[go: up one dir, main page]
More Web Proxy on the site http://driver.im/

CN113283581A - Multi-fusion graph network collaborative multi-channel attention model and application method thereof - Google Patents

Multi-fusion graph network collaborative multi-channel attention model and application method thereof Download PDF

Info

Publication number
CN113283581A
CN113283581A CN202110526469.9A CN202110526469A CN113283581A CN 113283581 A CN113283581 A CN 113283581A CN 202110526469 A CN202110526469 A CN 202110526469A CN 113283581 A CN113283581 A CN 113283581A
Authority
CN
China
Prior art keywords
layer
fusion
channel attention
graph
network
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.)
Granted
Application number
CN202110526469.9A
Other languages
Chinese (zh)
Other versions
CN113283581B (en
Inventor
蒋国平
程曼茹
宋玉蓉
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Nanjing University of Posts and Telecommunications
Original Assignee
Nanjing University of Posts and Telecommunications
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Nanjing University of Posts and Telecommunications filed Critical Nanjing University of Posts and Telecommunications
Priority to CN202110526469.9A priority Critical patent/CN113283581B/en
Publication of CN113283581A publication Critical patent/CN113283581A/en
Application granted granted Critical
Publication of CN113283581B publication Critical patent/CN113283581B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • 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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • 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
    • 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
    • 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
    • G08G1/0137—Measuring and analyzing of parameters relative to traffic conditions for specific applications
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
    • Y02T10/00—Road transport of goods or passengers
    • Y02T10/10—Internal combustion engine [ICE] based vehicles
    • Y02T10/40—Engine management systems

Landscapes

  • Physics & Mathematics (AREA)
  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • General Engineering & Computer Science (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Biomedical Technology (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Chemical & Material Sciences (AREA)
  • Analytical Chemistry (AREA)
  • Traffic Control Systems (AREA)

Abstract

The invention relates to a multi-fusion graph network cooperation multi-channel attention model and an application and application method thereof. The method comprises the steps of constructing a potential graph structure by using the existing data, constructing a multi-attribute fused GCN module, transmitting the graph structure containing a plurality of attributes into the GCN module, and fully mining graph structure information; the ECA model is used for traffic flow prediction for the first time, and the attention mechanism of the ECA model is applied, so that the nonlinear dynamic variability of modeling time is better.

