CN115240424A

CN115240424A - Multi-view flow prediction method and system based on data driving

Info

Publication number: CN115240424A
Application number: CN202210881914.8A
Authority: CN
Inventors: 刘长征; 王媛源; 张荣华; 高嘉; 刘陕南; 边正龙
Original assignee: Shihezi University
Current assignee: Shihezi University
Priority date: 2022-07-26
Filing date: 2022-07-26
Publication date: 2022-10-25

Abstract

The invention relates to the technical field of traffic prediction, in particular to a data-driven multi-view flow prediction method and a data-driven multi-view space-time prediction model MTGNN. In the MTGNN model, the invention extracts local spatial features from the existing adjacent edge matrix and the data-driven self-adaptive adjacent edge matrix respectively through a diffusion convolution method, then performs fusion through a multi-graph coupling module, extracts global spatial features, and captures time dynamics through a recurrent neural network of an encoder-decoder structure. Finally, a large number of experimental results are obtained on four real data sets, the superiority of the model relative to other nine baselines is verified, and the problem that the current mainstream flow prediction model mostly extracts the spatial characteristics of the network from the spatial topological structure of the nodes, so that the dependency relationship among the nodes similar in time sequence is ignored is solved.

Description

Multi-view flow prediction method and system based on data driving

Technical Field

The invention relates to the technical field of traffic prediction, in particular to a data-driven multi-view traffic prediction method and system.

Background

On one hand, along with the acceleration of the urbanization process, a large number of people are rushed into cities, and on the other hand, along with the improvement of the living standard of people, the number of private cars is also rapidly increased, which brings great pressure to the original urban traffic system, so that the establishment of a novel intelligent traffic system is urgent.

The traffic prediction is a key component of an advanced traffic management system, and can help traffic management departments to better perform traffic planning, traffic management and traffic control. The traffic flow prediction is a key factor in the successful deployment of an Intelligent Traffic System (ITS), and the real-time and accurate traffic flow prediction is beneficial to flow control and route planning, and can also help road users to make better trip decisions and alleviate traffic congestion.

With the improvement of traffic infrastructure, various sensors can generate a large amount of traffic data every day, including probes on roads, cameras, mobile global positioning systems, smart cards for taking buses and subways, social media and the like, so that the current traffic management and control are increasingly driven by data. However, traffic flow prediction has always been a challenging task due to the complex spatio-temporal dependence of traffic data.

Spatially, the change of the traffic flow is mainly influenced by the adjacent nodes in the topological structure of the urban network, for example, the traffic flow at a certain point is influenced by the transfer effect of the traffic condition of the upstream road, and is also influenced by the feedback effect of the traffic condition of the downstream road, and even is influenced by the farther nodes.

In the field of transportation, researchers have first conducted research using statistical methods, including ARIMA and its variants [1], kalman [2], which, while having a relatively firm theoretical basis, have poor practical results due to their non-compliance with the highly nonlinear and dynamic nature of traffic data.

To extract more complex correlations in traffic data, researchers have attempted to use machine learning methods to accomplish more complex modeling of data, including SVM [3], K-nearest [4], bayesian models [5].

However, in an actual big data scene, since these methods only consider the dynamic change of traffic conditions and ignore the spatial dependence of the road network, the state at the future time cannot be accurately predicted.

In recent years, with the rapid development of deep learning, researchers have begun to use recurrent neural networks to capture spatiotemporal dependencies in traffic data. In these research systems, researchers often use RNNs or variants thereof to extract temporal correlations in traffic data, CNNs to capture spatial correlations of grid-based traffic networks, or a fusion of the two to extract spatiotemporal features of data.

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 there is still a limitation on traffic networks with complex topological structures.

In recent years, with the development of Graphical Neural Networks (GNNs) [6], more and more researchers have attempted to fuse GCNs with RNNs to capture spatio-temporal features of traffic data. However, most of these researches are based on a topological structure formed by distances measured in a physical space, and such a graph structure not only has errors in measurement, but also cannot well reflect the spatial influence among nodes.

In recent years, due to the great success of deep learning methods (e.g., CNN, RNN), a series of networks were derived for generalization of neural networks into arbitrary graph structure data, and such networks were classified as Graph Neural Networks (GNN). The graph convolutional network is a most commonly used one of graph neural networks, which is derived from a conventional convolutional neural network, and generates new feature information through feature aggregation and nonlinear transformation for each node in the graph, which can be classified into a space-based method and a spectrum-based method according to different transformation formulas.

