependence of traffic flow at different time scales. Taking the input as an example, the features of the short time scale are processed first: the is convolved by the with the convolution kernel size of 1x2, and then the activation function is used, followed by the convolution with the convolution kernel size of 1x3, and finally the normalization operation is used to obtain , which can capture the short-term change pattern of the flow, such as the flow fluctuation of an intersection within a red-green light cycle. At the same time, the features of the long time scale are processed: the is convolved by the with the convolution kernel size of 1x5, and then the activation and normalization are used to obtain , which can capture the long-term flow pattern such as the flow change trend from the morning peak to the flat peak. Then and are spliced, and then the convolution with the convolution kernel size of 1x1 is used to obtain the output of the multi-scale time convolution module, which integrates the flow features of different time scales.

[0040] Then, for the node layer, the time dynamic graph convolution operation is performed in the parameter sharing mode. As shown in formula (13), the output of the multi-scale time convolution module Time dynamic graph of the node layer at time t Combined. Taking the temporal changes in urban traffic as an example, time t might be during the morning rush hour. This reflects the temporal dependencies between nodes, such as the temporal correlation of traffic flow changes between adjacent nodes during the morning rush hour. Through this operation, the model can simultaneously consider multi-scale traffic flow change patterns and dynamic temporal dependencies between nodes in the time dimension, thereby accurately capturing the complex characteristics of traffic flow in the time dimension and providing reliable time-dimensional information support for subsequent traffic flow prediction.

[0041] It also includes the following steps: The transformation of the feature dimension is independent of the graph convolution operation; for the node layer, the feature transformation in parameter sharing mode is Equation (14). = ;(14); in, This indicates a feature transformation convolution operation.

[0042] It also includes the following steps: The node-level fusion features output by each spatiotemporal module. Perform a skip connection summation, and process it through a convolutional layer as shown in equation (15). ; (15) in, This is the final prediction result.

[0043] First, there's the transformation of the feature dimensions, which is independent of graph convolution operations. For feature transformations at the node layer in parameter-sharing mode, convolution operations are performed through feature transformations. Hidden state of node layer This process involves adjusting the feature dimensions. For example, in traffic flow prediction, the flow features at the node layer may include multiple dimensions such as vehicle volume and average vehicle speed. The operation can map these features to a new dimensional space, making them more suitable for subsequent processing or fusion with other features, while maintaining independence from graph convolution operations and ensuring the targeted nature of feature transformation.

[0044] The next step is to generate the final prediction results, which requires fusing the node-level features output by each spatiotemporal module. The process involves summing skip connections and then processing the data through convolutional layers. First, each layer... After the summation Activation This skip connection step integrates features learned from different spatiotemporal modules. For example, short-term traffic fluctuation features captured by one layer and long-term trend features captured by another layer can be summed to obtain a comprehensive representation. Then, for... Two convolution operations are sequentially performed and ) and are matched with activation, and finally a prediction result is obtained . Taking urban traffic flow prediction as an example, through such an operation, the model can fuse multi-scale and multi-level features, thereby outputting accurate future traffic flow prediction results, and providing a strong basis for traffic management and other decisions.

[0045] The device embodiments described above are merely illustrative, wherein the units described as separate components can or can not be physically separated, and the components displayed as units can or can not be physical units, i.e., they can be located in one place, or distributed on multiple network units. Part or all of the modules can be selected to achieve the purpose of the present embodiment scheme according to actual needs. Those skilled in the art can understand and implement it without creative labor.

[0046] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be realized by means of software and the necessary general hardware platform, and of course, it can also be realized by hardware. Based on such understanding, the above technical solutions, essentially or in other words, the part that contributes to the prior art can be embodied in the form of a software product, which can be stored in a computer readable storage medium, such as a ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute the methods described in each embodiment or some parts of the embodiments.

