CN115482656B - Traffic flow prediction method by using space dynamic graph convolutional network - Google Patents
Traffic flow prediction method by using space dynamic graph convolutional network Download PDFInfo
- Publication number
- CN115482656B CN115482656B CN202210559673.5A CN202210559673A CN115482656B CN 115482656 B CN115482656 B CN 115482656B CN 202210559673 A CN202210559673 A CN 202210559673A CN 115482656 B CN115482656 B CN 115482656B
- Authority
- CN
- China
- Prior art keywords
- dynamic
- graph
- network
- node
- matrix
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 29
- 238000009792 diffusion process Methods 0.000 claims abstract description 32
- 230000007246 mechanism Effects 0.000 claims abstract description 19
- 230000004927 fusion Effects 0.000 claims abstract description 15
- 239000011159 matrix material Substances 0.000 claims description 67
- 230000006870 function Effects 0.000 claims description 27
- 230000003068 static effect Effects 0.000 claims description 22
- 230000004913 activation Effects 0.000 claims description 13
- 238000001914 filtration Methods 0.000 claims description 13
- 238000010586 diagram Methods 0.000 claims description 12
- 230000003044 adaptive effect Effects 0.000 claims description 9
- 238000006243 chemical reaction Methods 0.000 claims description 9
- 238000005096 rolling process Methods 0.000 claims description 9
- 230000000694 effects Effects 0.000 claims description 8
- 238000010276 construction Methods 0.000 claims description 7
- 230000002457 bidirectional effect Effects 0.000 claims description 6
- 230000008569 process Effects 0.000 claims description 6
- 238000011144 upstream manufacturing Methods 0.000 claims description 6
- 230000007704 transition Effects 0.000 claims description 5
- 238000005065 mining Methods 0.000 claims description 3
- 230000002441 reversible effect Effects 0.000 claims description 3
- MKXZASYAUGDDCJ-NJAFHUGGSA-N dextromethorphan Chemical compound C([C@@H]12)CCC[C@]11CCN(C)[C@H]2CC2=CC=C(OC)C=C21 MKXZASYAUGDDCJ-NJAFHUGGSA-N 0.000 claims description 2
- 230000009466 transformation Effects 0.000 claims description 2
- 230000007774 longterm Effects 0.000 abstract description 9
- 230000002123 temporal effect Effects 0.000 abstract description 7
- 238000013528 artificial neural network Methods 0.000 description 17
- 230000000306 recurrent effect Effects 0.000 description 11
- 238000013527 convolutional neural network Methods 0.000 description 10
- 230000000875 corresponding effect Effects 0.000 description 9
- 125000004122 cyclic group Chemical group 0.000 description 9
- 239000000284 extract Substances 0.000 description 5
- 230000001276 controlling effect Effects 0.000 description 4
- 238000004880 explosion Methods 0.000 description 4
- 230000009286 beneficial effect Effects 0.000 description 2
- 230000008033 biological extinction Effects 0.000 description 2
- 230000001419 dependent effect Effects 0.000 description 2
- 238000010606 normalization Methods 0.000 description 2
- 206010039203 Road traffic accident Diseases 0.000 description 1
- 230000002596 correlated effect Effects 0.000 description 1
- 238000007405 data analysis Methods 0.000 description 1
- 230000008034 disappearance Effects 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 238000009472 formulation Methods 0.000 description 1
- 239000000203 mixture Substances 0.000 description 1
- 230000001537 neural effect Effects 0.000 description 1
- 210000002569 neuron Anatomy 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 230000001902 propagating effect Effects 0.000 description 1
- 230000006403 short-term memory Effects 0.000 description 1
- 238000013068 supply chain management Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G08—SIGNALLING
- G08G—TRAFFIC CONTROL SYSTEMS
- G08G1/00—Traffic control systems for road vehicles
- G08G1/01—Detecting movement of traffic to be counted or controlled
- G08G1/0104—Measuring and analyzing of parameters relative to traffic conditions
- G08G1/0125—Traffic data processing
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G08—SIGNALLING
- G08G—TRAFFIC CONTROL SYSTEMS
- G08G1/00—Traffic control systems for road vehicles
- G08G1/01—Detecting movement of traffic to be counted or controlled
- G08G1/0104—Measuring and analyzing of parameters relative to traffic conditions
- G08G1/0125—Traffic data processing
- G08G1/0129—Traffic data processing for creating historical data or processing based on historical data
Landscapes
- Physics & Mathematics (AREA)
- Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Analytical Chemistry (AREA)
- Theoretical Computer Science (AREA)
- Chemical & Material Sciences (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Traffic Control Systems (AREA)
Abstract
The invention discloses a traffic flow prediction method by using a space dynamic graph convolution network, which firstly provides a novel attention fusion network comprising a space attention and a gating mechanism so as to generate a dynamic graph at a time node. The network aims to more effectively capture spatial dependencies through dynamic node embedding. Secondly, a dynamic diffusion map convolution loop module is presented that combines bi-directional diffusion convolution with a gated loop unit to capture both global spatial and temporal correlations. After the modules, the invention adds a residual connection switching network with zero initialization to extract long-term features. Finally, the space dynamic graph convolution network is evaluated on two real traffic data sets through a large number of experiments, and the experiments show that the space dynamic graph convolution network is superior to the current most advanced method.
