Traffic flow prediction method based on dynamic multi-view space-time fusion graph convolutional network

Abstract

The invention discloses a traffic flow prediction method based on a dynamic multi-view space-time fusion graph convolutional network, and the method comprises the following steps: collecting road network information, obtaining original data, and carrying out the normalization of the original data; a dynamic multi-view space-time fusion graph convolutional network is constructed, and the network structure is composed of a time decomposition layer, a time gating convolution (TGC) module and an adaptive multi-graph convolution (AMC) module; the time decomposition layer integrates various types of features into dynamic graph embedding, so that the defects of a traditional adjacent matrix in the aspect of simulating a real road network are relieved; the TGC module filters irrelevant information in a plurality of periods through a gating mechanism, so that the model can learn a long-term time mode in traffic data; and the AMC module provides diversified features for the model by using multi-graph embedding, so that the traffic flow prediction precision is further improved.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent transportation and deep learning, and specifically is a traffic flow prediction method based on a dynamic multi-view spatiotemporal fusion graph convolutional network. Background Art

[0002] Traffic prediction is the cornerstone of modern intelligent transportation systems (ITS). By collecting a large amount of data and conducting training and prediction, it is possible to accurately predict traffic conditions in real time, provide decision support for traffic management agencies, and thus optimize road operations, alleviate urban traffic congestion, and reduce accident risks.

[0003] Traffic forecasting is based on the historical records of multiple nodes distributed in the road network to predict future traffic conditions. However, due to the uneven configuration of the urban road network structure, traffic flow data has complex temporal and spatial correlations. There are several problems in the current research in this field: first, data periodicity. The periodicity of traffic data refers to the repetitive pattern or regularity of traffic flow over a period of time. By grasping the periodic changes in data, it is helpful to improve the accuracy of the model; interference from abnormal data. There are some fluctuating data in the data of a day. Studying the characteristic patterns of such abnormal fluctuations can help the model learn the complete traffic data pattern more accurately; spatial heterogeneity. Accurately learning the traffic pattern of a specific section of road is a key issue to improve the efficiency of model prediction. Summary of the invention

[0004] In view of the above problems, the present invention proposes a traffic flow prediction method based on a dynamic multi-view spatiotemporal fusion graph convolutional network. The method collects road network information to obtain raw data, normalizes the raw data through preprocessing, and divides the data set into a training set, a validation set, and a test set according to a certain ratio; constructs a dynamic multi-view spatiotemporal fusion graph convolutional network, the network structure consists of a time decomposition layer, a time gated convolution (TGC) module, and an adaptive multi-graph convolution (AMC) module; the time decomposition layer integrates various types of features into dynamic graph embedding, thereby alleviating the shortcomings of traditional adjacency matrices in simulating real road networks; the TGC module filters irrelevant information in multiple cycles through a gating mechanism, so that the model can learn long-term time patterns in traffic data; the adaptive multi-graph convolution (AMC) module uses multi-graph embedding to provide the model with diversified features, thereby further improving the accuracy of traffic flow prediction.

[0005] Technical solution: A traffic flow prediction method based on dynamic multi-view spatiotemporal fusion graph convolutional network, including the following steps:

[0006] Step 1) Collect road network information to obtain raw data, normalize the raw data through preprocessing, and divide the data set into training set, verification set and test set according to a certain ratio;

[0007] Step 2) construct a convolutional network based on dynamic spatiotemporal correlation graph, the network structure consists of a time decomposition layer, a TGC module and an AMC module;

[0008] Step 3) construct the temporal decomposition layer, the NCE module consists of multi-temporal feature extraction, multi-temporal feature extraction and dynamic graph embedding fusion;

[0009] Step 4) The TGC module filters irrelevant information in multiple cycles through a gating mechanism, enabling the model to learn long-term temporal patterns in traffic data;

[0010] Step 5) The AMC module uses multi-graph embedding to provide the model with diverse features to capture the complex information in the real road network;

[0011] Step 6) Use a dynamic multi-view spatiotemporal fusion graph convolutional network to further improve the accuracy of traffic flow prediction.

[0012] Furthermore, in step 1), given N city nodes in the urban road network, they are connected into a weighted directed graph, defined as G = (V, E, A), where V represents the N nodes of the city, E represents the edge set of the graph, and A∈R N*N Represented as the adjacency matrix of the graph, the problem in this paper is to predict the traffic flow data of the next T steps given the graph G and the information of the historical P steps. The input original data is The output objective function is F and F′ are represented as feature dimensions, Represents the graph structure.