[0047] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements for some technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A traffic flow prediction method of a double-layer multi-scale dynamic graph convolution network, characterized in that, The method comprises the following steps: A double-layer structure of a traffic network is constructed, including a node layer composed of traffic sensor nodes and a region layer formed by node clusters in the node layer, and node layer traffic flow data and region layer traffic flow data are extracted; The node layer traffic flow data and the region layer traffic flow data are respectively mapped to a latent space to obtain initial hidden states of the node layer and the region layer; The initial hidden states of the node layer and the region layer are input into a sequence containing at least one space-time module for processing to capture dynamic space-time dependencies; each space-time module includes a parameter independent processing unit and a parameter shared processing unit for transforming the hidden states of the node layer and the region layer in the spatial dimension, the temporal dimension and the feature dimension, respectively; Based on an attention mechanism, the node layer hidden state and the region layer hidden state processed by the space-time module sequence are dynamically fused to obtain a fusion feature; The fusion feature is processed to output a prediction result of future traffic flow. 2.The traffic flow prediction method of the double-layer multi-scale dynamic graph convolution network according to claim 1, wherein, The region layer traffic flow data is obtained by the following method: Traffic sensor nodes are clustered using a K-means clustering algorithm; The average value of the traffic flow features of all nodes in each cluster is calculated to obtain the traffic flow features of the corresponding region. 3.The traffic flow prediction method of the double-layer multi-scale dynamic graph convolution network according to claim 1, characterized in that, Based on an attention mechanism, the node layer hidden state and the region layer hidden state processed by the space-time module sequence are dynamically fused to obtain a fusion feature, which comprises the following steps: A matrix describing the relationship between node i and region j is constructed with elements defined as in equation (1); ； （1） Building dynamic transformation matrix based on attention mechanism with the calculation process as formula (2) and formula (3); ; （2） （3） wherein, denotes the hidden features of the node, denotes the hidden features of the region, denotes a convolution operation on the hidden features of the node layer with a 1x1 convolution kernel, denotes a convolution operation on the hidden features of the region layer with a 1x1 convolution kernel, σ denotes a sigmoid activation function, * denotes matrix multiplication, and denotes element-wise multiplication. The fusion features are calculated by formula (4) ; ; （4） wherein represents a concatenation operation.

4. The traffic flow prediction method of claim 3, wherein, processing of the node layer hidden state and the region layer hidden state includes: Use the independent parameters of the lth layer to perform spatial, temporal and feature dimension transformation; Use the cross-layer shared parameters to perform spatial, temporal and feature dimension transformation; Fuse each hidden state obtained by parameter independent processing and parameter shared processing through equation (5) and equation (6); ; （5） ; （6） wherein, are the hidden states of the node layer transformed in the spatial, temporal, and feature dimensions under the parameter independent and parameter shared modes, respectively, are the hidden states of the region layer transformed in the spatial, temporal, and feature dimensions under the parameter independent and parameter shared modes, respectively. Residual connection and batch normalization are performed through equation (7) and equation (8); （7） （8） Where ReLU represents an activation function, and BN represents a batch normalization operation.

5. The traffic flow prediction method of claim 4, wherein, Further comprising the following steps: Spatial dimension modeling is achieved by spatial dynamic graph convolution, and the spatial dynamic graph is constructed based on Tucker decomposition; for the node layer, the spatial dynamic graph convolution operation in the parameter sharing mode is equation (9); ; （9） wherein, denotes a convolution operation, denotes the spatial dynamic graph of the node layer at time t. 6.The traffic flow prediction method of the double-layer multi-scale dynamic graph convolution network according to claim 5, characterized in that, Further comprising the following steps: The modeling of the time dimension is realized by combining a multi-scale time convolution module and a time dynamic graph convolution module; the multi-scale time convolution module processes the input as formulas (10), (11) and (12). ; （10） ; （11） ; （12）; , , , denote convolution operations with kernel size 1x2, 1x3, 1x5 and 1x1, respectively, denotes a normalization operation; For the node layer, the time dynamic graph convolution operation in the parameter sharing mode is equation (13); ； （13） wherein, represents the time dynamic graph of the node layer at time t.

7. The traffic flow prediction method of claim 6, wherein, Further comprising the following steps: The transformation of the feature dimension is independent of the graph convolution operation; for the node layer, the feature transformation in the parameter sharing mode is equation (14); = ； （14）； wherein, denotes a feature transform convolution operation. 8.The traffic flow prediction method of the double-layer multi-scale dynamic graph convolution network according to claim 7, characterized in that, Further comprising the following steps: Fusing features of node layers output by each layer of spatiotemporal modules The skip connection summation is processed through the convolutional layer as formula (15); ； （15） wherein, is the final prediction result.

Images