Description
Technical Field
The invention relates to the technical field of traffic data analysis, in particular to a traffic flow prediction method by using a space dynamic graph convolution network.
Background
In intelligent transportation systems (Intelligent Transport System, ITS), spatiotemporal prediction is an important task in real life. Everyone is working to improve the accuracy of predicting future traffic flows. It is beneficial for many applications including autopilot operation, energy and smart grid optimization, logistics and supply chain management. In this context, we have successfully improved predictive performance by addressing dynamic time complexity.
To capture complex time-dependence of historical time series of traffic nodes, recurrent neural networks (Recurrent Neural Network, RNNs) are introduced to predict traffic flow. At the same time, there is a significant spatial correlation between adjacent nodes of the traffic network. Researchers apply convolutional neural networks (Convolutional Neural Networks, CNN) to extract spatial correlations. CNNs can handle mesh structures well. The space-time cyclic convolutional network (Spatiotemporal recurrent convolutional network, SRCN) inherits the advantages of the deep convolutional neural network (deep convolutional neural networks, DCNN) and long short-term memory (LSTM) neural network.
Road networks are a typical non-European type of data, i.e. topology. These methods are poor at capturing non-euclidean correlations to predict traffic time series. To address this problem, most methods use graph rolling networks (Graph Convolutional Networks, GCN) to process spatio-temporal information. The graph convolutional neural network effectively extracts non-grid local spatial features through a Laplacian matrix. The network of graph convolution in combination with a recurrent neural network or Gate-CNN (gated convolutional network) is widely adopted in space-time graph modeling in traffic networks. The network achieves better performance in spatial and temporal modeling than convolutional and recurrent neural networks.
However, in the above prior art, there are the following main drawbacks.
First, the dynamic of efficiently extracting traffic data in both the temporal and spatial dimensions remains challenging. Most methods of graph convolution employ static graphs and fixed adjacency matrices to capture the spatial correlation of node feature representations. However, each node of the traffic network is dynamically changing. Graph convolution based on static graphs cannot dynamically capture the characteristics of nodes. The space-time diagram convolutional network based on the laplacian matrix (space-temporal graph convolutional networks, STGCN) proposed by big Yu et al ignores many useful spatial relationships hidden in traffic data. Gating time convolutional layer (Gated temporal convolution layer, gated TCN) and adaptive adjacency matrix are proposed in Graph WaveNet network to extract spatial and long distance time information, respectively, but it cannot mine correlation between different dynamic graphs
Second, the current network does not have sufficient time series acquisition. The cyclic neural based network is prone to problems of gradient extinction and explosion, so the capture effect is not ideal in long sequence models. Attention-based networks perform well in capturing long-term complexity, but do not efficiently extract local information.
The main reason for the above-mentioned disadvantages is that:
most of the applications of graph convolutional neural networks and recurrent neural networks in traffic flow prediction use a pre-defined matrix and static graph structure, which is constructed based on the distance between sensors in the road nodes. Traffic flow data is highly nonlinear and dynamically changes from time to time due to a variety of factors, so that the correlation between road networks dynamically changes over time, such as: for a certain period of time, two road nodes are physically connected, but in logical space they are weakly connected, weakly correlated. The dynamic graph obtained according to the current road data can truly reflect the strong and weak correlation between nodes, and plays a key role in learning expression of space nodes.
The cyclic neural network involves a large number of derivative operations on time sequences during back propagation, whether the gradient disappears or the gradient explodes, and is derived from the fact that the network structure is too deep, so that the network weight is unstable, and basically, the cyclic neural network is due to the continuous multiplication effect in the back propagation of the gradient. The long-short distance neural network forgets a part of the gradient through the gate function, so that the probability of gradient disappearance can be greatly reduced, but the problem of gradient explosion can not be solved yet. Attention-based networks focus more on the similarity of sequences of time nodes before and after than on local time similarities.
Disclosure of Invention
The technical problem to be solved by the embodiment of the invention is to provide a traffic flow prediction method by using a space dynamic graph convolution network. The method can effectively solve the problems that in traffic flow prediction, nodes of a node relation restriction map fixed in a static map further feature learning and long-time dependency capturing of a circulating neural network are insufficient.