Description

Multi-fusion graph network collaborative multi-channel attention model and application method thereof
Technical Field
The invention relates to the technical field of communication, in particular to a multi-fusion graph network cooperation multi-channel attention model and a traffic flow prediction method applying the multi-fusion graph network cooperation multi-channel attention model.
Background
Traffic systems are one of the important components of modern cities, providing support for the daily commute and holiday travel of hundreds of millions of people. With the continuous acceleration of urbanization and the rapid increase of population, traffic systems become increasingly complex, and traffic jam and traffic accidents are frequent. If the urban traffic condition can be predicted, scheme deployment is made in advance, traffic pressure is relieved, the travel time and cost of people can be further reduced while traffic accidents are avoided, and environmental pollution is reduced. The traffic flow is one of the main parameters reflecting the traffic state, and the traffic flow prediction is more accurate, thereby being beneficial to the development of an intelligent traffic system.
Along with the development of IPv6, wireless communication technology and sensing technology, intelligent transportation integrates short-range wireless communication technology, microelectronic sensing technology, embedded sensing network and other technologies, so that continuous collection of various large-scale traffic data becomes possible, such as traffic, vehicle speed, lane occupancy rate and the like, a large amount of abundant data such as traffic time with geographic information and the like are accumulated, and a good data basis is provided for traffic prediction.
Currently, a large number of scholars have made relevant prediction attempts. The early prediction model requires relatively stable data and linear change, has the obvious defects of poor applicability, poor real-time performance and the like, and is difficult to adapt to actual requirements. Although the traditional machine learning developed later can also model complex data, the traffic flow data is influenced by a plurality of variable attributes, the traffic flow data usually has the characteristics of nonlinearity and mutability, and is difficult to predict accurately, and the prediction effect of the method greatly depends on feature engineering, which is usually dependent on the prior knowledge of domain experts.
The problem of prediction of spatio-temporal related data has become increasingly important in the field of spatio-temporal data mining. And traffic flow prediction is a typical spatio-temporal data prediction problem. Many classical models such as ARIMA and SVM only consider time information, such as ZL201711352952X, which discloses a traffic state prediction method based on an improved SVM algorithm, which is based on time intervals, but this method faces huge challenges in complex spatial dependence problems.
The ZL2018112789581 discloses an LSTM _ CNN-based urban road network traffic state prediction method, the prediction method combines the CNN and the LSTM to respectively model spatial correlation and temporal correlation, the effect of the prediction method is verified in urban pedestrian volume prediction problems, but the traffic prediction problems belong to typical graph structure correlation prediction, and the models do not further capture more graph structure information.
ZL2020115437933 discloses a traffic flow prediction method for training a convolutional neural network by using a dynamic space-time diagram, the method constructs the convolutional neural network and trains the convolutional neural network by using the dynamic space-time diagram formed by traffic flow data, and the invention has higher accuracy of predicting the traffic flow and can better capture the structural information of the dynamic space-time diagram by adopting diagram volume and attention.
ZL 2020103153127 a high-speed traffic flow prediction method based on multimode fusion and an image attention machine mechanism, time and space characteristics in a traffic flow model are comprehensively considered, multiple factors are fused through a coding mode, the speed prediction model is built through the methods of the image attention machine mechanism, hole convolution and the like, a model is built by using measured data of a certain highway section and is verified, and the prediction result and the measured result are compared to show that the vehicle speed prediction method provided by the research has a good effect.
In recent years, researchers have been trying to model spatial correlations in spatio-temporal network data using graph convolution methods. The DCRNN introduces a graph convolution network into the prediction of spatio-temporal data, and uses a diffusion convolution network to describe the diffusion process of information in a spatial network. STGCN models temporal correlation with CNN, and GCN models spatial correlation. The ASTGCN uses two layers of attention to capture spatially dependent and temporally dependent dynamics. Graph WaveNet designs an adaptive matrix to account for the impact variations between a node and its neighbors, models temporal correlation using extended causal convolution, and increases the acceptance field in exponential order. STG2Seq differs from the above method in that it does not use two different components to capture spatial and temporal correlations separately, but only simulates spatio-temporal correlations by using two attention-driven gated residual GCN blocks.
However, the above method only considers the distance or only establishes a simple connecting edge when constructing the space diagram structure, and does not consider the influence of other correlations such as road intersection structures, public transportation facilities, regional function attributes and the like on the diagram structure.