The spectral-based graph convolution method [7], from the perspective of graph signal processing, introduces a filter to define the graph convolution, so that the spectral-based graph convolution can be understood as removing noise from the graph signal. To simplify the calculation, kipf et al [8] uses the Chebyshev extension to reduce the complexity of the Laplace calculation, and Kipf [9] et al in turn simplifies the ChebyNet to a simpler form. The graph convolution method [10] based on space rearranges the vertexes into a certain grid form, and then collects the information of neighbor nodes through conventional convolution operation to construct graph convolution, namely, the method smoothes the signals of the nodes through gathering and converting neighborhood information of the nodes. For example, veli \711ckovi' c [11] et al employ an attention mechanism to learn weights between two nodes, and GarphSAGE [12] generates node embedding by locally sampling and aggregating features.

The spatio-temporal data is a common data type in daily data, and the spatio-temporal data prediction problem is also a very important research subject in spatio-temporal data mining. Early time sequence prediction methods based on traditional parameter models, such as HA, ARIMA, etc., almost do not consider the mutual influence between different positions, so the prediction performance is poor. A space-time sequence prediction method based on a traditional machine learning model, such as a decision tree DT [13], a support vector machine SVM [14] and the like, can better process space-time sequence data with unstable space, but can not fuse complex space dependency into the model (integrated inter).

The spatio-temporal prediction method based on deep learning can be roughly divided into two types, one is a method oriented to grid data, and the other is a method oriented to extraction of graph data. The method for grid data is suitable for spatio-temporal data of Euclidean space, usually spatial information is extracted through CNN, for example, convLSTM [15] extracts spatial information by using CNN, and models spatial and temporal correlations respectively by combining RNN; ST-Res Net [16] introduces residual error module, ensures prediction accuracy not to be reduced because of increase of the number of layers of convolution module; ST-3DNet 17 utilizes 3D convolution to extract features from both the spatial and temporal dimensions and models the short and long term, respectively.

The method of facing graph data, which is applicable to non-euclidean spatial data, has also received increasing attention over the last few years to model the nodes and their connectivity by aggregating information from the graph structure [18]. The DCRNN employs a diffusion graph convolution network to describe the information diffusion process in a spatial network and captures temporal dynamics through gated round-robin units (GRUs). STGCN uses ChebNet to extract spatial features in the spatial dimension from the spatial domain perspective and models temporal correlation in the temporal dimension using CNN.

GraphWaveNet designs an adaptive matrix to learn the variation in influence between nodes in anticipation of neighboring nodes and uses a dilated random convolution to model the time correlation. The ASTGCN designs a spatio-temporal convolution module consisting of a graph convolution to capture spatial features and a time-domain convolution to describe time dependencies.

Inspired by the above insight, we propose a multi-view network (MTGNN) for traffic flow prediction, aiming at fully capturing the global spatial features of the traffic network from different views.

Disclosure of Invention

Aiming at the defects of the prior art, the invention discloses a data-driven multi-view traffic prediction method and a data-driven multi-view traffic prediction system, which are used for solving the problem that the current mainstream traffic prediction model mostly extracts the spatial characteristics of a network from the spatial topological structure of nodes, so that the dependency relationship between the nodes similar in time sequence is ignored.

The invention is realized by the following technical scheme:

in a first aspect, the present invention provides a data-driven multi-view traffic prediction method, including the following steps:

acquiring traffic map data, constructing a data-driven self-adaptive adjacent edge matrix according to the node measurement angle, and inputting the self-adaptive adjacent edge matrix into a model for training after SVD (singular value decomposition);

extracting local spatial features from the existing adjacent edge matrix and the data-driven self-adaptive adjacent edge matrix by a diffusion convolution method respectively;

extracting global spatial features by coupling information of each node with spatial neighbor nodes and time sequence similar nodes of each node;

the time domain characteristics of the network are learned by stacking a plurality of GRU models, and the flow prediction of the future time step is realized.

Further, the method is based on data driving, analyzes the correlation among the nodes from the historical data of the nodes, and defines an adaptive adjacency matrix A of the traffic map ^adp ∈R ^N×N ：

A ^adp ＝Softmax(ReLU(E ₁ ,E ₂ ))

Wherein E is ₁ ,E ₂ Are trainable parameters.