In order to solve the technical problems, an embodiment of the present invention provides a traffic flow prediction method using a space dynamic graph convolutional network, which is characterized by comprising the following steps:
step 1: construction network diagram structureGEstablishing an adjacency matrix by calculating the distance between the sensors and the paired road network and utilizing a threshold Gaussian kernelA；
Step 2: graph fusion network construction, traffic flow data for current time nodeX t First, capturing a global graph node characteristic MAH under the current time by using a spatial attention mechanism t Secondly, adopting a gating mechanism to make each node characteristic H of the previous time period t-1 Node characteristic MAH obtained by spatial attention mechanism under current time t Performing adaptive fusion to obtain attention-based image filtering AF t ；
Step 3: a dynamic diffusion convolution module construction using bi-directional diffusion convolution to capture more of the effects from upstream and downstream traffic flows;
step 4: the time correlation module is constructed by a graph rolling circulation unit and a residual error connection conversion network with zero initialization, wherein the graph rolling circulation unit captures time dependence by using a gating circulation unit and replaces a matrix multiplication part in the gating circulation unit by using a dynamic diffusion convolution module, and the output H of each graph rolling circulation unit at time t t The characteristics of long-time dependence of the residual connection conversion network mining node with zero initialization can be input;
step 5: prediction layer output, HO derived from last layer residual connection transition network output with zero initialization ’(i) And obtaining the final traffic flow prediction output through linear transformation.
2. The traffic flow prediction method using a spatially dynamic graph convolutional network according to claim 1, wherein the adjacency matrix of step 1AThe formula expression is:
where V represents a set of nodes, E represents a set of edges,representative node v i And node v j The distance between K is the threshold value for controlling the sparsity of the adjacency matrix A, sigma is the standard deviation of the distance, +.>An exponential function based on a natural constant e is represented.
Further, the formula expression of the spatial attention mechanism in the step 2 is:
wherein the traffic flow dataLinearly transformed into three matrices Q t =X t W Q ，K t =X t W K And V is equal to t =X t W V ；W Q ，W K ，W V Three learnable parameter matrices; softmax (·) represents the normalized exponential function; attention (&) is a function represented by an implementation Attention mechanism to obtain node characteristics; w (W) o Is a learnable projection matrix; />Indicating the output of the ith attention at time t and Concat (·) indicating the splice operation.
Further, the graph filtering AF in the step 2 t The formula expression of (2) is:
Wz,1，W z,2 ，W h ，b h ，W z,l are all learnable parameters; sigma (·) represents the activation function, which will randomly initialize the model parametersAttention filtering AF t The corresponding elements are multiplied to obtain a dynamic graph node, and the formula is as follows:
the parameter beta is a super parameter for controlling the saturation rate of the activation function;<>representing the Hadamard inner product; tanh (·) is the activation function;seen as source dynamic node embedding, +.>The dynamic adjacency matrix DA is obtained by calculating the similarity between the source dynamic node embedding and the target dynamic node embedding t
ReLU (·) is an activation function, and by constructing a graph fusion network, a corresponding dynamic adjacency matrix is generated at each time node.
Further, the formulation in step 3 to capture more effects from upstream and downstream traffic flows using bi-directional diffusion convolution is:
wherein ,representing a transfer matrix; />Are non-outbound diagonal matrices; a represents a static adjacency matrix; /> and />Representing a transfer matrix P f ，P b Power series of (a); θ k,f ，θ k,b Is a learnable parameter that extracts relationships on spatial nodes from multiple angles, as shown below:
representing a dynamic forward transfer matrix; />Representing a dynamic reverse transfer matrix; θ k,f ，θ k,b ，θ k,df ，θ k,db Are all learnable parameters.
The embodiment of the invention has the following beneficial effects:
1. the static graph adjacency matrix can not completely reflect the node relation under each time node, and has certain limitation on graph node characteristic extraction. The attention fusion network provided by the invention generates attention force diagram filtering at each time node, and the filtering and the learning parameters are combined to generate dynamic node embedding; by calculating the similarity of embedding the dynamic nodes, a dynamic adjacency matrix under each time node can be obtained. Compared with a static fixed adjacency matrix, the dynamic adjacency matrix under each time node reflects the road network node relation according to the current traffic data, so that node characteristic learning is performed.
2. The traditional cyclic neural network has poor effect of capturing long-time dependence, and gradient explosion can occur. The present invention captures the dependency of each node in the time dimension by combining a graph convolution loop unit with a connection transition network with zero initialization residuals. In the graph convolution loop unit, the time and space dependency capture can be performed simultaneously, and the neuron forward propagation capacity of the converter is improved by connecting the zero initialization residual error to the converter so as to explore the deeper long-time dependency.
Drawings
FIG. 1 is a structure of a spatially dynamic graph convolutional neural network;
FIG. 2 is a schematic diagram of an attention fusion network;
fig. 3 is a transducer architecture with zero initialization residual connection.
Detailed Description
The present invention will be described in further detail with reference to the accompanying drawings, for the purpose of making the objects, technical solutions and advantages of the present invention more apparent.
As shown in FIG. 1, the traffic flow prediction method using the space dynamic graph convolutional network can effectively solve the problems of further feature learning of nodes of a node relation restriction graph fixed in a static graph and insufficient long-time dependency capturing of a cyclic neural network in traffic flow prediction. The invention realizes traffic flow prediction according to the following steps:
step 1: constructing a network diagram structure g= (V, E, a). The adjacency matrix A is established by calculating the distance between the sensors and the paired road network and utilizing a threshold Gaussian kernel, and the formula is expressed as follows:
(1)
v represents the set of nodes, E represents the set of edges,representative node v i And node v j The distance between K is the threshold value for controlling the sparsity of the adjacency matrix A, sigma is the standard deviation of the distance, +.>An exponential function based on a natural constant e is represented.
