Data-driven multi-view traffic prediction method and system
Through the multi-view network (MTGNN) method, the data-driven adaptive neighbor matrix and GRU model are used to solve the limitations of the existing traffic flow prediction model in capturing spatiotemporal dependence, and more accurate traffic flow prediction is achieved, suitable for Various scenarios.
Patent Information
- Application Number
- CN202210881914.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-07-26
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2042-07-26
AI Technical Summary
The existing traffic flow prediction model has limitations in capturing the spatiotemporal dependence of the traffic network, especially ignoring the spatial dependence between nodes, resulting in poor prediction results.
A multi-view network (MTGNN) method is proposed to construct a data-driven adaptive neighbor matrix, combine diffusion convolution and GRU model to extract global spatial features and time domain features to realize traffic prediction at future time steps.
This method can more accurately capture the global spatial characteristics and time domain characteristics of the traffic network, improve the accuracy of traffic flow prediction, is suitable for traffic networks and data sets of different sizes, and shows superiority on actual data sets.
Smart Images

Figure CN115240424B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of traffic prediction, and in particular to a data-driven multi-view traffic prediction method and system. Background Art
[0002] On the one hand, with the acceleration of urbanization, a large number of people have poured into cities. On the other hand, with the improvement of people's living standards, the number of private cars is also increasing rapidly. These have brought tremendous pressure to the original urban transportation system. Therefore, it is urgent to establish a new type of intelligent transportation system.
[0003] Traffic forecasting is a key component of advanced traffic management systems, which can help traffic management departments better carry out traffic planning, traffic management and traffic control. Traffic flow forecasting is a key factor in the successful deployment of intelligent transportation systems (ITS). Real-time and accurate traffic flow forecasting helps with flow control and route planning, and can also help road users make better travel decisions and alleviate traffic congestion.
[0004] With the improvement of transportation infrastructure, various sensors can generate a large amount of traffic data every day, including road probes, cameras, mobile global positioning systems, smart cards for buses and subways, social media, etc., which makes current traffic management and control increasingly data-driven. However, due to the complex spatiotemporal dependencies of traffic data, traffic flow prediction has always been a challenging task.
[0005] Spatially, changes in traffic flow are mainly affected by neighboring nodes in the urban network topology. For example, the traffic flow at a certain point is affected by the transfer effect of upstream road traffic conditions, as well as the feedback effect of downstream traffic conditions, and may even be affected by more distant nodes.
[0006] In the field of transportation, researchers first used statistical methods to conduct research, including ARIMA and its variant, Kalman. Although these methods have a relatively solid theoretical foundation, they do not conform to the highly nonlinear and dynamic characteristics of traffic data, so the actual effect of this method is very poor.
[0007] In order to extract more complex correlations in traffic data, researchers began to try to use machine learning methods to complete more complex data modeling, including SVM, K-nearest neighbor, and Bayesian models.
[0008] However, in actual big data scenarios, since these methods only consider the dynamic changes of traffic conditions and ignore the spatial dependence of road networks, they cannot accurately predict the status at future moments.
[0009] In recent years, with the rapid development of deep learning, researchers have begun to use recurrent neural networks to capture the spatiotemporal dependencies in traffic data. In these research systems, researchers usually use RNN or its variants to extract the temporal correlation in traffic data, use CNN to capture the spatial correlation of grid-based traffic networks, or combine the two to extract the spatiotemporal features of data.
[0010] Although the CNN method can extract the spatial dependency of data to a certain extent, it is essentially applicable to Euclidean spaces such as images and regular grids, and still has limitations for transportation networks with complex topological structures.
[0011] In recent years, with the development of graph neural networks (GNNs), more and more researchers have tried to integrate GCN with RNN to capture the spatiotemporal characteristics of traffic data. However, most of these studies are based on the topological structure of distances measured in physical space. Such graph structures not only have measurement errors, but also cannot well reflect the spatial influence between nodes.
[0012] In recent years, due to the great success of deep learning methods (such as CNN, RNN), a series of networks have been derived to generalize neural networks to arbitrary graph structure data. These networks are classified as graph neural networks (GNN). Graph convolutional networks are the most commonly used type of graph neural networks. They are derived from traditional convolutional neural networks and generate new feature information for each node in the graph through feature aggregation and nonlinear transformation. According to different conversion formulas, they can be divided into space-based methods and spectrum-based methods.