[0013] Furthermore, in the step 2), the dynamic multi-view spatiotemporal fusion graph convolutional network has a network structure consisting of a time decomposition layer, a TGC module and an AMC module.

[0014] In step 2), the dynamic multi-view spatiotemporal fusion graph convolutional network consists of a time decomposition layer, a TGC module and an AMC module. The input of the network is X∈R N*P*F .

[0015] Furthermore, in step 3), a temporal decomposition layer is constructed to integrate various types of features into dynamic graph embedding, thereby alleviating the shortcomings of traditional adjacency matrices in simulating real road networks.

[0016] Step 3-1): Define the diffusion process to extract the hidden time series features contained in the original input X:

[0017] The multi-phase feature extraction module processes the input signal through a linear layer, processes it with average pooling and maximum pooling respectively, and outputs the generated signal through a linear layer. The output signal is and Then the signal E is generated by the Hadamard product and activation function T , the steps are as follows:

[0018] X normal =AvgPooling(X)

[0019] X abnormal =MaxPooling(X)

[0020]

[0021]

[0022]

[0023] Where X normal and X abnormal Represents stable signals and abnormal signals, AvgPooling is the average pooling operation, MaxPooling is the maximum pooling operation, MLP represents the fully connected operation, and Tanh represents the activation function.

[0024] Step 3-2): The multi-temporal feature extraction module generates a multi-temporal feature extraction through linear transformation, which is converted into three subspaces, including a query vector Q, a key vector K and a value vector V. The correlation is calculated by calculating the dot product of Q and K, and the attention weight is calculated using softmax; further, multi-head self-attention is used to connect the various attention results together to obtain the output of the multi-head self-attention module. The steps are as follows:

[0025]

[0026]

[0027]

[0028]

[0029]

[0030]

[0031] Where Q s ,K s ,V s Representing the three subspaces of spatial attention, Q t ,K t ,V t Representing the three subspaces of temporal attention, SpatialAttention(Q s , K s , V s) represents the spatial attention processing process, TemporalAttention(Q t , K t , V t ) represents the temporal attention processing process, d k s and d k t represents the scaling factor, All represent learnable parameters, softmax represents the activation function, Concat represents the concatenation operation, and X sa represents spatial attention, X ta represents temporal attention, E A s Indicates the spatial level of long X sa The splicing result, F A t Indicates time-level long X ta The result of splicing, each header represents an X sa or X ta Calculation; then the processing results of spatiotemporal attention are fused:

[0032]

[0033] in b represents a learnable parameter.

[0034] Step 3-3): Then combine the processing results of spatiotemporal attention and abnormal features:

[0035]

[0036] Finally, we get the time decomposition layer result E F Used for subsequent generation of adaptive matrix.

[0037] Furthermore, in step 4), a TGC module is constructed to filter irrelevant information in multiple cycles through a gating mechanism, so that the model can learn long-term temporal patterns in traffic data. The specific construction method is:

[0038] Step 4-1): Define the diffusion process. The TGC module filters irrelevant information of two different diffusion convolutions in multiple cycles through a gating mechanism:

[0039] X res =Dconv(X t )

[0040] X gate =Dconv(X t )

[0041] Among them, X trepresents a single-step X input, Dconv represents a dilated causal convolution operation, X res and X gate They represent two different diffuse convolution processing results. The gating mechanism is then used to fuse the two information:

[0042] z=σ(W zx X res +W zg X gate )

[0043] r=σ(W rx X res +W rg X gate )

[0044] s=ReLU(W sx X gate +W sg (r*X res ))

[0045] TC=z*X res +(1-z)*s

[0046] Where W zx ,W zg ,W rx ,W rg ,W sx ,W sg All represent learnable parameters, z, r, s all represent intermediate variables of the gating mechanism, ReLU represents the activation function, σ() represents the activation function, and TC represents the output of TGC.