Disclosure of Invention
In order to solve the problems, the invention provides a multi-fusion graph network cooperation multi-channel attention model and a method for predicting traffic flow by using the model, wherein a space-time convolution deep learning model with a plurality of attribute characteristics is comprehensively considered, a multi-fusion graph structure is constructed by using the existing data information, space-time correlation and nonlinear time dynamic variability are fully mined by using a graph convolution neural network, a convolution neural network and an ECA model, and the ECA module avoids dimension reduction and effectively captures cross-channel interaction information.
In order to achieve the purpose, the invention is realized by the following technical scheme:
the invention relates to a multi-fusion graph network cooperation multi-channel attention model, which comprises a multi-channel attention layer (ECA), a multi-fusion graph volume layer (MF-GCN), a convolution neural network layer (CNN), a residual error network (residual) and a RELU layer, wherein the multi-channel attention layer ECA, the multi-fusion graph volume layer (MF-GCN) and the convolution neural network layer (CNN) are stacked to construct a multi-fusion time space BLOCK (MF-ST-BLOCK), the multi-fusion time space BLOCK (MF-ST-BLOCK) is combined with the residual error network (residual) and the RELU layer to obtain a multi-fusion graph network cooperation multi-channel attention BLOCK (MFG-ECA-BLOCK), a plurality of multi-fusion graph network cooperation multi-channel attention BLOCKs (MFG-ECA-BLOCK) are stacked, the output of the last multi-fusion graph network cooperation multi-channel attention BLOCK (MFG-ECA-BLOCK) is convoluted through a neural network (CNN) layer, and obtaining a multi-fusion graph network cooperative multi-channel attention model.
A construction method of a multi-fusion graph network collaborative multi-channel attention model comprises the following steps:
step 1: constructing a multi-fusion graph by using the existing data and combining a graph convolution neural network (GCN) to construct a multi-fusion graph convolution layer (MF-GCN);
step 2: stacking a multi-channel attention layer (ECA), a multi-fusion graph convolution layer (MF-GCN) and a convolutional neural network layer (CNN) to construct a multi-fusion time space BLOCK (MF-ST-BLOCK), and combining a residual error network (residual) to finally obtain a multi-fusion graph network cooperation multi-channel attention BLOCK (MFG-ECA-BLOCK) through a RELU layer;
and step 3: and (3) stacking a plurality of multi-fusion graph network cooperation multi-channel attention BLOCKs (MFG-ECA-BLOCKs) constructed in the step (2), and enabling the output of the last multi-fusion graph network cooperation multi-channel attention module to pass through a layer of convolutional neural network layer (CNN) to obtain the final sequence output.
The invention is further improved in that: the multi-fusion graph network cooperation multi-channel attention model can be used for traffic flow prediction, and the prediction method comprises the following steps:
step (1), inputting traffic data into a multi-channel attention layer (ECA), and performing scoring calculation through the multi-channel attention layer (ECA) to obtain attention output;
step (2), calculating the acquired traffic flow, average vehicle speed, average road occupation rate and node space information to obtain a related graph structure;
step (3), feeding the output of the step (1) and the adjacent matrix data in the step (2) into a multi-fusion graph convolution layer (MF-GCN) together for convolution operation to obtain space-time correlation;
step (4), the output of the multi-fusion map convolution layer (MF-GCN) is subjected to a layer of time convolution neural network to obtain the space-time correlation characteristic and the nonlinear dynamic variability of traffic data which are extracted again;
step (5), combining a residual error network (residual) to prevent information loss, and obtaining a multi-fusion graph network collaborative multi-channel attention BLOCK (MFG-ECA-BLOCK) through a RELU layer;
and (6) stacking a plurality of multi-fusion graph network cooperative multi-channel attention BLOCKs (MFG-ECA-BLOCKs), extracting more information, and passing the output of the last multi-fusion graph network cooperative multi-channel attention BLOCK (MFG-ECA-BLOCK) through a layer of Convolutional Neural Network (CNN) layer to obtain the final sequence output.
The invention has the beneficial effects that:
(1) the method comprises the steps of constructing a potential graph structure by using the existing data, constructing a multi-attribute fused GCN module (MF-GCN), transmitting a graph structure containing a plurality of attributes into the GCN module, and fully mining graph structure information;
(2) according to the invention, the ECA model is used for traffic flow prediction for the first time, the time dimension is used as a channel, and the attention mechanism of the ECA model is applied, so that the nonlinear dynamic variability of modeling time is better;
(3) according to the invention, the ECA module, the MF-GCN module and the time dimension CNN module are stacked and fused to fully mine the time-space related information.
A multi-fusion graph network in cooperation with a multi-channel attention model (MFG-ECA) is proposed herein: taking the time dimension as a channel, and capturing the space-time dynamics of traffic flow data by using an efficient channel attention model (ECA); the existing data is used for constructing a multi-fusion graph structure, spatial dependence is extracted through a graph convolution neural network, and then the convolution neural network is stacked to fully excavate space-time correlation and nonlinear dynamic change characteristics.