[0013] The spectrum-based graph convolution method introduces filters to define graph convolution from the perspective of graph signal processing. Therefore, the spectrum-based graph convolution can be understood as removing noise from graph signals. In order to simplify the calculation, Kipf et al. used Chebyshev expansion to reduce the complexity of Laplace calculation, and Kipf et al. simplified ChebyNet to a simpler form. The space-based graph convolution method rearranges the vertices into a certain grid form, and then constructs a graph convolution by gathering the information of neighboring nodes through conventional convolution operations. That is, this method smoothes the signal of the node by aggregating and transforming the neighborhood information of the node. For example, Veliˇckovi′c et al. use the attention mechanism to learn the weight between two nodes, and GarphSAGE generates node embeddings by local sampling and aggregation features.
[0014] Spatiotemporal data is a common data type in daily data, and the problem of spatiotemporal data prediction is also a very important research topic in spatiotemporal data mining. Early time series prediction methods based on traditional parameter models, such as HA and ARIMA, hardly consider the mutual influence between different locations, so the prediction performance is poor. Spatiotemporal series prediction methods based on traditional machine learning models, such as decision trees DT and support vector machines SVM, although they are good at handling spatially non-stationary spatiotemporal series data, cannot integrate complex spatial dependencies into the model.
[0015] The spatiotemporal prediction methods based on deep learning can be roughly divided into two categories, one is the method for grid data, and the other is the method for graph data. Among them, the method for grid data is suitable for spatiotemporal data in Euclidean space, and usually extracts spatial information through CNN. For example, ConvLSTM uses CNN to extract spatial information and combines RNN to model spatial and temporal correlations respectively; ST-Res Net introduces a residual module to ensure that the prediction accuracy will not decrease due to the increase in the number of convolutional module layers; ST-3DNet uses 3D convolution to extract features from spatial and temporal dimensions, and models short-term and long-term respectively.
[0016] Graph-oriented methods are applicable to non-Euclidean spatial data and have received increasing attention in the past few years. They aggregate information from graph structures and model nodes and their connectivity. DCRNN uses a diffusion graph convolutional network to describe the information diffusion process in spatial networks and captures temporal dynamics through gated recurrent units (GRUs). STGCN uses ChebNet to extract spatial features in the spatial dimension and uses CNN to model temporal correlations in the temporal dimension from the perspective of the spatial domain.
[0017] GraphWaveNet designs an adaptive matrix to learn the changes in the influence between nodes' expected neighbor nodes and uses dilated random convolutions to simulate temporal correlations. ASTGCN designs a spatiotemporal convolution module that consists of graph convolutions that capture spatial features and time-domain convolutions that describe temporal dependencies.
[0018] Inspired by the above insights, we propose a Multi-View Network (MTGNN) for traffic flow prediction, which aims to comprehensively capture the global spatial characteristics of the traffic network from different views. Summary of the invention
[0019] In view of the shortcomings of the prior art, the present invention discloses a data-driven multi-view traffic prediction method and system, which is used to solve the problem that most of the current mainstream traffic prediction models extract the spatial characteristics of the network from the spatial topological structure of the nodes, thereby ignoring the dependencies between nodes that are similar in time sequence.
[0020] The present invention is achieved through the following technical solutions:
[0021] In a first aspect, the present invention provides a data-driven multi-view traffic prediction method, comprising the following steps:
[0022] Obtain traffic map data, construct a data-driven adaptive neighbor matrix based on node metrics, and input it into model training after SVD decomposition;
[0023] The local spatial features are extracted from the existing neighboring edge matrix and the data-driven adaptive neighboring edge matrix by using the diffusion convolution method respectively;
[0024] By coupling the information of each node with its spatial neighbor nodes and nodes with similar time series, global spatial features are extracted;
[0025] By stacking multiple GRU models, the time domain features of the network are learned to achieve traffic prediction in future time steps.
[0026] Furthermore, the method is data-driven, analyzes the correlation between nodes from the historical data of the nodes, and defines the adaptive adjacency matrix A of the traffic graph. adp ∈R N×N :
[0027] A adp =Softmax(ReLU(E1,E2))
[0028] Among them, E1 and E2 are trainable parameters.
[0029] Furthermore, in the method, based on Norm 2 Define the distance matrix A between nodes D ∈R N×N :
[0030]
[0031] m,p,n=SVD(A D )
[0032] in is the binary norm of the historical data difference between node i and node j, m∈R N×N ,p∈R N ,n∈R N×N They are the node distance matrices A DThe singular decomposition value of .