[0047] Furthermore, in step 5), the AMC module is constructed to provide diversified features for the model by using multi-graph embedding, so as to capture the complex information in the real road network. The specific construction method is:

[0048] Step 5-1): Define the diffusion process, extract spatial feature information, and use diffusion graph convolution to learn road network feature information, where the diffusion graph convolution is expressed as:

[0049]

[0050]

[0051] Among them, W k1 , W k2 , W k3 are learnable parameters; P f is the forward transfer matrix, P b is the backward transfer matrix, K is the diffusion order, X temp Represents the output result of the time residual learner; represents the generated adaptive adjacency matrix, E1 and E2 represent learnable matrices;

[0052] Step 5-2): The AMC module is an embedding generated by the time decomposition layer and a randomly generated embedding to jointly construct an adjacency matrix to make up for the shortcomings of the traditional fixed adjacency matrix:

[0053]

[0054]

[0055]

[0056] SC=ReLU(W sc (Z1+Z2)+b sc )

[0057] in represents the multivariate adaptive adjacency matrix, W dk1 ,W dk2 ,W dk3 ,W sc ,b sc represents the learnable parameters. Z1 and Z2 are constructed using two adaptive matrices respectively. SC represents the result of the adaptive multi-graph convolution module, which provides diversified features for the model by fusing multi-graph embedding.

[0058] Furthermore, in step 6), the traffic data set collected from the road network is used to train and test a dynamic multi-view spatiotemporal fusion graph convolutional network, and used for traffic flow prediction. The specific process is as follows:

[0059] Step 6-1): Initialize the number of hidden layer nodes, learning rate, batch size and other parameters;

[0060] Step 6-2): Determine the ratio of training set, validation set and test set, and use mean absolute error as the loss function;

[0061] Step 6-3): Use the time decomposition layer to analyze the hidden features in the historical data, and use the spatiotemporal feature extraction module to simulate the traffic operation mode of the road network;

[0062] Step 6-4): Use the training set to train the dynamic multi-view spatiotemporal fusion graph convolutional network, stop training when the maximum number of iterations is reached, and use the test set to test the model accuracy;

[0063] Step 6-5): Load the optimal model weights and use them for traffic flow prediction.

[0064] Beneficial effects: The traffic flow prediction method of the present invention is aimed at the existing methods that mostly do not consider various hidden features. The present invention is based on a dynamic multi-view spatiotemporal fusion graph convolutional network, including a time decomposition layer, TGC and AMC modules, which improves the prediction accuracy of the model.

[0065] The present invention discloses a traffic flow prediction method based on a dynamic multi-view spatiotemporal fusion graph convolutional network, comprising the following steps: collecting road network information to obtain raw data and normalizing it; constructing a dynamic multi-view spatiotemporal fusion graph convolutional network, wherein the network structure consists of a time decomposition layer, a time gated convolution (TGC) module and an adaptive multi-graph convolution (AMC) module; the time decomposition layer integrates various types of features into dynamic graph embedding, thereby alleviating the shortcomings of traditional adjacency matrices in simulating real road networks; the TGC module filters irrelevant information in multiple cycles through a gating mechanism, so that the model can learn long-term time patterns in traffic data; the adaptive multi-graph convolution (AMC) module uses multi-graph embedding to provide the model with diversified features, thereby further improving the accuracy of traffic flow prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 A schematic diagram of the steps of a traffic flow prediction method based on a dynamic multi-view spatiotemporal fusion graph convolutional network of the present invention;

[0067] Figure 2 This is a model structure diagram of a traffic flow prediction method based on a dynamic multi-view spatiotemporal fusion graph convolutional network of the present invention;

[0068] Figure 3 This is a comparison chart of real data and predicted data of a test set of a traffic flow prediction method based on a dynamic multi-view spatiotemporal fusion graph convolutional network of the present invention. DETAILED DESCRIPTION

[0069] The technical method of the present invention will be further described in detail below in conjunction with the accompanying drawings.

[0070] like Figure 1 The traffic flow prediction method based on dynamic multi-view spatiotemporal fusion graph convolutional network comprises the following steps:

[0071] Step 1) Collect the original traffic flow data of the road network, align the data set according to the time scale after preprocessing, and split it into a training set and a test set according to a fixed ratio;

[0072] In step 1), in the urban road network, N city nodes are given and connected into a weighted directed graph, defined as G = (V, E, A), where V represents the N nodes of the city, E represents the edge set of the graph, and A∈R N*NRepresented as the adjacency matrix of the graph, the problem in this paper is to predict the traffic flow data of the next T steps given the graph G and the information of the historical P steps. The input original data is The output objective function is F and F′ are represented as feature dimensions, Represents the graph structure.