Step 2: and (5) constructing a graph fusion network. The graph fusion network is used to generate a dynamic adjacency matrix. Traffic flow data for current time nodeFirst, using the spatial attention mechanism, as shown in fig. 2, the global graph node feature MAH at the current time is captured as follows t 。
Here, traffic flow dataLinearly transformed into three matrices Q t =X t W Q ，K t =X t W K And V is equal to t =X t W V ；W Q ，W K ，W V Three learnable parameter matrices; softmax (·) represents the normalized exponential function; attention (&) is a function represented by an implementation Attention mechanism to obtain node characteristics; w (W) o Is a learnable projection matrix; />Indicating the output of the ith attention at time t and Concat (·) indicating the splice operation.
Then, the gating mechanism is adopted to characteristic each node in the previous time periodNode characteristic MAH obtained by spatial attention mechanism under current time t Performing adaptive fusion to obtain attention-based image filtering AF t The formula is as follows:
（4）
Wz,1，W z,2 ，W h ，b h ，W z,l are all learnable parameters; sigma (·) represents the activation function, which will randomly initialize the model parametersAttention filtering AF t Corresponding elementMultiplying to obtain a dynamic graph node, wherein the formula is as follows:
（5）
the parameter beta is a super parameter for controlling the saturation rate of the activation function;<>representing the Hadamard inner product; tanh (·) is the activation function;seen as source dynamic node embedding, +.>The dynamic adjacency matrix DA is obtained by calculating the similarity between the source dynamic node embedding and the target dynamic node embedding t 。
ReLU (·) is the activation function. And 2, constructing a fusion network through the step 2, and generating a corresponding dynamic adjacency matrix under each time node. The following steps describe how to combine static and dynamic graphs to perform graph node feature learning.
Step 3: and (5) constructing a dynamic diffusion convolution module. The diffusion convolution recurrent neural network (Diffusion Convolutional Recurrent Neural Network, DCRNN) uses bi-directional diffusion convolution to capture more of the effects from upstream and downstream traffic flows, as shown in (7).
（7）
wherein ,representing a transfer matrix; />Are non-outbound diagonal matrices; as shown in (1), A represents a static adjacency matrix; /> and />Representing a transfer matrix P f ，P b Power series of (a); θ k,f ，θ k,b Is a learnable parameter. The invention improves on the basis of (7), considers the diffusion convolution of the dynamic graph, extracts the relation on the space nodes from a plurality of angles, and has the following formula:
(8)
representing a dynamic forward transfer matrix; />Representing a dynamic reverse transfer matrix; θ k,f ，θ k,b ，θ k,df ，θ k,db All the learnable parameters are learnable parameters. Equation (8) shows that the bidirectional diffusion convolution of the static diagram and the bidirectional diffusion convolution process of the dynamic diagram together form a dynamic diffusion convolution module.
Step 4: and (5) constructing a time correlation module. The time correlation module consists of a graph convolution loop unit (Graph Convolution Recurrent Unit, GCRU) and a residual connection conversion network (rezero transformer) with zero initialization, as shown in fig. 3. In the graph convolution loop unit, the present invention uses a gating loop unit to capture the time dependence. Wherein the dynamic diffusion convolution module replaces a matrix multiplication part in the gating cyclic unit. The process will be the current timeTraffic flow data->The graph convolution loop unit outputs H at time t-1 t-1 And (3) performing splicing, and putting the spliced yarns into a graph convolution circulation unit, wherein the formula is expressed as follows:
(9)
r (t) is a reset gate, u (t) Is an update gate. ⋆ G represents a dynamic diffusion convolution as shown in (8).Is a learnable parameter of the diffusion convolution layer. The temporal and spatial features of the node can be captured simultaneously by means of a graph convolution loop unit.
Next, the output H of each graph convolution loop unit at time t t The characteristics of long-term dependence of the residual connection switching network mining node with zero initialization are input. Residual connection conversion networks (rezerotransformers) with zero initialization are improved over conversion networks (transformers). It promotes forward propagation of the conversion network (transformer) by means of the dynamic equidistant idea to promote its dependency capture at long-term nodes. It removes the original converter normalization process while preserving the position coding, multi-headed attention mechanism and forward propagating modules and uses residual connection with zero initialization (recro). Position-coding PE t Is represented by a sine function and a cosine function:
H t,i representing the output of the graph convolution unit at time t node i, d model Representing the dimensions of the hidden layer. Initially, the nodes at each time are spliced after position codingAs input to the layer 1 rezero switching network. In->Layer rezero converter, input H ,(i-1) Output HO which is the upper layer i-1 ,(i-1) Input is firstly processed by a multi-head attention mechanism, and then residual connection is performed to obtain H ,i And obtaining final output HO of the layer through forward propagation and residual connection ,(i) 。
Alpha is a learnable parameter that the multi-headed attention mechanism shares with forward propagation.