Further, in the method, the Norm-based ² Defining a distance matrix A between nodes ^D ∈R ^N×N ：

m,p,n＝SVD(A ^D )

Wherein

Is a two-norm of the difference between the historical data of node i and node j, and m belongs to R ^N×N ,p∈R ^N ,n∈R ^N×N Respectively, are node distance matrix A ^D The singular decomposition values of.

Further, in the method, E ₁ ,E ₂ The initialization is as follows:

during the training, E ₁ ∈R ^N×N ,E ₂ ∈R ^N×N It is automatically updated to learn the hidden dependency relationship between different traffic sequences and to generate an adaptive adjacency matrix for graph convolution.

Furthermore, the method respectively extracts the K-step diffusion results of the edge matrix of the traffic map and the self-adaptive adjacency matrix of the map node, and fuses the edge matrix and the self-adaptive adjacency matrix through a multilayer coupling module to obtain the final spatial domain characteristics of the node:

where sigma denotes a linear transformation function, the transformation function,

respectively representing the K-step diffusion results of the nodes on the edge matrix and the self-adaptive adjacent matrix of the graph, beta is a hidden characteristic dimension,

it is the final spatial feature extraction result, which will be the input of the time domain GRU.

Further, the method uses a multi-view based recursive network, which is defined as follows:

h _t ＝u _t *h _t-1 +(1-u _t )*c _t

wherein h is _t ∈R ^N×β Hidden state, h, representing the current layer output _t-1 Indicating the hidden state of the previous layer, u _t To refresh the door, r _t To reset the gate, W and b represent the weight and bias terms, respectively.

Furthermore, the method is based on a multi-view cyclic recursive network, and the traffic prediction of the future time step is realized by stacking a plurality of MGRU layers to capture a spatial model of the characteristic node.

Further, the method selects RMSE as a training target, and predicts losses with L2 optimization multi-step.

Furthermore, the method extracts the physical space characteristics and the time sequence space characteristics of the nodes through K-step diffusion, and finally captures the time dynamics through a recurrent neural network of an encoder-decoder structure.

In a second aspect, the present invention provides a data-driven-based multi-view traffic prediction system, which includes 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 the processor implements the steps of the data-driven-based multi-view traffic prediction method according to the first aspect when executing the computer program.

The beneficial effects of the invention are as follows:

in order to obtain a graph structure for supplementing the correlation of nodes outside a real space graph, a data-driven self-adaptive adjacent edge matrix is constructed from the perspective of node measurement, and the potential similarity relation between the nodes can be captured. And secondly, relative to other multi-view network models which are sent from the perspective of the long and short periods of data, the adjacent edge matrixes are constructed from the perspective of the spatial distance measurement and the node distance measurement respectively, so that more complete global spatial characteristics are obtained. And the method also has universal applicability to traffic networks of different sizes, data sets of different sizes and data loss situations in reality.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the embodiments or the prior art descriptions will be briefly described below, it is obvious that the drawings in the following description are only some embodiments of the present invention, and other drawings can be obtained by those skilled in the art without creative efforts.

FIG. 1 is a schematic diagram of the steps of a data-driven multi-view traffic prediction method;

FIG. 2 is a node spatial domain feature extraction diagram according to an embodiment of the present invention;

FIG. 3 is a diagram of multi-time-step traffic prediction in accordance with an embodiment of the present invention;

FIG. 4 is a graph comparing the performance of different prediction steps performed on a PEMS04 data set according to an embodiment of the present invention;

FIG. 5 is a graph comparing the performance of various prediction steps performed on a PEMS08 data set according to an embodiment of the present invention;

FIG. 6 is a graph of performance comparison and variation for the PEMS04 data set according to an embodiment of the present invention;

FIG. 7 is a graph of a comparison of performance of an embodiment of the present invention and a variation of the PEMS08 data set;

FIG. 8 is a graph comparing the performance of different MTGNN variants on the PEMS04 data set according to an embodiment of the present invention;

figure 9 is a graph comparing the performance of different MTGNN variants on the PEMS08 data set according to embodiments of the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are some, but not all, embodiments of the present invention. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

Example 1

Referring to fig. 1, the present embodiment provides a data-driven multi-view traffic prediction method, including the following steps:

acquiring traffic map data, constructing a data-driven self-adaptive adjacent edge matrix according to the node measurement angle, decomposing the self-adaptive adjacent edge matrix through SVD (singular value decomposition), and inputting the self-adaptive adjacent edge matrix into a model for training;

First, in order to obtain a graph structure for supplementing the correlation between nodes outside the real space graph, the present embodiment constructs a data-driven adaptive neighboring matrix from the perspective of node measurement, which can capture a potential similarity relationship between nodes.

Secondly, compared with other multi-view network models, the method starts from the perspective of long and short periods of data, the method pays attention to the non-strict periodicity of traffic data, and constructs an adjacent edge matrix from the perspective of spatial distance measurement and node distance measurement respectively, so that more complete global spatial features are obtained.

Example 2

On the basis of embodiment 1, the embodiment provides a data-driven adaptive adjacency matrix construction method, and compared with the traditional research that a traffic network is divided into a plurality of segments or grids, the spatial features of the traffic network extracted based on graph convolution are more consistent with the characteristics of a traffic network topological structure, and the spatial locality can be better captured. However, the measured geographic distance cannot reflect the spatial influence among the nodes, so that many scholars analyze the correlation among the nodes from the historical data of the nodes by setting trainable parameters based on data driving, and the method not only greatly increases the training time of the model, but also increases the difficulty of model interpretation.

In the embodiment, from the two aspects of mining the spatial correlation among the nodes and enhancing the interpretability of the model, the spatial mode of the nodes is mined from the historical data, and the spatial domain characteristics of the traffic flow are better extracted by constructing a multi-view traffic network.

Therefore, the embodiment is based on data driving, analyzes the correlation among nodes from the historical data of the nodes, and defines the adaptive adjacency matrix A of the traffic map ^adp ∈R ^N×N ：

A ^adp ＝Softmax(ReLU(E ₁ ,E ₂ ))

Wherein E ₁ ,E ₂ Are trainable parameters. To solve the problems of difficult convergence and unstable value, the embodiment is based on Norm ² Defining a distance matrix A between nodes ^D ∈R ^N×N ：

m,p,n＝SVD(A ^D )

Wherein

Is a two-norm of the difference between the historical data of node i and node j, and m belongs to R ^N×N ,p∈R ^N ,n∈R ^N×N Respectively, are node distance matrix A ^D Singular decomposition value of [19 ]]. Thus E ₁ ,E ₂ It can be initialized as:

this embodiment is in the course of training, E ₁ ∈R ^N×N ,E ₂ ∈R ^N×N It is automatically updated to learn hidden dependencies between different traffic sequences and to generate an adaptive adjacency matrix for graph convolution.

Example 3

On the basis of embodiment 1, this embodiment provides a spatial feature extraction method based on multiple views, and in order to better extract spatial features of nodes and improve accuracy of a model, this embodiment extracts neighborhood features of nodes from a perspective of multi-view space-time fusion based on an edge matrix of a traffic map and an adaptive adjacency matrix of map nodes respectively.

As shown in fig. 2, in this embodiment, for the edge matrix of the traffic map and the adaptive adjacency matrix of the map node, K-step diffusion results are respectively extracted, and the two are fused by a multilayer coupling module to obtain the final spatial domain characteristics of the node:

where a represents a linear transformation function and where,

Example 4

On the basis of embodiment 1, this embodiment provides a multi-time-step traffic prediction method, and in order to extract the time dependence of the traffic network, this embodiment is based on a gated cyclic unit (GRU) in a Recurrent Neural Network (RNN), and this embodiment proposes a multi-view-based cyclic recursive network (MGRU), which is defined as follows:

h _t ＝u _t *h _t-1 +(1-u _t )*c _t

wherein h is _t ∈R ^N×β Hidden state, h, representing the current layer output _t-1 Indicating the hidden state of the previous layer, u _t To refresh the door, r _t To reset the gate, W and b represent the weight and bias terms, respectively. Finally, the MTGRU network is designed in this embodiment, and a spatial model of a characteristic node is captured by stacking a plurality of MGRU layers, so as to realize traffic prediction at a future τ time step, as shown in fig. 3.

Finally, the embodiment selects RMSE as the training target of the embodiment, and optimizes the multi-step prediction loss of the embodiment along with L2. Therefore, the present embodiment designs the loss function of the MTGRU network for multi-step traffic prediction as:

example 5

Based on examples 1-4, the present example was experimentally verified that the PEMS03, PEMS04, PEMS07, and PEMS08 data sets are collected from four different areas in california, and the spatially adjacent networks of each data set are constructed by actual road networks and are recorded in a summary manner in every five minutes, which means that 12 data points are generated every hour, and 288 data sets are collected every day. As shown in Table 1

TABLE 1 description of data sets and statistical data

The Edges of each data set are from an actual road network, and if two adjacent data detectors on the actual road are on the same road, two nodes corresponding to the Edges are connected.