[0073] Step 2) construct a dynamic multi-view spatiotemporal fusion graph convolutional network, which includes a time decomposition layer, TGC and AMC modules;

[0074] In step 2), a dynamic multi-view spatiotemporal fusion graph convolutional network is constructed, which includes a time decomposition layer, a TGC and an AMC module. The input of the network is X∈R N*P*F .

[0075] In step 3), the temporal decomposition layer uses spatiotemporal attention as well as maximum and average pooling to alleviate the shortcomings of the traditional adjacency matrix in simulating the real road network;

[0076] In step 3), the time decomposition layer inputs the original data X:

[0077] Step 3-1): Define the diffusion process and extract the original input X:

[0078] The multi-phase feature extraction module processes the input signal through a linear layer, processes it with average pooling and maximum pooling respectively, and outputs the generated signal through a linear layer. The output signal is and Then the signal E is generated by the Hadamard product and activation function T , the steps are as follows:

[0079] X normal =AvgPooling(X)

[0080] X abnormal =MaxPooling(X

[0081]

[0082]

[0083]

[0084] Where X normal and X abnormalRepresents stable signals and abnormal signals, AvgPooling is an average pooling operation, MaxPooling is a maximum pooling operation, MLP represents a fully connected operation, and Tanh represents an activation function. Step 3-2): Multi-phase feature extraction module generation converts multi-phase feature extraction into three subspaces through linear transformation, including query vector Q, key vector K and value vector V. The correlation is calculated by calculating the dot product of Q and K, and softmax is used to calculate the attention weight; further multi-head self-attention is used to connect the various attention results together to obtain the output of the multi-head self-attention module. The steps are as follows:

[0085]

[0086]

[0087]

[0088]

[0089]

[0090]

[0091] Where Q s ,K s ,V s Representing the three subspaces of spatial attention, Q t ,K t ,V t Representing the three subspaces of temporal attention, SpatialAttention(Q s , K s , V s ) represents the spatial attention processing process, TemporalAttention(Q t , K t , V t ) represents the temporal attention processing process, d k s and d k t represents the scaling factor, All represent learnable parameters, softmax represents the activation function, Concat represents the concatenation operation, and X sa represents spatial attention, X ta represents temporal attention, E A s Indicates the spatial level of long X sa The splicing result, E A t Indicates time-level long X taThe result of splicing, each header represents an X sa or X ta Calculation; then the processing results of spatiotemporal attention are fused:

[0092]

[0093] in b represents a learnable parameter.

[0094] Step 3-3): Then combine the processing results of spatiotemporal attention and abnormal features:

[0095]

[0096] Finally, we get the time decomposition layer result E F Used for subsequent generation of adaptive matrix.

[0097] Step 4) Construct a TGC module to filter irrelevant information in multiple cycles through a gating mechanism, so that the model can learn long-term temporal patterns in traffic data. The specific construction method is:

[0098] Defining the diffusion process, the TGC module filters irrelevant information of two different diffusion convolutions in multiple cycles through a gating mechanism:

[0099] X res =Dconv(X t )

[0100] X gate =Dconv(X t )

[0101] Among them, X t represents a single-step X input, Dconv represents a dilated causal convolution operation, X res and X gate They represent two different diffuse convolution processing results. The gating mechanism is then used to fuse the two information:

[0102] z=σ(W zx X res +W zg X gate )

[0103] r=σ(W rx X res +W rg X gate )

[0104] s=ReLU(W sx X gate +W sg (r*X res ))

[0105] TC=z*X res +(1-z)*s

[0106] Where W zx ,W zg ,W rx ,W rg ,W sx ,W sg All represent learnable parameters, z, r, s all represent intermediate variables of the gating mechanism, ReLU represents the activation function, σ() represents the activation function, and TC represents the output of TGC.