Step 5: and outputting a prediction layer. HO derived from last layer residual connection transition network output with zero initialization ,(i) Obtaining the final traffic flow prediction output through linear transformationThe formula is expressed as follows:
W o and bo Is a learnable parameter.
Finally, a space dynamic graph rolling network can be built through the 5 steps to conduct traffic flow prediction tasks.
Compared with the prior art, the invention has the following advantages:
most of the existing methods adopt a traffic road Network graph structure constructed based on distance, a space-time graph convolution Network (space-Temporal Graph Convolutional Networks, STGCN) carries out graph convolution learning node characteristics based on a static graph structure, meanwhile, a plurality of types of static graphs are constructed from a plurality of semantic angles, such as a road Network graph constructed based on time similarity, a road Network graph constructed based on POI (point of interest) and the like, a graph structure based on road risks is constructed by a geographic semantic space-time Network (A Geographical and Semantic spatial-temporal Network, GSNet), the graph structure based on distance and the graph structure based on POI carries out graph convolution to carry out traffic accident prediction. Such methods are summarized as follows:
(13) The formula is to construct the corresponding graph structure according to different semantic angles.Representing an adjacency matrix corresponding to the corresponding semantics; />Representing the corresponding similarity function. (14) is a graph convolution formula. The graph Fourier radix U is a matrix formed by eigenvectors of the normalized graph Laplace matrix L. Drawing Laplace matrix>，I n Is an identity matrix, D is a diagonal matrix, wherein +.>⋀ is a diagonal matrix, in which the eigenvalues of L are stored. />Representing a filter, and a represents the constructed adjacency matrix therein.
However, the graph structures are static graphs, and the limited node relation representation limits feature learning in the graph node space. To overcome the difficulty of a static graph with limited representation of node relationships, an adaptive matrix is proposed that can be updated in an end-to-end learning process. The graph WaveNet proposes an adaptive matrix to learn graph nodes, but the adaptive matrix cannot achieve dynamics, and in each cycle, the graph of each time node is static, and there is a possibility that node relations with large area missing exist. Graph wavenet uses diffusion convolution with adaptive adjacency matrix and static graph to perform graph node feature learning summarized as follows:
for formula (15), E 1 Embedding for the source node; e (E) 2 Embedding for the target node; sosftmax (·) is the normalization function and ReLU (·) is the activation function;is an adaptive adjacency matrix. For formula (16), p k A power series representing a transition matrix;is a learnable parameter; x is a traffic signal.
In long-term dependent capture, most use is made of recurrent neural networks, which are prone to gradient extinction or gradient explosion in long-sequence-dependent capture. A diffusion convolution recurrent neural network (Diffusion Convolutional Recurrnet Neural Network, DCRNN) employs a gated loop unit and replaces the matrix multiplied portion of the gated loop unit with diffusion convolution to capture both spatial and temporal dependencies. The time convolution network (Temporal Convolutional Network, TCN) is a generic structure that extracts features in the time dimension, and Graph WaveNet proposes to combine time convolution with a gatekeeper to capture the time dependence. However, to capture a long-term sequence over time, TCN requires the superposition of multiple convolutional layers to connect any two positions in the sequence. But this compromises the ability of TCNs to learn long-term dependencies.
In the space dynamic graph rolling network, the self-adaptive adjacency matrix of (15) is different from the static graph constructed by the formula (13). Capturing the characteristics of the current time node by using a door mechanism in an attention fusion network, adaptively fusing the characteristics of the current time node with the characteristics of the previous time node to obtain attention filtering, generating dynamic node embedding pairs by the attention filtering according to formulas (5) and (6), and obtaining a dynamic adjacent matrix DA corresponding to each time node by calculating the similarity of the dynamic node embedding pairs t . As shown in equation (8), the dynamic adjacency matrix and the static adjacency matrix are applied in a bi-directional diffusion convolution to capture the correlation of the nodes of the upstream and downstream spatial effects. Meanwhile, in a graph convolution cyclic unit (GCRU), matrix multiplication of the gating cyclic unit is replaced with a dynamic diffusion convolution module to capture the correlation of time and space simultaneously. In long-term dependency capture, unlike the past methods, as shown in equation (10), a residual join (recro) with zero initialization is applied into the conversion network (transducer) to explore deep long-term dependencies.
The above disclosure is only a preferred embodiment of the present invention, and it is needless to say that the scope of the invention is not limited thereto, and therefore, the equivalent changes according to the claims of the present invention still fall within the scope of the present invention.