Drawings
FIG. 1 is a spatiotemporal correlation of traffic flow data of the present invention.
Fig. 2 is a view showing a structure of a traffic space according to the present invention.
Fig. 3 is a traffic flow data line graph in which the horizontal axis represents time steps and the vertical axis represents traffic flow values according to the present invention.
Fig. 4 shows the time correlation of traffic flow according to the present invention, and the thickness of the arrows indicates the difference in the degree of correlation.
FIG. 5 is a model structure diagram of MF-GCN according to the present invention.
FIG. 6 is a diagram of the MFG-ECA model architecture of the present invention.
FIG. 7 is a graph comparing experimental results of the spatiotemporal methods of the present invention, comparing the results on the two data of pems04 and pems 08.
FIG. 8 is a comparison of the MFG-ECA model predicted results of the present invention with real data, compared to results on both pems04 and pems 08.
Detailed Description
In the following description, for purposes of explanation, numerous implementation details are set forth in order to provide a thorough understanding of the embodiments of the invention. It should be understood, however, that these implementation details are not to be interpreted as limiting the invention. That is, in some embodiments of the invention, such implementation details are not necessary.
As can be seen from observing fig. 1, at the same time step, due to the accessibility of the physical distance, the traffic flow of one node is easily affected by its neighboring nodes; the size of the traffic flow of the same node has strong correlation with the size of the traffic flow of the previous time step; at different time steps, traffic flows between different nodes mutually influence to form complex space-time correlation; in addition, the public transport means setting, the regional function attribute and the flow among the nodes with similar road fork structures have certain correlation, and the correlation is reflected by traffic data such as average speed, average road occupation rate, traffic flow and the like; secondly, the occurrence of some major events or traffic accidents can cause traffic data to have great fluctuation, the occurrence of the major events or the traffic accidents has strong uncertainty, and the influence of a plurality of factors forms extremely strong traffic data with nonlinear dynamic change. As can be seen, the prediction of traffic flow is important and difficult.
As shown in fig. 2, the spatial distance between the point a and the point B is close, so that there is a possibility that the traffic flow between the point a and the point B may affect each other; although the spatial distance between the point a and the point C is relatively far, the traffic data between the point a and the point C may have a certain degree of similarity due to certain similarity in terms of functional structures, road junction structures, and public transportation facility deployments.
In the aspect of time correlation, as shown in fig. 4, traffic data at a point a at a time t is influenced by a flow rate at the time t-1, and may have a certain correlation with traffic data at the same time before a day or a week; traffic data has strong spatio-temporal correlation, and the lapse of time affects the change of space and then affects the traffic data of our time.
Meanwhile, as can be observed from fig. 3, the traffic data is not constant, and has extremely strong nonlinear dynamic variability, we propose a multi-fusion graph network and multi-channel attention model (MFG-ECA) to comprehensively consider the traffic data attributes for traffic flow prediction.
Fig. 5 shows a multi-fusion graph network collaborative multi-channel attention model (MFG-ECA) proposed herein for modeling spatiotemporal correlations of predicted targets, dynamic variability of spatiotemporal data, and other correlations such as road intersection structures, public transportation facility deployment, regional functional attributes, which are established on the characteristics of the above correlations, the collaborative energy models proposed in the present invention respectively include: the ECA module is used for capturing complex traffic data dynamics, for example, as shown in fig. 6, the MF-GCN module comprehensively considers relevant characteristics of road network distance, road intersection structure, public transport facility arrangement, regional function attributes and the like, captures complex potential graph structure information, and finally learns time series information through the TimeCNN module, each module takes its own role, each module is dedicated to the respective function to learn complex traffic data, and the three modules are stacked and fused to learn complex spatiotemporal relevant information together.
More specifically, the attention model of the invention comprises a multi-channel attention layer (ECA), a multi-fusion map convolutional layer (MF-GCN), a convolutional neural network layer (CNN), a residual error network (residual) and a RELU layer, wherein the multi-channel attention layer ECA, the multi-fusion map convolutional layer (MF-GCN) and the convolutional neural network layer CNN are stacked to construct a multi-fusion time space BLOCK (MF-ST-BLOCK), the multi-fusion time space BLOCK (MF-ST-BLOCK) is combined with the residual error network (residual) and the RELU layer to obtain a multi-fusion map network cooperation multi-channel attention BLOCK (MFG-ECA-BLOCK), a plurality of multi-fusion map network cooperation multi-channel attention BLOCKs (MFG-ECA-BLOCK) are stacked, the output of the last multi-fusion map network cooperation multi-channel attention BLOCK (MFG-ECA-BLOCK) passes through the neural network (CNN) layer, and obtaining a multi-fusion graph network cooperative multi-channel attention model.