[0107] Step 5) Construct the AMC module to use multi-graph embedding to provide the model with diverse features, thereby capturing the complex information in the real road network. The specific construction method is:

[0108] In the step 5);

[0109] Step 5-1): Step 5-1): Define the diffusion process, extract spatial feature information, and use diffusion graph convolution to learn road network feature information; where the diffusion graph convolution is expressed as:

[0110]

[0111]

[0112] Among them, W k1 , W k2 , W k3 are learnable parameters; P f is the forward transfer matrix, P b is the backward transfer matrix, K is the diffusion order, X temp Represents the output result of the time residual learner; represents the generated adaptive adjacency matrix, E1 and E2 represent learnable matrices;

[0113] Step 5-2): The AMC module is an embedding generated by the time decomposition layer and a randomly generated embedding to jointly construct an adjacency matrix to make up for the shortcomings of the traditional fixed adjacency matrix:

[0114]

[0115]

[0116]

[0117] SC=ReLU(W sc (Z1+Z2)+b sc )

[0118] in represents the multivariate adaptive adjacency matrix, W dk1 ,W dk2 ,W dk3 ,W sc ,b sc represents the learnable parameters. Z1 and Z2 are constructed using two adaptive matrices respectively. SC represents the result of the adaptive multi-graph convolution module, which provides diversified features for the model by fusing multi-graph embedding.

[0119] Step 6) Utilize the road network to collect dynamic multi-view spatiotemporal fusion graph convolutional networks and use them for traffic flow prediction.

[0120] In step 6), the road network collection data set is used to train and test the dynamic multi-view spatiotemporal fusion graph convolutional network, and used for traffic flow prediction. The specific process is:

[0121] Step 6-1): Initialize the number of hidden layer nodes, learning rate, batch size and other parameters;

[0122] Step 6-2): Determine the ratio of training set, validation set and test set, and use mean absolute error as the loss function;

[0123] Step 6-3): Use the time decomposition layer to analyze the hidden features in the historical data, and use the spatiotemporal feature extraction module to simulate the traffic operation mode of the road network;

[0124] Step 6-4): Use the training set to train the dynamic multi-view spatiotemporal fusion graph convolutional network, stop training when the maximum number of iterations is reached, and use the test set to test the model accuracy;

[0125] Step 6-5): Load the optimal model weights and use them for traffic flow prediction.

[0126] The present invention discloses a traffic flow prediction method based on a dynamic multi-view spatiotemporal fusion graph convolutional network, comprising the following steps: collecting road network information to obtain raw data and normalizing it; constructing a dynamic multi-view spatiotemporal fusion graph convolutional network, wherein the network structure consists of a time decomposition layer, a time gated convolution (TGC) module and an adaptive multi-graph convolution (AMC) module; the time decomposition layer integrates various types of features into dynamic graph embedding, thereby alleviating the shortcomings of traditional adjacency matrices in simulating real road networks; the TGC module filters irrelevant information in multiple cycles through a gating mechanism, so that the model can learn long-term time patterns in traffic data; the adaptive multi-graph convolution (AMC) module uses multi-graph embedding to provide the model with diversified features, thereby further improving the accuracy of traffic flow prediction.

[0127] The above description is only a preferred implementation of the present invention under the highway network data set. The protection scope of the present invention is not limited to the above implementation. Any equivalent modifications and other modifications made by ordinary technicians in this field based on the contents disclosed by the present invention should be included in the protection scope recorded in the claims.

Claims

1. A traffic flow prediction method based on dynamic multi-view spatiotemporal fusion graph convolutional network, characterized in that: The following steps are involved: Collect urban road network information to obtain raw data, normalize the raw data through preprocessing, and divide the data set into training set, validation set and test set according to a certain ratio; A dynamic multi-view spatiotemporal fusion graph convolutional network is constructed, which includes a time decomposition layer, a TGC module and an AMC module; the time decomposition layer uses spatiotemporal attention and maximum and average pooling to alleviate the shortcomings of traditional adjacency matrix in simulating real road networks; the TGC module filters irrelevant information in multiple cycles through a gating mechanism, so that the model can learn long-term temporal patterns in traffic data; The AMC module uses multi-graph embedding to provide the model with diverse features, thereby capturing the complex information in the real road network; Using the data set to train the multi-view spatiotemporal attention graph convolutional network; Traffic flow prediction is performed using the trained dynamic multi-view spatiotemporal fusion graph convolutional network.