The experiments of this example were performed on a server with a system of ubuntu 20.04, which contained two video cards of NVDIA Tesla T4, a memory size of 256g and a cpu model of Intel Xeon Silver 4210.

For the sake of fairness, the present embodiment divides the data set into data sets as the other models, that is, according to 6:2:2, dividing the sequence samples into a training set, a verification set and a test set, setting the same time step length for each sequence sample, namely each sequence sample consists of 24 time steps, selecting the first 1 hour and 12 time steps as input, selecting the last 1 hour and 12 time steps as judgment labels, and enabling the processed data to obey the distribution rule that the mean value is 0 and the variance is 1 by a Z-Score standardization method, thereby ensuring the comparability of the data:

wherein

Is the mean of the raw data and S is the variance of the raw data.

To avoid the impact of the validation data and the test data, the initialization of the adaptive adjacency matrix relies only on the training dataset, and the present embodiment evaluates for each dataset more than 10 times based on the pytorech framework. In the parameter setting section, the present embodiment sets the diffusion step size of the graph convolution section to 3, the hidden feature dimension β to 50, dropout to 0.1, batch size to 32, early stop step to 10, the learning rate to 0.001, and trains the model with Adam optimizer in consideration of the calculation efficiency and prediction performance. And finally, respectively adopting the Mean Absolute Error (MAE), the Root Mean Square Error (RMSE) and the Mean Absolute Percentage Error (MAPE) as evaluation indexes of the final model:

wherein the content of the first and second substances,

represents the true value of the sample of the sequence,

represents the predicted values of the sequence samples, M represents the number of samples, and for MAE, RMSE and MAPE indices, a smaller value represents a better prediction.

This example compares the model of this example with the other 9 models on the PEMS03, PEMS04, PEMS07, PEMS08 data sets. Table 2 shows the average results of the traffic flow prediction performance for the 1 hour future. As can be seen from the data in the table, the model of this example achieved the best performance on all data sets in terms of evaluation index.

Meanwhile, the embodiment can also observe from the data in the table that the prediction results of the traditional time series analysis methods on most data sets are not ideal, because the methods have limited capability of modeling complex nonlinear traffic data.

In contrast, the deep learning-based method achieves better results as a whole, and models in which temporal and spatial correlations are considered simultaneously, such as DCRNN, STGCN and the like, are obviously superior to the traditional deep learning model FC-LSTM.

In addition, compared with the construction of the ASTGCN model for capturing the global heterogeneous dependency relationship component, the model of the embodiment needs less additional information; compared with the STSGCN model which only uses multiplication operation to extract the local space-time dependency relationship, the model of the embodiment has stronger robustness to the missing value of the data; compared with the Graph WaveNet model which updates the adaptive adjacency matrix through data driving, the adaptive adjacency matrix initialization method provided by the embodiment accelerates the convergence of the model and greatly reduces the training time of the model; compared with the CCRNN model for learning the hidden space-time dependency relationship through a hierarchical coupling mechanism, the multi-view spatial feature fusion method provided by the embodiment needs fewer training parameters.

TABLE 2 comparison of Performance of MTGNN and baseline models on PEMS datasets

For better illustration, the present embodiment further plots a total average Flow trend graph Flow of nodes in one day along with time change, where the average prediction results of the model and the CCRNN model of the present embodiment along with time change on

future steps

1, 6, and 12 on PEMS04 and PEMS08 data sets are shown in fig. 4 and fig. 5, respectively, where CCRNN (K = 2) and CCRNN (K = 3) are respectively the cases where the coupling layer in the CCRNN model is set to 2 and 3. It can be seen from the figure that when the flow rate has a significant variation trend of ascending or descending, the model of the embodiment has better performance in prediction with a longer step length, and is more stable than other models in the case of traffic jam.

In order to more intuitively show the performance improvement condition of the model of the embodiment in different time periods, fig. 6 and fig. 7 respectively show the average prediction results of the model of the embodiment and the CCRNN model in future 12 steps on PEMS04 and PEMS08 data sets. As can be seen from the following figures, the model of the embodiment can generally have better prediction performance at the beginning and the end of the traffic flow peak, which also shows that the model of the embodiment can better extract the spatial characteristics of the traffic network in the time period of the great change of the traffic flow.

To further study MTGNNs, this example was based on PEMS04 and PEMS08 datasets, which were compared to different variants of MTGNNs: 1) Only using an adjacent edge matrix; 2) Only adaptive neighbor matrix is used. 3) Random E1E 2, performing Random initialization meeting uniform distribution on E1 and E2 of the adaptive adjacent matrix; 4) Train adjacencies which performs the same training of the adaptive adjacency matrix.

In this embodiment, as shown in table 3, it can be concluded that the complete MTGCN achieves the best performance, and the multi-view-based spatial feature extraction is helpful for the embodiment to better extract spatial features in the traffic network.

Table 3 comparison of the Performance of different MTGNN variants

The present embodiment shows the average prediction results of 12 future steps on the PEMS04 and PEMS08 data sets in fig. 8 and fig. 9, and it can be seen from the data in the present embodiment that, compared with other variants of the model, the model of the present embodiment obtains better performance as a whole, and obtains better effect on prediction with a longer step length, which also confirms that the model of the present embodiment can obtain more spatio-temporal information hidden in the training set data without any external information.

In conclusion, the invention provides a data-driven multi-view space-time prediction model MTGNN, and the data-driven multi-view space-time prediction model MTGNN is successfully applied to a traffic flow prediction task. In the MTGNN model, the invention extracts local spatial features from the existing adjacent edge matrix and the data-driven self-adaptive adjacent edge matrix by a diffusion convolution method, then performs fusion by a multi-graph coupling module, extracts global spatial features, and finally captures time dynamics by a recurrent neural network of an encoder-decoder structure. Finally, the invention verifies the superiority of the model relative to other nine baselines through a large number of experimental results on four real data sets.

In order to obtain a graph structure for supplementing the node correlation outside a real space graph, a data-driven self-adaptive adjacent edge matrix is constructed from the node measurement perspective, and the potential similarity relation between nodes can be captured. Secondly, relative to other multi-view network models, the method is sent from the perspective of the long and short periods of data, and the adjacent edge matrixes are constructed from the perspective of the space distance measurement and the node distance measurement respectively, so that more complete global space characteristics are obtained,

the invention does not require more additional information nor a large training data set. And the method also has universal applicability to traffic networks of different sizes, data sets of different sizes and data loss situations in reality.

The documents cited in the present invention are specifically as follows:

[1]G.Yu and C.Zhang,‘Switching ARIMA model based fo recasting for traffic flow’,in 2004IEEE International Conf erence on Acoustics,Speech,and Signal Processing,May 2004,vol.2,pp.ii–429.doi:10.1109/ICASSP.2004.1326286.

[2]I.Okutani and Y.J.Stephanedes,‘Dynamic predicti on of traffic volume through Kalman filtering theory’,Trans portation Research Part B:Methodological,vol.18,no.1,pp.1–11,Feb.1984,doi:10.1016/0191-2615(84)90002-X.

[3]C.-H.Wu,J.-M.Ho,and D.T.Lee,‘Travel-time pre diction with support vector regression’,IEEE Transactions on Intelligent Transportation Systems,vol.5,no.4,pp.276–281,Dec.2004,doi:10.1109/TITS.2004.837813.

[4]M.May,D.Hecker,C.

S.Scheider,and D.Schulz,‘A Vector-Geometry Based Spatial kNN-Algorithm for Traffic Frequency Predictions’,in 2008IEEE International Co nference on Data Mining Workshops,2008,pp.442–447.doi:10.1109/ICDMW.2008.35.

[5]S.Sun,C.Zhang,and G.Yu,‘A bayesian network ap proach to traffic flow forecasting’,IEEE Transactions on In telligent Transportation Systems,vol.7,no.1,pp.124–132,Mar.2006,doi:10.1109/TITS.2006.869623.

[6]F.Scarselli,M.Gori,A.C.Tsoi,M.Hagenbuchner,and G.Monfardini,‘The Graph Neural Network Model’,IEEE Transactions on Neural Networks,vol.20,no.1,pp.61–80,Jan.2009,doi:10.1109/TNN.2008.2005605.

[7]J.Bruna,W.Zaremba,A.Szlam,and Y.LeCun,‘Spectral Networks and Locally Connected Networks on Graphs’,Dec.2013,Accessed:Nov.20,2021.[Online].Available:https://openreview.net/forumid＝DQNsQf-UsoDBa

[8]M.Defferrard,X.Bresson,and P.Vandergheynst,‘Convolutional Neural Networks on Graphs with Fast Localized Spectral Filtering’,in Advances in Neural Information Processing Systems,2016,vol.29.Accessed:Nov.20,2021.[Online].Available:https://proceedings.neurips.cc/paper/2016/hash/04df4d434d481c5bb723be1b6df1ee65-Abstract.html

[9]T.N.Kipf and M.Welling,‘Semi-Supervised Classification with Graph Convolutional Networks’,arXiv:1609.02907[cs,stat],Feb.2017,Accessed:Nov.20,2021.[Online].Available:http://arxiv.org/abs/1609.02907

[10]J.Atwood and D.Towsley,‘Diffusion-Convolutional Neural Networks’,in Advances in Neural Information Processing Systems,2016,vol.29.Accessed:May 25,2022.[Online].Available:https://proceedings.neurips.cc/paper/2016/hash/390e982518a50e280d8e2b535462ec1f-Abstract.html

[11]P.

G.Cucurull,A.Casanova,A.Romero,P.Liò,and Y.Bengio,‘Graph Attention Networks’,arXiv:1710.10903[cs,stat],Feb.2018,Accessed:Nov.20,2021.[Online].Available:http://arxiv.org/abs/1710.10903

[12]W.Hamilton,Z.Ying,and J.Leskovec,‘Inductive Representation Learning on Large Graphs’,in Advances in Neural Information Processing Systems,2017,vol.30.Accessed:May 25,2022.[Online].Available:https://proceedings.neurips.cc/paper/2017/hash/5dd9db5e033da9c6fb5ba83c7a7ebea9-Abstract.html

[13]T.Kim,Y.Yue,S.Taylor,and I.Matthews,‘A Decision Tree Framework for Spatiotemporal Sequence Prediction:21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining’,KDD’15 Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining,pp.577–586,Aug.2015,doi:10.1145/2783258.2783356.

[14]H.Drucker,C.J.C.Burges,L.Kaufman,A.Smola,and V.Vapnik,‘Support Vector Regression Machines’,in Advances in Neural Information Processing Systems,1996,vol.9.Accessed:May 19,2022.[Online].Avai lable:https://proceedings.neurips.cc/paper/1996/hash/d38901788c533e8286cb6400b40b386d-Abstract.html

[15]X.Shi,Z.Chen,H.Wang,D.-Y.Yeung,W.Wong,andW.Woo,‘Convolutional LSTM Network:a machine learning approach for precipitation nowcasting’,in Proceedings of the 28th International Conference on Neural Information Processing Systems-Volume 1,Montreal,Canada,Dec.2015,pp.802–810.

[16]J.Zhang,Y.Zheng,and D.Qi,‘Deep spatio-temporal residual networks for citywide crowd flows prediction’,inProceedings of the Thirty-First AAAI Conference on Artificial Intelligence,San Francisco,California,USA,Feb.2017,pp.1655–1661.

[17]S.Guo,Y.Lin,S.Li,Z.Chen,and H.Wan,‘Deep Spatial–Temporal 3D Convolutional Neural Networks for Traffic Data Forecasting’,IEEE Transactions on Intelligent Transportation Systems,vol.20,no.10,pp.3913–3926,2019,doi:10.1109/TITS.2019.2906365.

[18]J.Zhou et al.,‘Graph neural networks:A review of methods and applications’,AI Open,vol.1,pp.57–81,Jan.2020,doi:10.1016/j.aiopen.2021.01.001.

[19]J.Ye,L.Sun,B.Du,Y.Fu,and H.Xiong,‘Coupled Layer-wise Graph Convolution for Transportation Demand Prediction’,arXiv,arXiv:2012.08080,Dec.2020.doi:10.48550/arXiv.2012.08080.

[20]B.E.Hansen,‘TIME SERIES ANALYSISJames D.Hamilton Princeton University Press,1994’,Econometric Theory,vol.11,no.3,pp.625–630,Jun.1995,doi:10.1017/S0266466600009440.

[21]I.Sutskever,O.Vinyals,and Q.V.Le,‘Sequence to Sequence Learning with Neural Networks’,in Advances in Neural Information Processing Systems,2014,vol.27.Accessed:May 14,2022.[Online].Available:https://proceedings.neurips.cc/paper/2014/hash/a14ac55a4f27472c5d894ec1c3c743d2-Abstract.html

[22]Y.Li,R.Yu,C.Shahabi,and Y.Liu,‘Diffusion Convolutional Recurrent Neural Network:Data-Driven Traffic Forecasting’,arXiv:1707.01926[cs,stat],Feb.2018,Accessed:Nov.20,2021.[Online].Available:http://arxiv.org/abs/1707.01926

[23]B.Yu,H.Yin,and Z.Zhu,‘Spatio-Temporal Graph Convolutional Networks:A Deep Learning Framework for TrafficForecasting’,Proceedings of the Twenty-Seventh International Joint Conference on Artificial Intelligence,pp.3634–3640,Jul.2018,doi:10.24963/ijcai.2018/505.

[24]S.Guo,Y.Lin,N.Feng,C.Song,and H.Wan,‘Attention Based Spatial-Temporal Graph Convolutional Networks for Traffic Flow Forecasting’,Proceedings of the AAAI Conference on Artificial Intelligence,vol.33,no.01,Art.no.01,Jul.2019,doi:10.1609/aaai.v33i01.3301922.

[25]Z.Wu,S.Pan,G.Long,J.Jiang,and C.Zhang,‘Graph wavenet for deep spatial-temporal graph modeling’,in Proceedings of the 28th International Joint Conference on Artificial Intelligence,Macao,China,Aug.2019,pp.1907–1913.

[26]C.Song,Y.Lin,S.Guo,and H.Wan,‘Spatial-Temporal Synchronous Graph Convolutional Networks:A New Framework for Spatial-Temporal Network Data Forecasting’,Proceedings of the AAAI Conference on Artificial Intelligence,vol.34,no.01,Art.no.01,Apr.2020,doi:10.1609/aaai.v34i01.5438.。

the above examples are only intended to illustrate the technical solution of the present invention, but not to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, it should be understood by those of ordinary skill in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some technical features may be equivalently replaced; and such modifications or substitutions do not depart from the spirit and scope of the corresponding technical solutions of the embodiments of the present invention.

Claims

1. A multi-view flow prediction method based on data driving is characterized by comprising the following steps:

extracting local spatial features from an adjacent edge matrix based on spatial distance and a data-driven self-adaptive adjacent edge matrix respectively by a diffusion convolution method;

2. The method according to claim 1, wherein the method is based on dataDriving, analyzing the correlation between the nodes from the historical data of the nodes, and defining the adaptive adjacent matrix A of the traffic map ^adp ∈R ^N×N ：

A ^adp ＝Softmax(ReLU(E ₁ ,E ₂ ))

Wherein E is ₁ ,E ₂ Are trainable parameters.

3. The method according to claim 2, wherein the method is based on Norm ² Defining a distance matrix A between nodes ^D ∈R ^N×N ：

m,p,n＝SVD(A ^D )

Wherein X ⁱ ∈R ^τ×d ，X ⁱ ∈R ^τ×d ，L ₂ (X ⁱ -X ^j ) Is a two-norm of the difference between the historical data of node i and node j, and m belongs to R ^N ^×N ,p∈R ^N ,n∈R ^N×N Are respectively a node distance matrix A ^D The singular decomposition values of.

4. The method according to claim 3, wherein E is the maximum value of the predicted multi-view traffic ₁ ,E ₂ The initialization is as follows:

during the training, E ₁ ∈R ^N×N ,E ₂ ∈R ^N×N The information can be automatically updated according to the information,to learn hidden dependencies between different traffic sequences and to generate an adaptive adjacency matrix for graph convolution.

5. The data-driven-based multi-view traffic prediction method according to claim 1, characterized in that the method extracts K-step diffusion results of an edge matrix of a traffic map and an adaptive adjacency matrix of a map node respectively, and fuses the two through a multilayer coupling module to obtain final spatial domain characteristics of the node:

where a represents a linear transformation function and where,

6. The method according to claim 1, wherein the method uses a multi-view based cyclic recursive network, which is defined as follows:

h _t ＝u _t *h _t-1 +(1-u _t )*c _t

7. The method as claimed in claim 6, wherein the method is based on a multi-view cyclic recursive network, and captures a spatial model of a characteristic node by stacking multiple MGRU layers, so as to realize traffic prediction at a future time step.

8. The data-driven-based multi-view flow prediction method of claim 1, wherein the method selects RMSE as a training target, and is accompanied by L2 optimization multi-step prediction loss.

9. The method according to claim 1, wherein the method extracts physical spatial features and temporal spatial features of nodes through K-step diffusion, and finally captures temporal dynamics through a recurrent neural network of an encoder-decoder structure.

10. A data-driven-based 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 the processor executes the computer program to implement the steps of the data-driven-based multi-view traffic prediction method according to any one of claims 1 to 9.