The construction method of the model comprises the following steps:
step 1: constructing a multi-fusion graph by using the existing data and combining a graph convolution neural network (GCN) to construct a multi-fusion graph convolution layer (MF-GCN);
step 2: stacking a multi-channel attention layer (ECA), a multi-fusion graph convolution layer (MF-GCN) and a convolutional neural network layer (CNN) to construct a multi-fusion time space BLOCK (MF-ST-BLOCK), and combining a residual error network (residual) to finally obtain a multi-fusion graph network cooperation multi-channel attention BLOCK (MFG-ECA-BLOCK) through a RELU layer;
and step 3: and (3) stacking a plurality of multi-fusion graph network cooperation multi-channel attention BLOCKs (MFG-ECA-BLOCKs) constructed in the step (2), and enabling the output of the last multi-fusion graph network cooperation multi-channel attention module to pass through a layer of convolutional neural network layer (CNN) to obtain the final sequence output.
The multi-fusion graph network cooperation multi-channel attention model is applied to space-time data prediction, particularly traffic flow prediction, and the specific traffic flow prediction method applying the multi-fusion graph network cooperation multi-channel attention model comprises the following steps:
step (1), inputting traffic data into a multi-channel attention layer (ECA), and performing scoring calculation through the multi-channel attention layer (ECA) to obtain attention output, specifically:
feeding historical traffic flow data X into an ECA module to obtain an attention score, and setting X as the same as RF×N×TAs input:
Figure BDA0003066066790000083
wherein
Figure BDA0003066066790000081
Is the global average pooling of channels, σ is the activation function, let z be g (x),
Figure BDA0003066066790000084
ReLU represents a linear modification unit activation function, and given an aggregate feature z, channel attention can be learned by:
ω＝σ(Wz) (2)
W∈RT×Tis a matrix; in order to make learning more efficient, to make all channels share learning parameters,
Figure BDA0003066066790000082
and (3) realizing information interaction between channels by using a 1-dimensional convolution method with a convolution kernel size of k:
ω＝σ(C1Dk(z)) (3)
C1D represents a one-dimensional convolution, multiplying the attention matrix by x to obtain the output:
xout＝x*ω (4)
then x is putout∈RF×N×TAs output, into the MF-GCN layer, and then captures spatio-temporal related information.
Step (2), the obtained traffic flow, average speed, average road occupation rate and node space information are calculated to obtain a related graph structure,
Figure BDA0003066066790000091
Figure BDA0003066066790000092
Figure BDA0003066066790000093
where μ 1, μ 2, μ 3, ε 1, ε 2, ε 3 are hyper-parameters, pij＝mean(|ci-cj|)，bij＝mean(|si-sj|，hij＝mean(|oi-ojL, by fij，sij，oijFlow connecting edges between the vertex i and the vertex j, average speed connecting edges and average lane occupation rate connecting edges are respectively formed to construct a matrix C, S, O;
distance matrix:
Figure BDA0003066066790000094
Qijand (4) connecting the distances between the vertex i and the vertex j to obtain a distance matrix Q.
And (3) feeding the output of the step (1) and the adjacent matrix data in the step (2) into a multi-fusion graph convolution layer (MF-GCN) together for convolution operation to obtain space-time correlation.
The above four graph structures are fused together to form a new graph:
A＝wdQ+wcC+wsS+woO (9)
wherein Wd，Wc，Ws，WoIs a super ginseng and wd+wc+ws+woConstructing a Laplace matrix from an adjacent matrix A of a multi-fused graph structure as 1, combining the output of the step 1 and inputting the combined output into a GCN,
the graph convolution uses a linear operator defined diagonalized in the fourier domain to equivalently replace the classical convolution operator, thereby implementing the convolution operation:
Figure BDA0003066066790000101
where L is the Laplace matrix of the graph, L ═ D-A, normalized
Figure BDA0003066066790000102
Wherein A is an adjacency matrix, INIs an identity matrix, D ∈ RN×NIs a degree matrix, a diagonal matrix composed of degrees of nodes, Dij＝∑Aij。L＝UΛUTIs the result of eigen decomposition of the laplacian matrix, where Λ ═ diag ([ λ [ # ] ]0，...，λN-1])∈RN×NIs a diagonal matrix composed of eigenvalues of L; gθIs a graph-convolution filter that is,
when a large graph is convolved, it is expensive to directly perform eigenvalue decomposition on the laplacian matrix, where the solution is solved using chebyshev polynomial approximation:
Figure BDA0003066066790000103
θK∈RKis a coefficient of the chebyshev polynomial,
Figure BDA0003066066790000104
is the maximum eigenvalue, T, of the Laplace matrixK(x)＝2xTK-1(x)-TK-2(x) And (4) performing approximate expansion and solving by using a Chebyshev polynomial, namely extracting the information of 0-K-1 order neighbors taking the node as the center in the graph by using a convolution kernel.
Constructing an adjacent matrix A by using a multi-fusion graph structure, constructing a Laplace matrix L, assuming that K is 2, extracting 0-1 order neighbor information, and approximating the maximum eigenvalue lambda of the Laplace operator matrix in the next stepmax，λmax2, equation (11) can be expressed as:
Figure BDA0003066066790000105
θ0，θ1is two shared parameters, θ0，θ1May be replaced by θ, θ ═ θ0＝-θ1To avoid the problems of unstable value, disappearance of gradient, explosion of gradient, etc
Figure BDA0003066066790000106
Then equation (12) can be simplified as:
Figure BDA0003066066790000111
stacking multiple layers:
Figure BDA0003066066790000112
Figure BDA0003066066790000113
is an input of the r-th layer, Cr-1Is the number of channels in the r-th layer, r e 1,.., l being the number of layers of convolution,
Figure BDA0003066066790000114
the convolution kernel parameters are obtained, and each node contains the information of 0-k-1 neighbor nodes through layer-by-layer aggregation.
Step (4), the output of the multi-fusion map convolution layer (MF-GCN) is subjected to a layer of time convolution neural network to obtain the space-time correlation characteristic and the nonlinear dynamic variability of traffic data which are extracted again;
and (5) combining a residual error network (residual) to prevent information loss, and obtaining a multi-fusion graph network cooperation multi-channel attention BLOCK (MFG-ECA-BLOCK) through a RELU layer.
After the convolution operation of the MF-GCN block, each node on the graph captures the relevant information of the neighbor nodes, and further stacking the time dimension standard convolution layers to update the signals of the nodes by combining the information on the adjacent time steps:
Figure BDA0003066066790000115
where Φ is the convolution kernel parameter in the time dimension, representing the convolution operation, ReLU is the linear modification unit activation function.
And (6) stacking a plurality of multi-fusion graph network cooperative multi-channel attention BLOCKs (MFG-ECA-BLOCKs), extracting more information, and passing the output of the last multi-fusion graph network cooperative multi-channel attention BLOCK (MFG-ECA-BLOCK) through a layer of Convolutional Neural Network (CNN) layer to obtain the final sequence output.
To verify the accuracy of the model, this embodiment refers to two sets named PEMS04 and PEMS08, these traffic data are collected every 5 minutes as data of one time slice, which means that each hour will contain 12 time slices of information, the data set contains three-dimensional characteristics of traffic flow, average vehicle speed, average road occupation rate and geographical location information of the relevant detectors, and the traffic data of the next hour is predicted using the traffic data of the past 1 hour, the detailed information is shown in table 1:
table 1 data set description
Figure BDA0003066066790000121
Experimental comparison of the MFG-ECA model proposed in this example with the above reference model, table 2 shows experimental results comparison of different methods on two datasets Pems04 and Pems 08. Compared with other 7 reference methods, the MFG-ECA model provided by the invention has good performance on evaluation indexes of the two data sets.
TABLE 2 comparison of traffic flow prediction Performance across different approaches
Figure BDA0003066066790000122
VAR, SVM, LSTM only consider temporal correlation and do not join one of the important dependencies in spatio-temporal networks: spatial dependence. The DCRNN, the STGCN, the STG2Seq, the Graph WaveNet and the MFG-ECA model provided by the invention all add the important dependence of Graph structure information, so that the performance of the model is better than that of a time series prediction model only considering the time performance from the experimental result.
However, the model of the invention considers the dependence of common spatial distance, and simultaneously adds other attribute information, a new graph structure is constructed by fusion, and meanwhile, an attention module ECA is introduced to capture the influence of different time steps and the inherent dynamic variability of traffic data by taking the time dimension as a channel, and also considers the space-time correlation, and the performance of the model of the invention is better as shown in FIG. 7. Wherein STG2Seq also introduces an attention module while considering spatio-temporal correlations to capture the impact of the most influential historical time-steps on the predicted demand and the dynamics inherent to these relationships. But it does not consider the potential graph structure information and ignores other potential important information that affects traffic flow. Meanwhile, the attention ECA module is more effective and reasonable in capturing the dynamic property, so that the model can concentrate on learning of important information.
In order to better analyze the performance of the MFG-ECA model of the present invention, a detection point is randomly extracted and a visual comparison operation is performed on the predicted value of the detection point on the test set, and as can be seen from the visual result of fig. 7, the MFG-ECA model can better sense the dynamic variability of the data time dimension and can better detect the nonlinear fluctuation. The invention integrates ECA module while considering potential graph structure, integrates more graph structure information to traffic flow prediction, gives full play to the importance of attention mechanism in traffic flow prediction, uses the output of ECA module with time attention mechanism as input data, integrates graph information of multiple graph network structures, convolutes through GCN module, captures time-space correlation attribute, convolutes through TimeCNN module, learns time-space data again, extracts effective characteristic information, combines modules, superposes and fuses, and enables the model to focus on important information learning and sense dynamic variability of data and capture time-space correlation information together.
The above description is only an embodiment of the present invention, and is not intended to limit the present invention. Various modifications and alterations to this invention will become apparent to those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the scope of the claims of the present invention.

Claims (8)

1. The multi-fusion graph network cooperation multi-channel attention model is characterized in that: comprises a multi-channel attention layer, a multi-fusion graph convolution layer, a convolution neural network layer, a residual error network and a RELU layer,
stacking a multi-channel attention layer, a multi-fusion graph convolution layer and a convolution neural network layer to construct a multi-fusion time empty block;
combining the multi-fusion space-time block with the residual error network and a RELU layer to obtain a multi-fusion graph network collaborative multi-channel attention block;
and stacking a plurality of multi-fusion graph network cooperative multi-channel attention blocks, and passing the output of the last multi-fusion graph network cooperative multi-channel attention block through a convolutional neural network layer to obtain a multi-fusion graph network cooperative multi-channel attention model.
2. The multi-fusion graph network-collaborative multi-channel attention model of claim 1, wherein: the multi-fusion graph convolution layer is constructed by constructing a multi-fusion graph by using the existing data and combining a graph convolution neural network.
3. The method for constructing the multi-fusion graph network-collaborative multi-channel attention model according to claim 2, wherein: the construction method comprises the following steps:
step 1: constructing a multi-fusion graph by using the existing data and combining a graph convolution neural network to construct a multi-fusion graph convolution layer;
step 2: stacking a multi-channel attention layer, a multi-fusion graph convolution layer and a convolution neural network layer to construct a multi-fusion time empty block, combining a residual error network, and finally obtaining a multi-fusion graph network collaborative multi-channel attention block through a RELU layer;
and step 3: and (3) stacking a plurality of multi-fusion graph network cooperation multi-channel attention blocks constructed in the step (2), and enabling the output of the last multi-fusion network cooperation multi-channel attention module to pass through a layer of convolution neural network layer to obtain the final sequence output.
4. Use of the multi-fusion graph network in cooperation with a multi-channel attention model according to any one of claims 1-3 in traffic flow prediction.
5. The traffic flow prediction method applying the multi-fusion graph network in cooperation with the multi-channel attention model according to claim 4, characterized in that: the prediction method comprises the following steps:
step (1), inputting traffic data into a multi-channel attention layer, and performing scoring calculation through the multi-channel attention layer to obtain attention output;
step (2), calculating the acquired traffic flow, average vehicle speed, average road occupation rate and node space information to obtain a related graph structure;
step (3), feeding the output of the step (1) and the adjacent matrix data in the step (2) into a multi-fusion graph convolution layer together for convolution operation to obtain space-time correlation;
step (4), the output of the convolution layer of the multiple fusion graphs is subjected to a layer of time convolution neural network to obtain the space-time correlation characteristic and the nonlinear dynamic variability of the traffic data which are extracted again;
step (5), combining a residual error network to prevent information loss, and obtaining a multi-fusion graph network collaborative multi-channel attention block through a RELU layer;
and (6) stacking a plurality of multi-fusion graph network cooperative multi-channel attention blocks, extracting more information, and passing the output of the last multi-fusion graph network cooperative multi-channel attention block through a convolutional neural network layer to obtain final sequence output.
6. The multi-fusion graph network-collaborative multi-channel attention model of claim 5, wherein: in the step (1):
feeding historical traffic flow data X into a multi-channel attention layer to obtain attention scores, and obtaining the attention scores
Figure FDA0003066066780000021
As input:
Figure FDA0003066066780000022
wherein
Figure FDA0003066066780000023
Is the global average pooling of channels, sigma is the activation function, and
Figure FDA0003066066780000024
Figure FDA0003066066780000025
ReLU represents a linear modification unit activation function, and given an aggregate feature z, channel attention can be learned by:
ω＝σ(Wz) (2)
W∈RT×Tis a matrix; in order to make learning more efficient, to make all channels share learning parameters,
Figure FDA0003066066780000031
and (3) realizing information interaction between channels by using a 1-dimensional convolution method with a convolution kernel size of k:
ω＝σ(C1Dk(z)) (3)
C1D represents a one-dimensional convolution of the attention matrix with
Figure FDA0003066066780000032
Multiplying to obtain an output:
Figure FDA0003066066780000033
then will be
Figure FDA0003066066780000034
As output, into the multi-fused graph convolution layer, and then captures spatiotemporal correlation information.
7. The multi-fusion graph network-collaborative multi-channel attention model of claim 5, wherein: in the step (2):
Figure FDA0003066066780000035
Figure FDA0003066066780000036
Figure FDA0003066066780000037
where μ 1, μ 2, μ 3, ε 1, ε 2, ε 3 are hyper-parameters, pij＝mean(|ci-cj|)，bij＝mean(|si-sj|，hij＝mean(|oi-ojL, by fij，sij，oijFlow connecting edges between the vertex i and the vertex j, average speed connecting edges and average lane occupation rate connecting edges are respectively formed to construct a matrix C, S, O;
distance matrix:
Figure FDA0003066066780000041
Qijand (4) connecting the distances between the vertex i and the vertex j to obtain a distance matrix Q.
8. The multi-fusion graph network-collaborative multi-channel attention model of claim 5, wherein: the step (3) is specifically as follows: fusing the four graph structures of the step (1) and the step (2) together to form a new graph:
A＝wdQ+wcC+wsS+wOO (9)
wherein Wd，Wc，Ws，WoIs a super ginseng and wd+wc+ws+woConstructing a Laplace matrix from an adjacent matrix A of a multi-fusion graph structure, and inputting the adjacent matrix A into a graph convolution neural network by combining the output of the step 1 to realize convolution operation:
gθ×Gx＝gθ(L)x＝gθ(UΛUT)x
＝U(gθ(Λ))UTx (10)
where L is the Laplace matrix of the graph, L ═ D-A, normalized
Figure FDA0003066066780000042
Wherein A is an adjacency matrix, INIs an identity matrix, D ∈ RN×NIs a degree matrix, a diagonal matrix composed of degrees of nodes, Dij＝∑Aij。L＝UΛUTIs the result of eigen decomposition of the laplacian matrix, where Λ ═ diag ([ λ [ # ] ]0，...，λN-1])∈RN×NIs a diagonal matrix composed of eigenvalues of L; gθIs a graph convolution filter, solved using chebyshev polynomial approximation expansion:
Figure FDA0003066066780000043
θK∈RKis a coefficient of the chebyshev polynomial,
Figure FDA0003066066780000044
λmaxis the maximum eigenvalue, T, of the Laplace matrixK(x)＝2xTK-1(x)-TK-2(x) Constructing an adjacent matrix A by using a multi-fusion graph structure, constructing a Laplace matrix L, assuming that K is 2, extracting 0-1 order neighbor information, and approximating the maximum eigenvalue lambda of the Laplace operator matrix in the next stepmax，λmax2, equation (11) can be expressed as:
Figure FDA0003066066780000051
θ0，θ1is two shared parameters, θ0，θ1May be replaced by θ, θ ═ θ0＝-θ1Let us order
Figure FDA0003066066780000052
Figure FDA0003066066780000053
Then equation (12) can be simplified as:
Figure FDA0003066066780000054
stacking multiple layers:
Figure FDA0003066066780000055
Figure FDA0003066066780000056
is an input of the r-th layer, Cr-1Is the number of channels in the r-th layer, r is the { 1., l }, l is the number of convolution layers,
Figure FDA0003066066780000057
the convolution kernel parameters are obtained, and each node contains the information of 0-k-1 neighbor nodes through layer-by-layer aggregation.
CN202110526469.9A 2021-05-14 2021-05-14 Multi-fusion graph network collaborative multi-channel attention model and application method thereof Active CN113283581B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110526469.9A CN113283581B (en) 2021-05-14 2021-05-14 Multi-fusion graph network collaborative multi-channel attention model and application method thereof

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110526469.9A CN113283581B (en) 2021-05-14 2021-05-14 Multi-fusion graph network collaborative multi-channel attention model and application method thereof

Publications (2)

Publication Number Publication Date
CN113283581A true CN113283581A (en) 2021-08-20
CN113283581B CN113283581B (en) 2022-08-09

Family

ID=77279014

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110526469.9A Active CN113283581B (en) 2021-05-14 2021-05-14 Multi-fusion graph network collaborative multi-channel attention model and application method thereof

Country Status (1)

Country Link
CN (1) CN113283581B (en)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114422211A (en) * 2021-12-30 2022-04-29 中国人民解放军战略支援部队信息工程大学 HTTP malicious traffic detection method and device based on graph attention network
CN114971007A (en) * 2022-05-20 2022-08-30 山东华科信息技术有限公司 Photovoltaic power generation prediction method and system based on multi-scale graph convolutional neural network
CN115482656A (en) * 2022-05-23 2022-12-16 汕头大学 Method for predicting traffic flow by using space dynamic graph convolution network
CN116597657A (en) * 2023-07-17 2023-08-15 四川省商投信息技术有限责任公司 Urban traffic prediction method, device and medium based on artificial intelligence

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111914726A (en) * 2020-07-28 2020-11-10 联芯智能(南京)科技有限公司 Pedestrian detection method based on multi-channel self-adaptive attention mechanism
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
US20210081717A1 (en) * 2018-05-18 2021-03-18 Benevolentai Technology Limited Graph neutral networks with attention

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20210081717A1 (en) * 2018-05-18 2021-03-18 Benevolentai Technology Limited Graph neutral networks with attention
CN111914726A (en) * 2020-07-28 2020-11-10 联芯智能(南京)科技有限公司 Pedestrian detection method based on multi-channel self-adaptive attention mechanism
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

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114422211A (en) * 2021-12-30 2022-04-29 中国人民解放军战略支援部队信息工程大学 HTTP malicious traffic detection method and device based on graph attention network
CN114422211B (en) * 2021-12-30 2023-07-18 中国人民解放军战略支援部队信息工程大学 HTTP malicious traffic detection method and device based on graph attention network
CN114971007A (en) * 2022-05-20 2022-08-30 山东华科信息技术有限公司 Photovoltaic power generation prediction method and system based on multi-scale graph convolutional neural network
CN115482656A (en) * 2022-05-23 2022-12-16 汕头大学 Method for predicting traffic flow by using space dynamic graph convolution network
CN115482656B (en) * 2022-05-23 2023-09-26 汕头大学 Traffic flow prediction method by using space dynamic graph convolutional network
CN116597657A (en) * 2023-07-17 2023-08-15 四川省商投信息技术有限责任公司 Urban traffic prediction method, device and medium based on artificial intelligence

Also Published As

Publication number Publication date
CN113283581B (en) 2022-08-09

Similar Documents

Publication Publication Date Title
CN113283581B (en) Multi-fusion graph network collaborative multi-channel attention model and application method thereof
CN109887282B (en) Road network traffic flow prediction method based on hierarchical timing diagram convolutional network
CN109697852B (en) Urban road congestion degree prediction method based on time sequence traffic events
US11270579B2 (en) Transportation network speed foreeasting method using deep capsule networks with nested LSTM models
Yan et al. Spatial-temporal chebyshev graph neural network for traffic flow prediction in iot-based its
CN113724504B (en) Urban area traffic prediction system and method oriented to vehicle track big data
CN111860951A (en) Rail transit passenger flow prediction method based on dynamic hypergraph convolutional network
CN111223301B (en) Traffic flow prediction method based on graph attention convolution network
CN110322064B (en) Urban trip demand prediction method
Duan et al. FDSA-STG: Fully dynamic self-attention spatio-temporal graph networks for intelligent traffic flow prediction
CN113313947A (en) Road condition evaluation method of short-term traffic prediction graph convolution network
CN115204478A (en) Public traffic flow prediction method combining urban interest points and space-time causal relationship
CN115578851A (en) Traffic prediction method based on MGCN
Dai et al. Spatio-temporal deep learning framework for traffic speed forecasting in IoT
Keyarsalan et al. Designing an intelligent ontological system for traffic light control in isolated intersections
Deng et al. The pulse of urban transport: Exploring the co-evolving pattern for spatio-temporal forecasting
CN115376317A (en) Traffic flow prediction method based on dynamic graph convolution and time sequence convolution network
CN115565369A (en) Hypergraph-based time-space hypergraph convolution traffic flow prediction method and system
CN115545758A (en) Method and system for self-adaptive incremental site selection of urban service facilities
CN114021809A (en) Unmanned driving is with picture level convolution neural network model of traffic flow prediction
CN115862324A (en) Space-time synchronization graph convolution neural network for intelligent traffic and traffic prediction method
CN113159371B (en) Unknown target feature modeling and demand prediction method based on cross-modal data fusion
Yang et al. MGSTCN: a multi-graph spatio-temporal convolutional network for metro passenger flow prediction
Fang et al. Traffic speed prediction based on LSTM-graph attention network (L-GAT)
CN118262516A (en) Vehicle traffic prediction method and system based on multiple attention neural networks

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