Claims (5)
1. A method for traffic flow prediction using a spatial dynamic graph convolutional network, comprising the steps of:
step 1: construction network diagram structureGBy calculating the distance between the sensors and the roads, an adjacency matrix is established by using a threshold Gaussian kernelA；
Step 2: graph fusion network construction, traffic flow data for current time nodeX t First, capturing a global graph node characteristic MAH under the current time by using a spatial attention mechanism t Secondly, adopting a gating mechanism to make each node characteristic H of the previous time period t-1 With the section obtained by the spatial attention mechanism at the current timePoint characteristics MAH t Performing adaptive fusion to obtain attention-based image filtering AF t The method comprises the steps of carrying out a first treatment on the surface of the Will randomly initialize model parametersAttention filtering AF t Multiplying corresponding elements to obtain dynamic graph nodes, and obtaining dynamic adjacency matrix DA according to the dynamic graph nodes t The method comprises the steps of carrying out a first treatment on the surface of the Through the construction of the graph fusion network, a corresponding dynamic adjacency matrix is generated under each time node;
step 3: a dynamic diffusion convolution module construction using bi-directional diffusion convolution to capture more of the effects from upstream and downstream traffic flows; obtaining the bidirectional diffusion convolution of the static diagram according to the static adjacent matrix A, obtaining the bidirectional diffusion convolution of the dynamic diagram according to the dynamic adjacent matrix DAt, and forming a dynamic diffusion convolution module together by the bidirectional diffusion convolution of the static diagram and the bidirectional diffusion convolution process of the dynamic diagram;
step 4: the time correlation module is constructed by a graph rolling circulation unit and a residual error connection conversion network with zero initialization, wherein the graph rolling circulation unit captures time dependence by using a gating circulation unit and replaces a matrix multiplication part in the gating circulation unit by using a dynamic diffusion convolution module, and the output H of each graph rolling circulation unit at time t t Is input with the characteristic of long-time dependence of the residual connection conversion network mining node with zero initialization to obtain output HO ’(i) ；
Step 5: prediction layer output, HO derived from last layer residual connection transition network output with zero initialization ’(i) And obtaining the final traffic flow prediction output through linear transformation.
2. The traffic flow prediction method using a spatially dynamic graph convolutional network according to claim 1, wherein the adjacency matrix of step 1AThe formula expression is:
；
where V represents a set of nodes, E represents a set of edges,representative node v i And node v j The distance between K is the threshold value for controlling the sparsity of the adjacency matrix A, sigma is the standard deviation of the distance, +.>An exponential function based on a natural constant e is represented.
3. The traffic flow prediction method using the spatial dynamic convolutional network according to claim 2, wherein the formula of the spatial attention mechanism in step 2 is:
；
；
wherein the traffic flow data X t Linearly transformed into three matrices Q t =X t W Q ，K t =X t W K And V is equal to t =X t W V ；W Q ，W K ，W V Three learnable parameter matrices; softmax (·) represents the normalized exponential function; attention (&) is a function represented by an implementation Attention mechanism to obtain node characteristics; w (W) o Is a learnable projection matrix;indicating the output of the ith attention at time t and Concat (·) indicating the splice operation.
4. Traffic flow prediction using spatially dynamic graph convolutional network of claim 3The method is characterized in that the image filtering AF in the step 2 t The formula expression of (2) is:
；
Wz,1，W z,2 ，W h ，b h are all learnable parameters; sigma (·) represents the activation function, which will randomly initialize the model parametersAttention filtering AF t The corresponding elements are multiplied to obtain a dynamic graph node, and the formula is as follows:
；
the parameter beta is a super parameter for controlling the saturation rate of the activation function;< >representing the Hadamard inner product; tanh (·) is the activation function;seen as source dynamic node embedding, +.>The dynamic adjacency matrix DA is obtained by calculating the similarity between the source dynamic node embedding and the target dynamic node embedding t
；
ReLU (·) is an activation function, and by constructing a graph fusion network, a corresponding dynamic adjacency matrix is generated at each time node.
5. The traffic flow prediction method using a spatial dynamic graph convolution network according to claim 4, wherein the formula for capturing more influence from upstream and downstream traffic flows using bi-directional diffusion convolution in step 3 is:
；