[0033] Furthermore, in the method, E1 and E2 are initialized as:
[0034]
[0035] During the training process, E1∈R N×N ,E2∈R N×N It is automatically updated to learn the hidden dependencies between different traffic sequences and generate an adaptive adjacency matrix for graph convolution.
[0036] Furthermore, the method extracts the K-step diffusion results of the edge matrix of the traffic graph and the adaptive adjacency matrix of the graph nodes respectively, and fuses the two through a multi-layer coupling module to obtain the final spatial domain features of the nodes:
[0037]
[0038] Where σ represents the linear transformation function, They represent the K-step diffusion results of nodes on the edge matrix and adaptive adjacency matrix of the graph, β is the hidden feature dimension, This is the final spatial feature extraction result, which will be used as the input of the time domain GRU.
[0039] Furthermore, the method uses a multi-view based recurrent network, which is defined as follows:
[0040]
[0041] h t =u t *h t-1 +(1-u t )*c t
[0042] where h t ∈R N×β represents the hidden state of the current layer output, h t-1 represents the hidden state of the previous layer, u t is the update gate, r t is the reset gate, W and b represent the weight and bias terms respectively.
[0043] Furthermore, the method captures the spatial model of characteristic nodes by stacking multiple MGRU layers based on a multi-view recurrent network to achieve traffic prediction in future time steps.
[0044] Furthermore, the method selects RMSE as the training objective and optimizes the multi-step prediction loss along with L2.
[0045] Furthermore, the method extracts the physical space features and temporal space features of the nodes through K-step diffusion, and finally captures the temporal dynamics through a recursive neural network with an encoder-decoder structure.
[0046] In a second aspect, the present invention provides a data-driven multi-view traffic prediction system, comprising a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the memory is coupled to the processor, and when the processor executes the computer program, the steps in the data-driven multi-view traffic prediction method described in the first aspect are implemented.
[0047] The beneficial effects of the present invention are:
[0048] In order to obtain a graph structure that supplements the node correlation outside the real space graph, the present invention constructs a data-driven adaptive neighbor matrix from the perspective of node measurement, which can capture the potential similarity relationship between nodes. Secondly, compared with other multi-view network models that start from the perspective of the long and short cycles of data, the neighbor matrix is constructed from the perspectives of spatial distance measurement and node distance measurement, respectively, so as to obtain more complete global spatial features. The present invention not only does not require more additional information, but also does not require a large amount of training data sets. And it is also generally applicable to traffic networks of different sizes, data sets of different sizes, and data missing situations in reality. BRIEF DESCRIPTION OF THE DRAWINGS
[0049] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0050] Figure 1 This is a schematic diagram of the steps of the data-driven multi-view traffic prediction method;
[0051] Figure 2 It is a node spatial domain feature extraction diagram of an embodiment of the present invention;
[0052] Figure 3 is a multi-time-step traffic prediction graph according to an embodiment of the present invention;
[0053] Figure 4 is a performance comparison diagram of different prediction steps performed on the PEMS04 data set by an embodiment of the present invention;
[0054] Figure 5 is a performance comparison diagram of different prediction steps performed on the PEMS08 data set by an embodiment of the present invention;
[0055] Figure 6 is a performance comparison and change diagram of the embodiment of the present invention on the PEMS04 data set;
[0056] Figure 7 is a performance comparison of the embodiments of the present invention and a change diagram of the PEMS08 data set;
[0057] Figure 8 is a graph comparing the performance of different MTGNN variants on the PEMS04 dataset according to an embodiment of the present invention;
[0058] Fig. 9 It is a graph comparing the performance of different MTGNN variants on the PEMS08 dataset according to an embodiment of the present invention. DETAILED DESCRIPTION
[0059] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0060] Example 1
[0061] Reference Figure 1 As shown, this embodiment provides a data-driven multi-view traffic prediction method, including the following steps:
[0062] Obtain traffic map data, construct a data-driven adaptive neighbor matrix based on node metrics, and input it into model training after SVD decomposition;
[0063] The local spatial features are extracted from the existing neighboring edge matrix and the data-driven adaptive neighboring edge matrix by using the diffusion convolution method respectively;
[0064] By coupling the information of each node with its spatial neighbor nodes and nodes with similar time series, global spatial features are extracted;
[0065] By stacking multiple GRU models, the time domain features of the network are learned to achieve traffic prediction in future time steps.
[0066] First, in order to obtain a graph structure that supplements the node correlation outside the real space graph, this embodiment constructs a data-driven adaptive neighbor matrix from the perspective of node measurement, which can capture the potential similarity relationship between nodes.