2. According to claim 1, a traffic flow prediction method based on dynamic multi-view spatiotemporal fusion graph convolutional network is characterized in that: The temporal decomposition layer includes multi-temporal feature extraction, multi-temporal feature extraction and dynamic graph embedding fusion; The multi-phase feature extraction module processes the input signal through a linear layer, processes it with average pooling and maximum pooling respectively, and outputs the generated signal through a linear layer. The output signal is and Then the signal E is generated by the Hadamard product and activation function T , the steps are as follows: X normal =AvgPooling(X) X abnormal =MaxPooling(X) Where X normal and X abnormmal Represents stable signals and abnormal signals, AvgPooling is the average pooling operation, MaxPooling is the maximum pooling operation, MLP represents the fully connected operation, and Tanh represents the activation function; The multi-temporal feature extraction module generates a multi-temporal feature extraction through linear transformation into three subspaces, including query vector Q, key vector K and value vector V. The correlation is calculated by calculating the dot product of Q and K, and the attention weight is calculated using softmax; further, multi-head self-attention is used to connect the various attention results together to obtain the output of the multi-head self-attention module. The steps are as follows: Where Q s , K s , V s Representing the three subspaces of spatial attention, Q t , K t , V t Representing the three subspaces of temporal attention, SpatialAttention(Q s , K s , V s ) represents the spatial attention processing process, TemporalAttention(Q t , K t , V t ) represents the temporal attention processing process, d k s and d k t represents the scaling factor, All represent learnable parameters, softmax represents the activation function, Concat represents the concatenation operation, and X sa represents spatial attention, X ta represents temporal attention, E A s Indicates the spatial level of long X sa The splicing result, E A t Indicates time-level long X ta The result of splicing, each header represents an X sa or X ta Calculation; then the processing results of spatiotemporal attention are fused: in b denotes learnable parameters. Step 3-3): Then combine the processing results of spatiotemporal attention and abnormal features: Finally, we get the time decomposition layer result E F Used for subsequent generation of adaptive matrix.

3. According to claim 2, a traffic flow prediction method based on dynamic multi-view spatiotemporal fusion graph convolutional network is characterized in that: The TGC module filters irrelevant information of two different diffusion convolutions in multiple cycles through a gating mechanism: X res =Dconv(X t ) X gate =Dconv(X t ) Among them, X t represents a single-step X input, Dconv represents a dilated causal convolution operation, X res and X gate They represent two different diffuse convolution processing results respectively; the gating mechanism is then used to fuse the information of the two: z=σ(W zx X res +W zg X gate ) r=σ(W rx X res +W rg X gate ) s=ReLU(W sx X gate +W sg (r*X res )) TC=z*X res +(1-z)*s Where W zx , W zg , W rx , W rg , W sx , W sg All represent learnable parameters, z, r, s all represent intermediate variables of the gating mechanism, ReLU represents the activation function, σ() represents the activation function, and TC represents the output of TGC.

4. According to claim 1, a traffic flow prediction method based on dynamic multi-view spatiotemporal fusion graph convolutional network is characterized in that: The AMC module includes diffusion graph convolution and fusion; wherein the diffusion graph convolution is expressed as: Among them, W k1 , W k2 , W k3 are learnable parameters; P f is the forward transfer matrix, P b is the backward transfer matrix, K is the diffusion order, X temp Represents the output result of the time residual learner; represents the generated adaptive adjacency matrix, E1 and E2 represent learnable matrices; The AMC module is an adjacency matrix constructed by embedding generated by the time decomposition layer and randomly generated embedding to make up for the shortcomings of the traditional fixed adjacency matrix: SC=ReLU(W sc (Z1+Z2)+b sc ) in represents the multivariate adaptive adjacency matrix, W dk1 , W dk2 , W dk3 , W sc , b sc represents the learnable parameters. Z1 and Z2 are constructed using two adaptive matrices respectively. SC represents the result of the adaptive multi-graph convolution module, which provides diversified features for the model by fusing multi-graph embedding.

5. According to claim 1, a traffic flow prediction method based on dynamic multi-view spatiotemporal fusion graph convolutional network is characterized in that: The data set is used to train the dynamic multi-view spatiotemporal fusion graph convolutional network, and the specific steps include: Initialize the number of hidden layer nodes, learning rate, and batchsize parameters; Determine the ratio of training set, validation set and test set, and use mean absolute error as the loss function; The time decomposition layer is used to analyze the hidden features in the historical data, and the spatiotemporal feature extraction module is used to simulate the traffic operation mode of the road network; The training set is used to train the dynamic multi-view spatiotemporal fusion graph convolutional network. The training is stopped when the maximum number of iterations is reached, and the model accuracy is tested using the test set.

Images