wherein ,representing a transfer matrix; />Are non-outbound diagonal matrices; /> and />Representing a transfer matrix P f ，P b Power series of (a); />Representing a dynamic forward transfer matrix; />Representing a dynamic reverse transfer matrix; θ k,f ，θ k,b ，θ k,df ，θ k,db Are all learnable parameters.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210559673.5A CN115482656B (en) | 2022-05-23 | 2022-05-23 | Traffic flow prediction method by using space dynamic graph convolutional network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210559673.5A CN115482656B (en) | 2022-05-23 | 2022-05-23 | Traffic flow prediction method by using space dynamic graph convolutional network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN115482656A CN115482656A (en) | 2022-12-16 |
| CN115482656B true CN115482656B (en) | 2023-09-26 |
Family
ID=84420776
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210559673.5A Active CN115482656B (en) | 2022-05-23 | 2022-05-23 | Traffic flow prediction method by using space dynamic graph convolutional network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115482656B (en) |
Families Citing this family (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115866658B (en) * | 2023-02-06 | 2023-05-16 | 湖北工业大学 | Flow prediction method and system based on dynamic space-time hypergraph convolutional network |
| CN115883424B (en) * | 2023-02-20 | 2023-05-23 | 齐鲁工业大学(山东省科学院) | Method and system for predicting flow data between high-speed backbone networks |
| CN116504060B (en) * | 2023-05-01 | 2024-05-14 | 兰州理工大学 | Diffusion diagram attention network traffic flow prediction method based on Transformer |
Citations (19)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111260131A (en) * | 2020-01-16 | 2020-06-09 | 汕头大学 | Short-term traffic flow prediction method and device |
| CN111696355A (en) * | 2020-06-29 | 2020-09-22 | 中南大学 | Dynamic graph convolution traffic speed prediction method |
| CN111899510A (en) * | 2020-07-28 | 2020-11-06 | 南京工程学院 | Intelligent traffic system flow short-term prediction method and system based on divergent convolution and GAT |
| CN112071065A (en) * | 2020-09-16 | 2020-12-11 | 山东理工大学 | Traffic flow prediction method based on global diffusion convolution residual error network |
| CN112183862A (en) * | 2020-09-29 | 2021-01-05 | 长春理工大学 | Traffic flow prediction method and system for urban road network |
| CN112241814A (en) * | 2020-10-20 | 2021-01-19 | 河南大学 | Traffic prediction method based on reinforced space-time diagram neural network |
| CN113283581A (en) * | 2021-05-14 | 2021-08-20 | 南京邮电大学 | Multi-fusion graph network collaborative multi-channel attention model and application method thereof |
| CN113313947A (en) * | 2021-05-31 | 2021-08-27 | 湖南大学 | Road condition evaluation method of short-term traffic prediction graph convolution network |
| CN113450568A (en) * | 2021-06-30 | 2021-09-28 | 兰州理工大学 | Convolutional network traffic flow prediction method based on space-time attention mechanism |
| CN113487088A (en) * | 2021-07-06 | 2021-10-08 | 哈尔滨工业大学（深圳） | Traffic prediction method and device based on dynamic space-time diagram convolution attention model |
| CN113763710A (en) * | 2021-09-16 | 2021-12-07 | 汕头大学 | Short-term traffic flow prediction method based on nonlinear adaptive system |
| CN114038200A (en) * | 2021-11-29 | 2022-02-11 | 东北大学 | Attention mechanism-based time-space synchronization map convolutional network traffic flow prediction method |
| CN114120652A (en) * | 2021-12-21 | 2022-03-01 | 重庆邮电大学 | Traffic flow prediction method based on dynamic graph neural network |
| CN114220271A (en) * | 2021-12-21 | 2022-03-22 | 南京理工大学 | Traffic flow prediction method, equipment and storage medium based on dynamic space-time graph convolution cycle network |
| WO2022061170A1 (en) * | 2020-09-18 | 2022-03-24 | Visa International Service Association | Dynamic graph node embedding via light convolution |
| CN114239974A (en) * | 2021-12-21 | 2022-03-25 | 清华大学 | Multi-agent position prediction method and device, electronic equipment and storage medium |
| CA3131843A1 (en) * | 2020-09-25 | 2022-03-25 | Royal Bank Of Canada | System and method for structure learning for graph neural networks |
| CN114372503A (en) * | 2021-12-03 | 2022-04-19 | 同济大学 | Cluster vehicle motion trail prediction method |
| CN114495500A (en) * | 2022-01-26 | 2022-05-13 | 北京工业大学 | Traffic prediction method based on dual dynamic space-time diagram convolution |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11217012B2 (en) * | 2018-11-16 | 2022-01-04 | Uatc, Llc | System and method for identifying travel way features for autonomous vehicle motion control |
| US20210383228A1 (en) * | 2020-06-05 | 2021-12-09 | Deepmind Technologies Limited | Generating prediction outputs using dynamic graphs |
-
2022
- 2022-05-23 CN CN202210559673.5A patent/CN115482656B/en active Active
Patent Citations (19)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111260131A (en) * | 2020-01-16 | 2020-06-09 | 汕头大学 | Short-term traffic flow prediction method and device |
| CN111696355A (en) * | 2020-06-29 | 2020-09-22 | 中南大学 | Dynamic graph convolution traffic speed prediction method |
| CN111899510A (en) * | 2020-07-28 | 2020-11-06 | 南京工程学院 | Intelligent traffic system flow short-term prediction method and system based on divergent convolution and GAT |
| CN112071065A (en) * | 2020-09-16 | 2020-12-11 | 山东理工大学 | Traffic flow prediction method based on global diffusion convolution residual error network |
| WO2022061170A1 (en) * | 2020-09-18 | 2022-03-24 | Visa International Service Association | Dynamic graph node embedding via light convolution |
| CA3131843A1 (en) * | 2020-09-25 | 2022-03-25 | Royal Bank Of Canada | System and method for structure learning for graph neural networks |
| CN112183862A (en) * | 2020-09-29 | 2021-01-05 | 长春理工大学 | Traffic flow prediction method and system for urban road network |