[0067] Secondly, compared with other multi-view network models, this embodiment starts from the perspective of long and short periods of data. This embodiment pays attention to the non-strict periodicity of traffic data, and constructs the neighbor matrix from the perspectives of spatial distance measurement and node distance measurement, so as to obtain more complete global spatial features. This method not only does not require more additional information, but also does not require a large amount of training data sets.
[0068] Example 2
[0069] Based on Example 1, this example provides a data-driven method for constructing an adaptive adjacency matrix. Compared with traditional studies that divide the traffic network into multiple segments or grids, the spatial characteristics of the traffic network extracted based on graph convolution are more in line with the characteristics of the topological structure of the traffic network and can better capture spatial locality. However, the measured geographical distance cannot reflect the spatial influence between nodes. Therefore, many scholars analyze the correlation between nodes from the historical data of the nodes based on data-driven analysis by setting trainable parameters. This method not only greatly increases the time for model training, but also increases the difficulty of model interpretation.
[0070] This embodiment starts from the two perspectives of mining the spatial correlation between nodes and enhancing the interpretability of the model, mines the spatial modes of nodes from historical data, and better extracts the spatial characteristics of traffic flow by constructing a multi-view traffic network.
[0071] Therefore, this embodiment is based on data-driven, analyzes the correlation between nodes from the historical data of the nodes, and defines the adaptive adjacency matrix A of the traffic graph adp ∈R N×N :
[0072] A adp =Softmax(ReLU(E1,E2))
[0073] Where E1 and E2 are trainable parameters. In order to solve the problems of convergence difficulty and numerical instability, this embodiment is based on Norm 2 Defines the distance matrix between nodes
[0074]
[0075] in X j ∈R τ×d , is the binary norm of the historical data difference between node i and node j, m∈R N×N ,p∈R N ,n∈R N×N The node distance matrix is The singular decomposition value of . Therefore, E1 and E2 can be initialized as:
[0076]
[0077] In this embodiment, during the training process, E1∈R N×N ,E2∈R N×N It is automatically updated to learn the hidden dependencies between different traffic sequences and generate an adaptive adjacency matrix for graph convolution.
[0078] Example 3
[0079] Based on Example 1, this embodiment provides a spatial feature extraction method based on multiple views. In order to better extract the spatial features of nodes and improve the accuracy of the model, this embodiment extracts the neighborhood features of nodes based on the edge matrix of the traffic graph and the adaptive adjacency matrix of the graph nodes from the perspective of multi-view spatiotemporal fusion.
[0080] In this embodiment, Figure 2 As shown, for the edge matrix of the traffic graph and the adaptive adjacency matrix of the graph nodes, this embodiment extracts their K-step diffusion results respectively, and fuses the two through a multi-layer coupling module to obtain the final spatial domain features of the nodes:
[0081]
[0082] Where σ represents the linear transformation function, They represent the K-step diffusion results of nodes on the edge matrix and adaptive adjacency matrix of the graph, β is the hidden feature dimension, This is the final spatial feature extraction result, which will be used as the input of the time domain GRU.
[0083] Example 4
[0084] Based on Example 1, this embodiment provides a multi-time-step traffic prediction method. In order to extract the time dependency of the traffic network, this embodiment is based on the gated recurrent unit (GRU) in the recurrent neural network (RNN). This embodiment proposes a multi-view based recurrent network (MGRU), which is defined as follows:
[0085]
[0086] h t =u t *h t-1 +(1-u t )*c t
[0087] where h t ∈R N×β represents the hidden state of the current layer output, h t-1represents the hidden state of the previous layer, u t is the update gate, r t is the reset gate, W and b represent the weight and bias respectively. Finally, this embodiment designs the MTGRU network, which captures the spatial model of characteristic nodes by stacking multiple MGRU layers to achieve traffic prediction for the next τ time step, such as Figure 3 shown.
[0088] Finally, this embodiment selects RMSE as the training target of this embodiment, and optimizes the multi-step prediction loss of this embodiment with L2. Therefore, this embodiment designs the loss function of the MTGRU network for multi-step traffic prediction as:
[0089]
[0090] Example 5
[0091] Based on Examples 1-4, this example is experimentally verified. This example uses PEMS03, PEMS04, PEMS07, and PEMS08 data sets collected from four different regions in California. The spatial adjacency network of each data set is constructed from the actual road network and recorded in a window of every five minutes, which means that 12 data points are generated every hour and 288 sets of data are collected every day. As shown in Table 1
[0092] Table 1 Description and statistics of the dataset
[0093]
[0094] The Edges of each data set come from the actual road network. If two adjacent data detectors on the actual road are on the same road, the two corresponding nodes in the Edges are connected.
[0095] The experiments in this embodiment are all completed on a server with the system Ubuntu 20.04, including two NVDIATesla T4 graphics cards, a memory size of 256G, and a CPU model of Intel Xeon Silver 4210.
[0096] For the sake of fairness, this embodiment divides the data set into the same data set as other models, that is, it is divided into training set, validation set and test set in a ratio of 6:2:2, and sets the same time step for each sequence sample, that is, each sequence sample consists of 24 time steps, and selects 12 time steps in the first hour as input, and 12 time steps in the last hour as judgment labels, and uses the Z-Score standardization method to make the processed data obey the distribution law with a mean of 0 and a variance of 1, thereby ensuring the comparability of the data:
[0097]
[0098] in is the mean of the original data, and S is the variance of the original data.
[0099] In order to avoid the influence of validation data and test data, the initialization of the adaptive adjacency matrix depends only on the training data set, and this embodiment evaluates each data set more than 10 times based on the Pytorch framework. In the parameter setting part, this embodiment takes into account the computational efficiency and prediction performance, sets the diffusion step of the graph convolution part to 3, the hidden feature dimension β to 50, the dropout to 0.1, the batch size to 32, the early stop step to 10, the learning rate to 0.001, and uses the Adam optimizer to train the model. Finally, the mean absolute error (MAE), root mean square error (RMSE) and mean absolute percentage error (MAPE) are used as the evaluation indicators of the final model:
[0100]
[0101] in, represents the true value of the sequence sample, Represents the predicted value of the sequence sample, M represents the number of samples, and for the MAE, RMSE and MAPE indicators, the smaller the value, the better the prediction effect.
[0102] This embodiment compares the model of this embodiment with other 9 models on the PEMS03, PEMS04, PEMS07, and PEMS08 data sets. Table 2 shows the average results of the traffic flow prediction performance for the next hour. It can be seen from the data in the table that in terms of the evaluation indicators, the model of this embodiment achieves the best performance on all data sets.
[0103] At the same time, it can also be observed from the data in the table that the prediction results of traditional time series analysis methods on most data sets are not ideal. This is because these methods have limited ability to model complex nonlinear traffic data.
[0104] Relatively speaking, deep learning-based methods have achieved better results overall, and models that consider both temporal and spatial correlations, such as DCRNN and STGCN, are significantly better than traditional deep learning models such as FC-LSTM.
[0105] In addition, compared with the ASTGCN model that constructs a component that captures global heterogeneous dependencies, the model of this embodiment requires less additional information; compared with the STSGCN model that only uses multiplication operations to extract local spatiotemporal dependencies, the model of this embodiment is more robust to missing data values; compared with the Graph WaveNet model that also updates the adaptive adjacency matrix through data-driven, the adaptive adjacency matrix initialization method proposed in this embodiment accelerates the convergence of the model and greatly reduces the training time of the model; compared with the CCRNN model that learns hidden spatiotemporal dependencies through a hierarchical coupling mechanism, the multi-view spatial feature fusion method proposed in this embodiment requires fewer training parameters.
[0106] Table 2 Performance comparison of MTGNN and baseline models on the PEMS dataset
[0107]
[0108]
[0109] To illustrate the prediction performance of the model in this embodiment at a longer step length, Figure 4 and Figure 5 The average prediction results of the model of this embodiment and the CCRNN model for the future 1, 6, and 12 steps over time on the PEMS04 and PEMS08 data sets are shown respectively, where CCRNN (K = 2) and CCRNN (K = 3) are the cases where the coupling layer in the CCRNN model is set to 2 and 3 respectively. For better explanation, this embodiment also draws the total average flow trend graph Flow of nodes over time within a day. It can be seen from the figure that when the flow has an obvious upward or downward trend, the model of this embodiment has better performance in the prediction of longer steps, and is more stable than other models in traffic congestion.
[0110] In order to more intuitively demonstrate the performance improvement of the model in this embodiment at different time periods, Figure 6 and Figure 7 The average prediction results of the model of this embodiment and the CCRNN model for the next 12 steps on the PEMS04 and PEMS08 data sets are shown respectively. In order to more clearly show the advantages of this embodiment, this embodiment also draws the average flow heat map Flow of each node over time within a day, and the heat map Difference after the CCRNN model and the model of this embodiment are made. As can be seen from the figure below, the model of this embodiment usually has better prediction performance at the beginning and end of the traffic flow peak, which also reflects that the model of this embodiment can better extract the spatial characteristics of the traffic network during the period of large changes in traffic flow.
[0111] In order to further study MTGNN, this example is based on the PEMS04 and PEMS08 datasets and compares them with different variants of MTGNN: 1) Only adjacent, using only the adjacent edge matrix; 2) Only adaptive, using only the adaptive adjacent edge matrix. 3) Random E1 E2, randomly initialize E1 and E2 of the adaptive adjacent matrix to satisfy uniform distribution; 4) Trainadjacent, which performs the same training on the adjacent edge matrix as the adaptive adjacent matrix.
[0112] In this embodiment, as shown in Table 3, it can be concluded that the complete MTGCN achieves the best performance, and the spatial feature extraction based on multi-view helps this embodiment to better extract spatial features in the traffic network.
[0113] Table 3 compares the performance of different MTGNN variants
[0114]
[0115] This embodiment Figure 8 and Fig. 9 The average prediction results for the next 12 steps on the PEMS04 and PEMS08 datasets are shown respectively. It can be seen from the data in the figure that, compared with other variants of the model, the model of this embodiment has achieved better performance overall and also achieved better results in predictions with longer step lengths, which also confirms that the model of this embodiment can obtain more spatiotemporal information hidden in the training set data without any external information.
[0116] In summary, the present invention provides a data-driven multi-view spatiotemporal prediction model MTGNN, and successfully applies it to the traffic flow prediction task. In the MTGNN model, the present invention extracts local spatial features from the existing neighbor matrix and the data-driven adaptive neighbor matrix through the diffusion convolution method, and then fuses them through the multi-graph coupling module to extract global spatial features, and finally captures the temporal dynamics through the recursive neural network of the encoder-decoder structure. Finally, the present invention uses a large number of experimental results on four real data sets to verify the superiority of our model over the other nine baselines.
[0117] In order to obtain the graph structure that supplements the node correlation outside the real space graph, this paper constructs a data-driven adaptive neighbor matrix from the perspective of node measurement, which can capture the potential similarity relationship between nodes. Secondly, compared with other multi-view network models that start from the perspective of long and short cycles of data, the neighbor matrix is constructed from the perspective of spatial distance measurement and node distance measurement, respectively, so as to obtain more complete global spatial features.
[0118] The present invention does not require additional information or a large amount of training data sets, and is also generally applicable to traffic networks of different sizes, data sets of different sizes, and situations where data is missing in reality.
[0119] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A data-driven multi-view traffic prediction method, characterized in that: The following steps are involved: Obtain traffic map data, construct a data-driven adaptive adjacency matrix from the node metric perspective, and input it into the multi-view spatiotemporal prediction model MTGNN training after SVD decomposition; constructing a data-driven adaptive adjacency matrix from the node metric perspective is as follows: Based on data-driven, the correlation between nodes is analyzed from the historical data of the nodes, and the adaptive adjacency matrix A of the traffic graph is defined. adp ∈R N×N : YOUR adp =Softmax(ReLU(E1,E2)) Among them, E1 and E2 are trainable parameters; Based on Norm 2 Defines the distance matrix between nodes in is the binary norm of the historical data difference between node i and node j, m∈R N×N , p∈R N×N , n∈R N×N The node distance matrix is The singular decomposition value of ; Initialize E1 and E2 as: During the training process, E1∈R N×N , E2∈R N×N It is automatically updated to learn the hidden dependencies between different traffic sequences and generate an adaptive adjacency matrix for graph convolution; The local spatial features are extracted from the neighboring edge matrix based on the spatial distance of traffic map data and the data-driven adaptive adjacency matrix by using the diffusion convolution method respectively; By coupling the information of each node with its spatial neighbor nodes and nodes with similar time sequence, the global spatial features are extracted and used as the time domain input of the GRU model; By stacking multiple GRU models, the time domain features of the multi-view spatiotemporal prediction model MTGNN network are learned to achieve traffic prediction in future time steps.
2. A data-driven multi-view traffic prediction system, characterized in that: The method comprises a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the memory is coupled to the processor, and when the processor executes the computer program, the steps in the data-driven multi-view traffic prediction method according to claim 1 are implemented.