| CN112241814A (en) * | 2020-10-20 | 2021-01-19 | 河南大学 | Traffic prediction method based on reinforced space-time diagram neural network |
| CN113283581A (en) * | 2021-05-14 | 2021-08-20 | 南京邮电大学 | Multi-fusion graph network collaborative multi-channel attention model and application method thereof |
| CN113313947A (en) * | 2021-05-31 | 2021-08-27 | 湖南大学 | Road condition evaluation method of short-term traffic prediction graph convolution network |
| CN113450568A (en) * | 2021-06-30 | 2021-09-28 | 兰州理工大学 | Convolutional network traffic flow prediction method based on space-time attention mechanism |
| CN113487088A (en) * | 2021-07-06 | 2021-10-08 | 哈尔滨工业大学（深圳） | Traffic prediction method and device based on dynamic space-time diagram convolution attention model |
| CN113763710A (en) * | 2021-09-16 | 2021-12-07 | 汕头大学 | Short-term traffic flow prediction method based on nonlinear adaptive system |
| CN114038200A (en) * | 2021-11-29 | 2022-02-11 | 东北大学 | Attention mechanism-based time-space synchronization map convolutional network traffic flow prediction method |
| CN114372503A (en) * | 2021-12-03 | 2022-04-19 | 同济大学 | Cluster vehicle motion trail prediction method |
| CN114120652A (en) * | 2021-12-21 | 2022-03-01 | 重庆邮电大学 | Traffic flow prediction method based on dynamic graph neural network |
| CN114220271A (en) * | 2021-12-21 | 2022-03-22 | 南京理工大学 | Traffic flow prediction method, equipment and storage medium based on dynamic space-time graph convolution cycle network |
| CN114239974A (en) * | 2021-12-21 | 2022-03-25 | 清华大学 | Multi-agent position prediction method and device, electronic equipment and storage medium |
| CN114495500A (en) * | 2022-01-26 | 2022-05-13 | 北京工业大学 | Traffic prediction method based on dual dynamic space-time diagram convolution |
Non-Patent Citations (5)
| Title |
|---|
| Graph Attention Convolutional Network: Spatiotemporal Modeling for Urban Traffic Prediction;Qingyu Song等;《2020 IEEE 23rd International Conference on Intelligent Transportation Systems (ITSC)》;1-6 * |
| ST-TrafficNet: A Spatial-Temporal Deep Learning Network for Traffic Forecasting;Huakang Lu等;《Electronics》;1-17 * |
| 基于图卷积网络的交通预测综述;王竟成等;《北京工业大学学报》(第8期);954-970 * |
| 基于异常值识别卡尔曼滤波器的短期交通流预测;白伟华等;《计算机应用研究》(第3期);817-821 * |
| 基于时空动态图卷积神经网络的交通流预测方法;郑雨馨;《中国优秀硕士学位论文全文数据库 工程科技Ⅱ辑》(第2期);C034-630 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN115482656A (en) | 2022-12-16 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN115482656B (en) | Traffic flow prediction method by using space dynamic graph convolutional network | |
| Saputra et al. | Learning monocular visual odometry through geometry-aware curriculum learning | |
| CN112071065A (en) | Traffic flow prediction method based on global diffusion convolution residual error network | |
| Chen et al. | 3D point cloud semantic segmentation toward large-scale unstructured agricultural scene classification | |
| CN115713679A (en) | Target detection method based on multi-source information fusion, thermal infrared and three-dimensional depth map | |
| CN115829171B (en) | Pedestrian track prediction method combining space-time information and social interaction characteristics | |
| CN113807318B (en) | Action recognition method based on double-flow convolutional neural network and bidirectional GRU | |
| CN112163601A (en) | Image classification method, system, computer device and storage medium | |
| CN114493014A (en) | Multivariate time series prediction method, multivariate time series prediction system, computer product and storage medium | |
| CN110570035A (en) | people flow prediction system for simultaneously modeling space-time dependency and daily flow dependency | |
| Wang et al. | Quality-aware dual-modal saliency detection via deep reinforcement learning | |
| Foroughi et al. | Mapsegnet: A fully automated model based on the encoder-decoder architecture for indoor map segmentation | |
| CN116596109A (en) | Traffic flow prediction model based on gating time convolution network | |
| Chen et al. | A hierarchical consensus attention network for feature matching of remote sensing images | |
| CN118116194A (en) | Traffic state prediction method based on self-adaptive dynamic space-time diagram convolutional network | |
| Xu et al. | JCa2Co: A joint cascade convolution coding network based on fuzzy regional characteristics for infrared and visible image fusion | |
| Yang et al. | SA‐FlowNet: Event‐based self‐attention optical flow estimation with spiking‐analogue neural networks | |
| CN115953902B (en) | Traffic flow prediction method based on multi-view space-time diagram convolutional network | |
| CN117392846A (en) | Traffic flow prediction method for space-time self-adaptive graph learning fusion dynamic graph convolution | |
| CN117095460A (en) | Self-supervision group behavior recognition method and system based on long-short time relation predictive coding | |
| Lu et al. | Monocular semantic occupancy grid mapping with convolutional variational auto-encoders | |
| CN115830707A (en) | Multi-view human behavior identification method based on hypergraph learning | |
| CN114120245B (en) | Crowd image analysis method, device and equipment based on deep neural network | |
| CN115937594A (en) | Remote sensing image classification method and device based on local and global feature fusion | |
| Gao et al. | Semantic segmentation of remote sensing images based on multiscale features and global information modeling |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |