CN115019504A - Short-term traffic flow prediction method based on new deep space time self-adaptive fusion graph network - Google Patents

Short-term traffic flow prediction method based on new deep space time self-adaptive fusion graph network Download PDF

Info

Publication number
CN115019504A
CN115019504A CN202210532872.7A CN202210532872A CN115019504A CN 115019504 A CN115019504 A CN 115019504A CN 202210532872 A CN202210532872 A CN 202210532872A CN 115019504 A CN115019504 A CN 115019504A
Authority
CN
China
Prior art keywords
time
space
fusion
adaptive
temporal
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.)
Pending
Application number
CN202210532872.7A
Other languages
Chinese (zh)
Inventor
周腾
杨舒敏
李桦樱
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.)
Shantou University
Original Assignee
Shantou University
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 Shantou University filed Critical Shantou University
Priority to CN202210532872.7A priority Critical patent/CN115019504A/en
Publication of CN115019504A publication Critical patent/CN115019504A/en
Pending legal-status Critical Current

Links

Images

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
    • 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
    • G08G1/0133—Traffic data processing for classifying traffic situation
    • 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

Abstract

The invention discloses a short-term traffic flow prediction method based on a new deep space time self-adaptive fusion graph network. And then, carrying out parallel fusion on the graphs constructed from different angles according to the traffic flow data to obtain a new space-time fusion matrix. And simultaneously carrying out graph diffusion convolution on the adaptive matrix and the space-time fusion matrix to capture hidden space-time dependency, and carrying out training on a deeper network model by using the finally captured characteristics to obtain a prediction result. The invention carries out test experiments on a plurality of traffic flow data sets, and the experimental result shows that the network performance is superior to the prior most advanced method.

Description

Short-term traffic flow prediction method based on new deep space time self-adaptive fusion graph network
Technical Field
The invention relates to the technical field of traffic flow prediction, in particular to a short-term traffic flow prediction method based on a new deep space time self-adaptive fusion graph network.
Background
An Intelligent Transport System (ITS) was proposed in the 90 s of the 20 th century, and is an optimal approach to solve urban traffic congestion and improve driving safety. The traffic flow prediction is the core content of an intelligent traffic system and is also an important basis for traffic information service, traffic control and guidance. The traffic flow prediction means that the collected dynamic traffic data is used for traffic flow prediction of a future road network. The urbanization process is accelerated, the living standard of people is improved, and the limited road network capacity in the city is difficult to match with vehicles with increasingly sharp roads nowadays. A reasonable, real-time and accurate traffic flow prediction system has great significance for relieving road network traffic pressure, planning travel of vehicles and ensuring efficient and safe operation of a traffic system.
Traffic flow prediction methods are mainly classified into two categories: knowledge-driven classes and data-driven classes. In early development, the knowledge-driven class mainly applies queuing theory and simulation behavior; in the later stage, the data-class driving method is rapidly developed under the promotion of hardware equipment, and comprises the traditional statistics and machine learning methods, such as: autoregressive Moving Average model (ARIMA), Vector Autoregressive model (VAR), support Vector regression model (SVR). The basic idea of such methods is to put a traffic data stream over a period of time into a regression model to predict the traffic information conditions for the next period of time. It is limited by the assumption of stationarity of time series, ignoring spatio-temporal correlations.
The traditional statistical method and the machine learning method need to establish prediction on the assumption that the time sequence is stationary. The predicted data precision is poor, meanwhile, the static prediction cannot reflect the dynamic change characteristics of the traffic flow, the uncertainty and nonlinearity of the traffic flow process are difficult to reflect, and the influence of an emergency on the traffic condition cannot be solved.
The combination of Graph Convolutional Neural Network (GCN) and Recurrent Neural Network (RNN) with traffic flow prediction has been widely studied. However, it is often difficult to accurately capture the spatial dependency of hidden road nodes and the temporal dependency of road nodes.
Most of the applications of the graph convolutional neural network and the cyclic neural network in traffic flow prediction use a matrix that has been defined in advance and a static graph structure, which is constructed based on the distances between sensors in road nodes. However, traffic flow data is highly nonlinear and dynamically changes from moment to moment due to various factors, and thus, the correlation between road networks dynamically changes with time, such as: at some time period, 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 among the nodes, and plays a key role in the learning expression of the space nodes. Furthermore, many methods choose to construct a graph that constructs temporal correlations based on road data to capture spatial node features, most methods use euclidean distances to compute the temporal correlations between road nodes. However, the similarity between data with similar shapes but unparallel time series cannot be well measured, so that a relatively accurate time correlation graph cannot be obtained, and the captured time dependency between nodes is greatly inconsistent with the actual time dependency.
Disclosure of Invention
The technical problem to be solved by the embodiment of the invention is to provide a short-term traffic flow prediction method based on a new deep space-time adaptive fusion graph network. The problems of difficult spatial and temporal dependency capture and concealment and inaccurate predicted value can be effectively solved.
In order to solve the technical problem, an embodiment of the present invention provides a short-term traffic flow prediction method based on a new deep space time adaptive fusion graph network, including the following steps:
s1: a structural diagram G = (N, E, a) of a known sensor network in the intelligent transportation system is constructed. Wherein, the first and the second end of the pipe are connected with each other,
Figure 86313DEST_PATH_IMAGE002
representing a collection of known sensor nodes.
Figure 501114DEST_PATH_IMAGE004
Representing a collection of links between known sensor nodes.
Figure 443662DEST_PATH_IMAGE006
Structure diagram for indicating traffic
Figure 881596DEST_PATH_IMAGE008
Of the adjacent matrix. If it is not
Figure 115263DEST_PATH_IMAGE010
Node and
Figure 587832DEST_PATH_IMAGE012
a link connection exists between the nodes, then
Figure DEST_PATH_IMAGE013
Description of the preferred embodiments
Figure 498020DEST_PATH_IMAGE014
Node and
Figure DEST_PATH_IMAGE015
there is an adjacency of traffic flow. If it is not
Figure 423250DEST_PATH_IMAGE016
Node and
Figure DEST_PATH_IMAGE017
if there is no link connection between nodes, then
Figure 953283DEST_PATH_IMAGE018
. At each time step t of the time sequence,
Figure DEST_PATH_IMAGE019
a dynamic feature matrix representing N sensors including D graphical signals (e.g., road network occupancy, traffic speed, etc.). The invention aims to learn a function by continuous iterative training
Figure 342676DEST_PATH_IMAGE020
It can be used
Figure DEST_PATH_IMAGE021
To predict
Figure 627027DEST_PATH_IMAGE022
：
Figure DEST_PATH_IMAGE023
Wherein
Figure 852603DEST_PATH_IMAGE024
Representing past traffic sequences and
Figure DEST_PATH_IMAGE025
representing a future traffic sequence.
S2: and constructing a space-time Adaptive Fusion Construction Module (STAFCM) block to establish a space-time Fusion adjacency matrix and an integrated space-time dependency relationship. The space-time adaptive fusion building block is shown in figure 2. Spatio-temporal fusion adjacency matrix
Figure 677340DEST_PATH_IMAGE026
From a time-adjacency matrix
Figure DEST_PATH_IMAGE027
Spatially adjacent matrix
Figure 858922DEST_PATH_IMAGE028
And time connectivity graph
Figure DEST_PATH_IMAGE029
And (5) matrix composition. In the present invention, K is set to 4 (as can be seen in fig. 1), which represents the size of the spatio-temporal fusion adjacency matrix. Fast Dynamic Time Warping (Fast-DTW) algorithm for constructing Time adjacency matrix
Figure 625759DEST_PATH_IMAGE030
And adding the similarity of the time trend to
Figure DEST_PATH_IMAGE031
In (1). Spatial adjacency matrix
Figure 587899DEST_PATH_IMAGE032
An adjacency matrix a representing a spatial adjacency of the traffic network. Time connectivity graph
Figure DEST_PATH_IMAGE033
And the connection of the same node in the latest time step is shown, and the approximate correlation of the latest time step is shown. At each node
Figure 967059DEST_PATH_IMAGE034
When is coming into contact with
Figure DEST_PATH_IMAGE035
And
Figure 3148DEST_PATH_IMAGE036
then, then
Figure DEST_PATH_IMAGE037
It represents the connection of nodes at adjacent time steps t. Each node in the network can be derived from the network by multiple matrix multiplications
Figure 691618DEST_PATH_IMAGE032
Aggregating spatial correlations, from
Figure 78737DEST_PATH_IMAGE030
Learning temporal pattern correlations and slaves
Figure 494544DEST_PATH_IMAGE033
Resulting in an approximate long correlation axis of itself. Construction of
Figure 650719DEST_PATH_IMAGE031
The combination of (a) is shown in the STAFCM of fig. 1.
S3: gated Convolution Modules (GCMs) were designed that capture long-term spatio-temporal information using large inflation rates. Since the gate convolution module uses a larger spreading rate, the gate convolution module is different from the TCN in GraphWaveNet and STGCN to extract more long-term spatio-temporal dependencies.
Known as input
Figure 447773DEST_PATH_IMAGE038
,
Figure 100002_DEST_PATH_IMAGE039
A sigmoid activation function is represented,
Figure 118926DEST_PATH_IMAGE040
representing the tanh activation function, the gated convolution module can be expressed as:
Figure 292418DEST_PATH_IMAGE041
wherein
Figure 303100DEST_PATH_IMAGE042
And
Figure 84105DEST_PATH_IMAGE043
represents a convolution operation for two convolution functions,
Figure 180237DEST_PATH_IMAGE044
is a Hadamard product, this convolution function sets the expansion coefficient, which is K-1, controlling the jump distance, enlarging the receptive field along the time axis.
S4: a spatio-Temporal Adaptive Fusion graph neural module (STAFM) is constructed for deep convolution of a learnable model, capturing hidden spatio-Temporal features, and supplementing incomplete spatio-Temporal connections. As shown in fig. 1, the spatio-temporal adaptive fused map neural module is composed of a fused adaptive convolutional layer and a stacked gate multiplication layer, and then uses a residual connection with zero initialization instead of the residual connection, and outputs through a maximum pool layer.
The neural module of the space-time self-adaptive fusion map adaptively learns the fusion adjacency matrix by fusing the self-adaptive convolution layer, continuously learns and supplements the fusion adjacency matrix by end-to-end training, and finally constructs the self-adaptive fusion adjacency matrix
Figure 626262DEST_PATH_IMAGE045
。
Figure 819346DEST_PATH_IMAGE046
Is defined as
Figure 958203DEST_PATH_IMAGE047
，
Figure 541631DEST_PATH_IMAGE048
And
Figure 102931DEST_PATH_IMAGE049
node embeddings representing the originating node and the target node, and also a learnable parameter. In a directed graph, the output Y of the fused adaptive convolutional layer definition can be formulated as:
Figure 88205DEST_PATH_IMAGE050
wherein
Figure 460281DEST_PATH_IMAGE051
And
Figure 531005DEST_PATH_IMAGE052
the respective representations are bi-directional diffusion matrices, including forward diffusion and backward diffusion. X is the input to this layer.
Figure 584411DEST_PATH_IMAGE053
、
Figure 237241DEST_PATH_IMAGE054
And
Figure 717901DEST_PATH_IMAGE055
representing the trainable parameters of the three k-th layers. The last layer is the maximum pool layer, which connects each hidden state
Figure 10342DEST_PATH_IMAGE056
。
The neural module of the space-time self-adaptive fusion diagram is also provided with a gating multiplication layer which is used for capturing hidden space-time correlation and integrating complex space-time correlation, a matrix multiplication is used for replacing a spectrum filter, and a gating linear unit can generalize global characteristics after nonlinear activation.
Figure 929756DEST_PATH_IMAGE057
Wherein
Figure 624043DEST_PATH_IMAGE058
Is a product of the Hadamard sum of the signals,
Figure 275604DEST_PATH_IMAGE059
a sigmoid function is represented as a function,
Figure 638364DEST_PATH_IMAGE060
is the first
Figure 33573DEST_PATH_IMAGE062
The hidden layer state of a layer.
Figure 644683DEST_PATH_IMAGE063
And
Figure 732725DEST_PATH_IMAGE064
two offsets are indicated.
Figure 999758DEST_PATH_IMAGE065
And
Figure 746129DEST_PATH_IMAGE066
two trainable matrices. The input of the spatio-temporal adaptive fusion map neural module may be from T slices to T-K +1 in time series. Is provided with
Figure 415007DEST_PATH_IMAGE067
Input states of intermediate time steps in the spatial-temporal adaptive fusion graph neural module
Figure 673950DEST_PATH_IMAGE068
The clipping connection can be formulated as:
Figure 225017DEST_PATH_IMAGE069
s5: based on the gated Convolution module of step 3 and the spatio-Temporal Adaptive Fusion map neural module of step 4, a spatio-Temporal Adaptive Fusion Convolution Layer (stfagnayer) is designed, which is a combination of the gated Convolution module and the spatio-Temporal Adaptive Fusion map neural module. As shown in fig. 1, their combination method is parallel connection and then integration by summation. The convolutional layer can increase the receptive field along the time axis and extract long-term spatio-temporal correlations in the gated convolution block. The layer can also establish an adaptive fusion adjacency matrix, complementary spatio-temporal connections in the spatio-temporal adaptive fusion map neural block.
S6: after each module is constructed, residual error connection (Rezero) with zero initialization is adopted to replace the residual error to connect each layer and each module so as to accelerate the speed and the convergence speed of the train. Residual connection with zero initialization is a simple modification to the deep residual network architecture that can facilitate dynamic equidistant and efficient training of extremely deep networks.
Figure 227608DEST_PATH_IMAGE070
Wherein
Figure 750994DEST_PATH_IMAGE071
Learnable parameters representing the remaining weights for the ith layer,
Figure 226843DEST_PATH_IMAGE072
as a function of the output of the i-th layer,
Figure 468469DEST_PATH_IMAGE073
is an implicit state of the (i + 1) th layer.
The short-term traffic flow prediction can be carried out by constructing a space-time self-adaptive fusion graph network through 6 steps.
S7: by connecting the residualsThe invention constructs a space-time Adaptive Fusion Graph Network (STFAGN) by connecting each module, and uses a Huber function as a loss function. The network being in S1
Figure 9171DEST_PATH_IMAGE074
A function. In order to learn parameters in the network, the invention uses a Huber loss function to calculate loss, and then uses a random gradient descent method to continuously iterate learning. Is provided with
Figure 387063DEST_PATH_IMAGE076
As the sensitivity of the over-parameter control square error, the Huber loss function calculation method is as follows:
Figure 50126DEST_PATH_IMAGE077
Figure 513468DEST_PATH_IMAGE078
wherein
Figure 670911DEST_PATH_IMAGE079
And (4) traffic flow data in the predicted future T time period. The multiple STAFM layers operate on the input signal in parallel, extracting the spatio-temporal correlation by gated multiplication. The input size of the final model is
Figure 168889DEST_PATH_IMAGE080
Output size of
Figure 737273DEST_PATH_IMAGE081
. The short-term traffic flow prediction can be carried out by constructing a space-time self-adaptive fusion graph network through 7 steps.
Wherein the spatio-temporal fusion adjacency matrix
Figure 953491DEST_PATH_IMAGE082
From a time-adjacency matrix
Figure 835996DEST_PATH_IMAGE083
Spatially adjacent matrix
Figure 500065DEST_PATH_IMAGE084
And time connectivity graph
Figure 177034DEST_PATH_IMAGE085
A matrix composition, the time adjacency matrix
Figure 614968DEST_PATH_IMAGE086
Using a fast dynamic time warping structure and adding similarity of time trend to the spatio-temporal fusion adjacency matrix
Figure 363481DEST_PATH_IMAGE087
The spatio-temporal fusion adjacency matrix
Figure 570472DEST_PATH_IMAGE088
Is of the size of
Figure 683921DEST_PATH_IMAGE089
。
Wherein, the space-time adaptive fusion convolution layer in the S3 is further used for establishing an adaptive fusion adjacency matrix, and space-time connection supplemented in the space-time adaptive fusion diagram neural module block.
Wherein the output Y of the gated convolution module in S4 is:
Figure 422201DEST_PATH_IMAGE090
wherein the input is
Figure 708826DEST_PATH_IMAGE091
,
Figure 347487DEST_PATH_IMAGE092
A sigmoid activation function is represented,
Figure 694154DEST_PATH_IMAGE093
the tan h activation function is expressed as,
Figure 841102DEST_PATH_IMAGE094
and
Figure 682150DEST_PATH_IMAGE095
represents a convolution operation for two convolution functions,
Figure 255212DEST_PATH_IMAGE097
is the Hadamard product.
In S5, the spatio-temporal adaptive fused graph neural module adaptively learns the fused adjacency matrix through the fused adaptive convolution layer, and continuously learns and supplements the fused adjacency matrix through end-to-end training, thereby finally constructing the adaptive fused adjacency matrix
Figure 772781DEST_PATH_IMAGE098
；
In the directed graph, the output Y defined by the fusion adaptive convolutional layer is:
Figure 485653DEST_PATH_IMAGE099
wherein
Figure 51764DEST_PATH_IMAGE100
And
Figure 150170DEST_PATH_IMAGE101
respectively, representing a diffusion matrix that is bi-directional, including forward and reverse directions, X being the input to the layer,
Figure 87908DEST_PATH_IMAGE102
、
Figure 271765DEST_PATH_IMAGE103
and
Figure 516932DEST_PATH_IMAGE104
representing trainable parameters for three kth layers;
a gating multiplication layer is designed in the space-time self-adaptive fusion graph neural module and is used for capturing hidden space-time correlation and integrated complex space-time correlation, a matrix multiplication is used for replacing a spectrum filter, and a gating linear unit can generalize global characteristics after nonlinear activation;
the output Y of the gated convolution module is:
Figure 984692DEST_PATH_IMAGE105
wherein
Figure 844063DEST_PATH_IMAGE107
Is a product of the Hadamard sum of the signals,
Figure 265948DEST_PATH_IMAGE108
a sigmoid function is represented by a function of,
Figure 236178DEST_PATH_IMAGE109
is the first
Figure 558444DEST_PATH_IMAGE110
The hidden layer state of a layer is,
Figure 588717DEST_PATH_IMAGE111
and
Figure 684849DEST_PATH_IMAGE112
two offset amounts are indicated, and,
Figure 209502DEST_PATH_IMAGE113
and
Figure 340270DEST_PATH_IMAGE114
two trainable matrices.
The embodiment of the invention has the following beneficial effects:
1. the fixed adjacency matrix graph in the prior art cannot be directly related to the prediction task, it lacks some spatial connection information, which may lead to considerable bias. The invention provides a self-adaptive fusion adjacency matrix, each node in the graph not only can integrate the spatial correlation, the time mode information and the correlation of the adjacent time axis, but also can self-adaptively perfect the space-time correlation, and the prediction result of the traffic sequence is more accurate.
2. Traditional traffic flow prediction is based on a spatial adjacency matrix, and time correlation is not considered when constructing the adjacency matrix. The invention uses a fast dynamic time warping algorithm to measure the time sequence correlation, constructs a time adjacency matrix and then further establishes a self-adaptive fusion adjacency matrix. In order to accelerate convergence, the invention also introduces ReLU connection to replace residual connection, and accelerates the speed of network information transmission and training in training.
Drawings
FIG. 1 is a network framework diagram of a spatio-temporal adaptive fusion graph network;
FIG. 2 is a schematic diagram of one layer in the spatio-temporal adaptive fusion map neural module.
Detailed Description
To make the objects, technical solutions and advantages of the present invention more apparent, the present invention will be described in further detail with reference to the accompanying drawings.
The short-term traffic flow prediction method based on the new deep space-time Adaptive Fusion Graph Network uses a new deep space-time Adaptive Fusion Graph model (STFAGN) to learn icons by constructing an Adaptive adjacency matrix and fusing the space-time Adaptive Fusion adjacency matrix, and the Network can effectively solve the problems of difficulty in capturing hidden space-time dependency and inaccurate predicted value.
The method can be continuously updated in end-to-end learning so as to dynamically update the relationship between the road nodes along with the change of time to effectively learn the node characteristics. For space-time fusion matrix
Figure 541444DEST_PATH_IMAGE031
The matrix includes an adjacency matrix of temporal similarities calculated by a Dynamic Time Warping algorithm (Fast-DTW)
Figure 859293DEST_PATH_IMAGE030
The dynamic time normalization algorithm can process the condition that the sequence is not parallel but the shape of the sequence is similar, and can process and calculate the similarity of traffic flow data with high nonlinearity and dynamic change to obtain an accurate time phase diagram; secondly, the space-time fusion matrix also comprises a space adjacency matrix constructed based on distance
Figure 374588DEST_PATH_IMAGE032
And connections based on step size of the same node at the latest time
Figure 677305DEST_PATH_IMAGE033
. A bipartite graph diffusion convolution is used to capture the spatial dependencies between graph nodes. The continuous updating of the adaptive matrix can overcome the defects that the provided graph information of the original adjacency matrix constructed based on the distance is incomplete and inaccurate. Spatio-temporal fusion matrix
Figure 987064DEST_PATH_IMAGE088
The relation among the road nodes is provided from multiple angles, the capability of the model can be improved to capture the characteristics of the road nodes, the potential road space dependency can be better captured, and the accuracy of the predicted value is improved.
The invention aims to provide a short-term traffic flow prediction method based on a new deep space time self-adaptive fusion graph network, which aims at the problem that incomplete and missing connection exists in a fixed graph structure and adopts end-to-end convolution network training to self-adaptively complement an adjacent matrix of fusion space-time correlation so as to capture the hidden space-time correlation in data and further complement incomplete information.
The space-time adaptive fusion map network consists of three modules, including a space-time adaptive fusion construction module (STAFCM), a space-time adaptive fusion map neural module (STAFM) and a Gate Convolution Module (GCM).
Firstly, a space-time self-adaptive fusion construction module establishes a space-time fusion adjacency matrix
Figure 792208DEST_PATH_IMAGE088
To integrate the spatiotemporal information.
Figure 907932DEST_PATH_IMAGE087
Time-adjacency matrix including computation by fast dynamic time warping
Figure 747712DEST_PATH_IMAGE086
Spatially adjacent matrix
Figure 228372DEST_PATH_IMAGE115
And time connectivity graph
Figure 599442DEST_PATH_IMAGE116
To represent a given spatio-temporal connection in the flow graph.
Figure 190960DEST_PATH_IMAGE088
The combination of (a) and (b) is shown in figure 1. Second, the spatio-temporal adaptive fusion construction module in the fusion adaptive convolutional layer is used to complete the fully fused adjacency matrix
Figure 150826DEST_PATH_IMAGE088
For hiding spatiotemporal features in Gated Multiplication Layers (Gated Multiplication Layers).
The spatio-temporal adaptive fused graph neural module basically consists of a fused adaptive convolution layer and a gated multiplication layer superposed by a maximum pool layer. The fused adaptive convolutional layer learns the adjacency matrix from the data through end-to-end supervised training to construct an adaptive fused adjacency matrix
Figure 864704DEST_PATH_IMAGE117
. Gated multiplication layer pass-through
Figure 644441DEST_PATH_IMAGE118
Matrix multiplication is performed to aggregate spatio-temporal dependencies.
Thirdly, the gating convolution module extracts long-range space-time correlation through a large expansion rate to serve as a gating mechanism of the recurrent neural network. The space-time self-adaptive fusion map neural module and the gated convolution module jointly form a space-time self-adaptive fusion map convolution layer so as to capture space-time characteristics. We stack K layers of spatio-temporal adaptive fusion graph neural modules to capture hidden spatio-temporal dependencies.
The invention solves the technical problems through the following technical scheme: the space-time self-adaptive fusion graph network. The construction steps are as follows:
step 1: a structural diagram G = (N, E, a) of a known sensor network in the intelligent transportation system is constructed. Where N represents a set of known sensor nodes. E denotes the set of links between known sensor nodes. A denotes an adjacency matrix of the traffic structure diagram G. If it is not
Figure 39650DEST_PATH_IMAGE119
Node and
Figure 165607DEST_PATH_IMAGE120
a link connection exists between the nodes, then
Figure 988069DEST_PATH_IMAGE122
Description of the invention
Figure 255103DEST_PATH_IMAGE123
Node and
Figure 516320DEST_PATH_IMAGE124
there is an adjacency of traffic flow. If it is used
Figure 185198DEST_PATH_IMAGE123
Node and
Figure 178562DEST_PATH_IMAGE124
if there is no link connection between nodes, then
Figure 745941DEST_PATH_IMAGE125
. At each time step t of the time sequence,
Figure 748532DEST_PATH_IMAGE126
a dynamic feature matrix representing N sensors, including DA graphical signal (e.g., road network occupancy, traffic speed, etc.). The invention aims to learn a function by continuous iterative training
Figure 271917DEST_PATH_IMAGE127
It can use
Figure 498499DEST_PATH_IMAGE128
To predict
Figure 740125DEST_PATH_IMAGE129
：
Figure 592412DEST_PATH_IMAGE130
Wherein
Figure 970304DEST_PATH_IMAGE131
Representing past traffic sequences and
Figure 571049DEST_PATH_IMAGE132
representing a future traffic sequence.
Step 2: and constructing a space-time Adaptive Fusion Construction Module (STAFCM) block to establish a space-time Fusion adjacency matrix and an integrated space-time dependency relationship. The space-time adaptive fusion building block is shown in figure 2. Spatio-temporal fusion adjacency matrix
Figure 96709DEST_PATH_IMAGE133
From a time-adjacency matrix
Figure 441102DEST_PATH_IMAGE134
Spatially adjacent matrix
Figure 486550DEST_PATH_IMAGE135
And time connectivity graph
Figure 320514DEST_PATH_IMAGE136
And (5) forming a matrix. In the present invention, K is set to 4 (FIG. 4)1) which represents the size of the spatio-temporal fusion adjacency matrix. Fast Dynamic Time Warping (Fast-DTW) algorithm for constructing Time adjacency matrix
Figure 271152DEST_PATH_IMAGE030
And adding the similarity of the time trend to
Figure 153657DEST_PATH_IMAGE137
In (1). Spatial adjacency matrix
Figure 817726DEST_PATH_IMAGE032
An adjacency matrix a representing a spatial adjacency of the traffic network. Time connectivity graph
Figure 760274DEST_PATH_IMAGE138
The connection of the same node in the latest time step is shown, and the approximate correlation of the latest time step is shown. At each node
Figure 260526DEST_PATH_IMAGE139
When it comes to
Figure 946722DEST_PATH_IMAGE140
And
Figure 153712DEST_PATH_IMAGE141
then, then
Figure 80211DEST_PATH_IMAGE142
It represents the connection of nodes at adjacent time steps t. Each node in the network can be driven by multiple matrix multiplications
Figure 5442DEST_PATH_IMAGE032
Aggregating spatial correlations, from
Figure 229750DEST_PATH_IMAGE030
Learning temporal pattern correlations and slaves
Figure 619143DEST_PATH_IMAGE033
Resulting in an approximate long correlation axis of itself. Construction of
Figure 637914DEST_PATH_IMAGE137
The combination of (a) is shown in the STAFCM of fig. 1.
And step 3: gated Convolution Modules (GCMs) were designed that capture long-term spatio-temporal information using large inflation rates. Since the gate convolution module uses a larger spreading rate, the gate convolution module is different from the TCN in GraphWaveNet and STGCN to extract more long-term spatio-temporal dependencies.
Known as input
Figure 50441DEST_PATH_IMAGE143
,
Figure 384165DEST_PATH_IMAGE144
A sigmoid activation function is represented,
Figure 565748DEST_PATH_IMAGE145
representing the tanh activation function, the gated convolution module can be represented as:
Figure 83317DEST_PATH_IMAGE146
wherein
Figure 717560DEST_PATH_IMAGE147
And
Figure 283671DEST_PATH_IMAGE148
represents a convolution operation for two convolution functions,
Figure 398389DEST_PATH_IMAGE149
is a Hadamard product, and the convolution function sets an expansion coefficient which is K-1, controls the jump distance and enlarges the receptive field along the time axis.
And 4, step 4: a spatio-Temporal Adaptive Fusion graph neural module (STAFM) is constructed for deep convolution of a learnable model, capturing hidden spatio-Temporal features, and supplementing incomplete spatio-Temporal connections. As shown in fig. 1, the spatio-temporal adaptive fused map neural module is composed of a fused adaptive convolutional layer and a stacked gate multiplication layer, and then uses a residual connection with zero initialization instead of the residual connection, and outputs through a maximum pool layer.
The neural module of the space-time self-adaptive fusion map adaptively learns the fusion adjacency matrix by fusing the self-adaptive convolution layer, continuously learns and supplements the fusion adjacency matrix by end-to-end training, and finally constructs the self-adaptive fusion adjacency matrix
Figure 24542DEST_PATH_IMAGE150
。
Figure 146082DEST_PATH_IMAGE117
Is defined as
Figure 578200DEST_PATH_IMAGE151
，
Figure 734375DEST_PATH_IMAGE152
And
Figure 843014DEST_PATH_IMAGE153
node embeddings representing the originating node and the target node, and also a learnable parameter. In a directed graph, the output Y of the fused adaptive convolutional layer definition can be formulated as:
Figure 451850DEST_PATH_IMAGE154
wherein
Figure 359763DEST_PATH_IMAGE155
And
Figure 432761DEST_PATH_IMAGE156
respectively, represent diffusion matrices that are bi-directional, including forward diffusion and backward diffusion. X is the input to the layer.
Figure 400717DEST_PATH_IMAGE157
、
Figure 496849DEST_PATH_IMAGE158
And
Figure 21503DEST_PATH_IMAGE159
representing the trainable parameters of the three k-th layers. The last layer is the maximum pool layer, which connects each hidden state
Figure 886691DEST_PATH_IMAGE160
。
The neural module of the space-time self-adaptive fusion diagram is also provided with a gating multiplication layer which is used for capturing hidden space-time correlation and integrating complex space-time correlation, a matrix multiplication is used for replacing a spectrum filter, and a gating linear unit can generalize global characteristics after nonlinear activation.
Figure 25548DEST_PATH_IMAGE161
Wherein
Figure 671293DEST_PATH_IMAGE162
Is a product of the Hadamard sum of the signals,
Figure 186588DEST_PATH_IMAGE163
a sigmoid function is represented as a function,
Figure 906282DEST_PATH_IMAGE164
is the first
Figure 527625DEST_PATH_IMAGE062
The hidden layer state of a layer.
Figure 598349DEST_PATH_IMAGE111
And
Figure 651756DEST_PATH_IMAGE112
two offsets are indicated.
Figure 553853DEST_PATH_IMAGE165
And
Figure 847562DEST_PATH_IMAGE166
two trainable matrices. The input of the spatio-temporal adaptive fusion map neural module may be from T slices to T-K +1 in time series. Is provided with
Figure 405582DEST_PATH_IMAGE167
Input states of intermediate time steps in the spatial-temporal adaptive fusion diagram neural module
Figure 997101DEST_PATH_IMAGE168
The clipping join can be formulated as:
Figure 753704DEST_PATH_IMAGE169
。
and 5: based on the gated Convolution module of step 3 and the spatio-Temporal Adaptive Fusion map neural module of step 4, a spatio-Temporal Adaptive Fusion Convolution Layer (stfagnayer) is designed, which is a combination of the gated Convolution module and the spatio-Temporal Adaptive Fusion map neural module. As shown in fig. 1, their combination method is parallel connection and then integration by summation. The convolutional layer can increase the receptive field along the time axis and extract long-term spatio-temporal correlations in the gated convolution block. The layer can also establish an adaptive fusion adjacency matrix, complementary spatio-temporal connections in the spatio-temporal adaptive fusion map neural block.
Step 6: after each module is constructed, residual error connection (Rezero) with zero initialization is adopted to replace the residual error to connect each layer and the module so as to accelerate the speed and the convergence speed of the train. Residual concatenation with zero initialization is a simple modification to the deep residual network architecture that facilitates dynamic equidistant and efficient training of extremely deep networks.
Figure 670845DEST_PATH_IMAGE170
Wherein
Figure 450582DEST_PATH_IMAGE171
Learnable parameters representing the remaining weights for the ith layer,
Figure 157376DEST_PATH_IMAGE172
as a function of the output of the i-th layer,
Figure 971748DEST_PATH_IMAGE173
is an implicit state of the (i + 1) th layer.
The short-term traffic flow prediction can be carried out by constructing a space-time self-adaptive fusion graph network through 6 steps.
And 7: by connecting each module through residual error, the invention constructs a space-time Adaptive Fusion Graph Network (STFAGN). Such a network is that of step 1
Figure 856527DEST_PATH_IMAGE174
A function. In order to learn parameters in the network, the invention uses a Huber loss function to calculate loss, and then uses a random gradient descent method to continuously iterate learning. Is provided with
Figure 123560DEST_PATH_IMAGE176
As the sensitivity of the over-parameter control square error, the Huber loss function calculation method is as follows:
Figure 135510DEST_PATH_IMAGE177
Figure 804389DEST_PATH_IMAGE178
wherein
Figure 797752DEST_PATH_IMAGE179
And (4) traffic flow data in the predicted future T time period. Multiple STAFM layers operating in parallelAnd (4) inputting signals, and extracting space-time correlation through gated multiplication. The input size of the final model is
Figure 614399DEST_PATH_IMAGE180
Output size of
Figure 351411DEST_PATH_IMAGE181
. The short-term traffic flow prediction can be carried out by constructing a space-time self-adaptive fusion graph network through 7 steps.
The work flow of the deep space time self-adaptive fusion graph model provided by the invention can be summarized as follows: an adaptive adjacency matrix is first constructed, which may be continuously updated during each iterative training. And then, carrying out parallel fusion on the graphs constructed from different angles according to the traffic flow data to obtain a new space-time fusion matrix. And simultaneously carrying out graph diffusion convolution on the adaptive matrix and the space-time fusion matrix to capture hidden space-time dependency, and carrying out training on a deeper network model by using the finally captured characteristics to obtain a prediction result. The invention carries out test experiments on a plurality of traffic flow data sets, and the experimental result shows that the network performance is superior to the prior most advanced method.
While the invention has been described in connection with what is presently considered to be the most practical and preferred embodiment, it is to be understood that the invention is not to be limited to the disclosed embodiment, but on the contrary, is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims.

Claims (5)

1. A short-term traffic flow prediction method based on a new deep space-time adaptive fusion graph network is characterized by comprising the following steps:
s1: constructing a structural diagram G = (N, E, A) of a known sensor network in the intelligent transportation system, wherein N represents a set of known sensor nodes, E represents a set of links between the known sensor nodes, A represents an adjacency matrix of the structural diagram G, and
Figure 76550DEST_PATH_IMAGE001
representing a dynamic feature matrix, using
Figure 99738DEST_PATH_IMAGE002
To predict
Figure 964926DEST_PATH_IMAGE003
：
Figure 103783DEST_PATH_IMAGE004
Wherein
Figure 749528DEST_PATH_IMAGE005
Representing past traffic sequences and
Figure 999244DEST_PATH_IMAGE006
representing a future traffic sequence;
s2: constructing a spatio-temporal adaptive fusion construction module for establishing a spatio-temporal fusion adjacency matrix
Figure 797567DEST_PATH_IMAGE007
And integrating spatiotemporal dependencies;
s3: establishing a gate control convolution module which captures long-term space-time information by using a large expansion rate so as to extract more long-term space-time dependency relations;
s4: constructing a space-time self-adaptive fusion map neural module for deeply convolving a learnable model, capturing hidden space-time characteristics and supplementing incomplete space-time connection;
s5: constructing a space-time self-adaptive fusion map neural module for a deep convolution learnable model, capturing hidden space-time characteristics and supplementing incomplete space-time connection, wherein the space-time self-adaptive fusion map neural module consists of a fusion self-adaptive convolution layer and a laminated gate multiplication layer, replaces residual connection with zero initialization by residual connection, and outputs through a maximum pool layer;
s6: connecting each layer and each module by using residual connection with zero initialization instead of the residual connection so as to accelerate the speed and the convergence speed of the train;
s7: after connecting all the modules through residual errors, a space-time self-adaptive fusion graph network is constructed, and a Huber loss function learning network is used.
2. The method for predicting short-term traffic flow based on new deep space-time adaptive fusion graph network according to claim 1, wherein the space-time fusion adjacency matrix
Figure 107326DEST_PATH_IMAGE008
From a time-adjacency matrix
Figure 178050DEST_PATH_IMAGE009
Spatially adjacent matrix
Figure 293773DEST_PATH_IMAGE010
And time connectivity graph
Figure 133553DEST_PATH_IMAGE011
A matrix composition, the time adjacency matrix
Figure 614213DEST_PATH_IMAGE012
Using a fast dynamic time warping construct and adding similarity of temporal trends to the spatio-temporal fusion adjacency matrix
Figure 224098DEST_PATH_IMAGE013
The spatio-temporal fusion adjacency matrix
Figure 81196DEST_PATH_IMAGE008
Is of a size of
Figure 775482DEST_PATH_IMAGE014
。
3. The method for predicting short-term traffic flow based on new deep space-time adaptive fusion map network according to claim 2, wherein the spatio-temporal adaptive fusion convolution layer in S3 is further used for establishing an adaptive fusion adjacency matrix, and spatio-temporal connections are supplemented in the spatio-temporal adaptive fusion map neural module.
4. The short-term traffic flow prediction method based on the new deep space-time adaptive fusion graph network according to claim 3, wherein the output Y of the gated convolution module in S4 is:
Figure 489360DEST_PATH_IMAGE015
wherein the input is
Figure 534677DEST_PATH_IMAGE016
,
Figure 929886DEST_PATH_IMAGE017
A sigmoid activation function is represented,
Figure 557308DEST_PATH_IMAGE018
the tan h activation function is expressed as,
Figure 645349DEST_PATH_IMAGE019
and
Figure 912383DEST_PATH_IMAGE020
represents a convolution operation for two convolution functions,
Figure 908020DEST_PATH_IMAGE022
is the Hadamard product.
5. The short-term traffic flow prediction method based on the new deep space-time adaptive fusion map network according to claim 4, wherein in S5, the spatio-temporal adaptive fusion map neural module adaptively learns the fusion adjacency matrix through the fusion adaptive convolution layer, and continuously learns and trains from end to endSupplementing the fusion adjacency matrix and finally constructing the self-adaptive fusion adjacency matrix
Figure 311320DEST_PATH_IMAGE023
；
In the directed graph, the output Y defined by the fusion adaptive convolutional layer is:
Figure 570263DEST_PATH_IMAGE024
wherein
Figure 636177DEST_PATH_IMAGE025
And
Figure 373189DEST_PATH_IMAGE026
respectively, the diffusion matrices, which are bi-directional, including forward and reverse,
Figure 896574DEST_PATH_IMAGE028
is the input to the layer or layers,
Figure 388735DEST_PATH_IMAGE029
、
Figure 364781DEST_PATH_IMAGE030
and
Figure 718534DEST_PATH_IMAGE031
representing trainable parameters for three kth layers;
a gating multiplication layer is designed in the space-time self-adaptive fusion graph neural module and is used for capturing hidden space-time correlation and integrated complex space-time correlation, a matrix multiplication is used for replacing a spectrum filter, and a gating linear unit can generalize global characteristics after nonlinear activation;
the output Y of the gated convolution module is:
Figure 362004DEST_PATH_IMAGE032
wherein
Figure 697171DEST_PATH_IMAGE033
Is a product of the Hadamard functions,
Figure 753989DEST_PATH_IMAGE034
a sigmoid function is represented as a function,
Figure 832803DEST_PATH_IMAGE035
is the hidden layer state of the l-th layer,
Figure 330780DEST_PATH_IMAGE036
and
Figure 882853DEST_PATH_IMAGE037
two offset amounts are indicated, and,
Figure 99071DEST_PATH_IMAGE038
and
Figure DEST_PATH_IMAGE039
two trainable matrices.
CN202210532872.7A 2022-05-17 2022-05-17 Short-term traffic flow prediction method based on new deep space time self-adaptive fusion graph network Pending CN115019504A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210532872.7A CN115019504A (en) 2022-05-17 2022-05-17 Short-term traffic flow prediction method based on new deep space time self-adaptive fusion graph network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210532872.7A CN115019504A (en) 2022-05-17 2022-05-17 Short-term traffic flow prediction method based on new deep space time self-adaptive fusion graph network

Publications (1)

Publication Number Publication Date
CN115019504A true CN115019504A (en) 2022-09-06

Family

ID=83068332

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210532872.7A Pending CN115019504A (en) 2022-05-17 2022-05-17 Short-term traffic flow prediction method based on new deep space time self-adaptive fusion graph network

Country Status (1)

Country Link
CN (1) CN115019504A (en)

Citations (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20160379486A1 (en) * 2015-03-24 2016-12-29 Donald Warren Taylor Apparatus and system to manage monitored vehicular flow rate
CN110021165A (en) * 2019-03-18 2019-07-16 浙江工业大学 A kind of traffic flow forecasting method based on Autoencoder-LSTM Fusion Model
WO2020024319A1 (en) * 2018-08-01 2020-02-06 苏州大学张家港工业技术研究院 Convolutional neural network based multi-point regression forecasting model for traffic flow forecasting
CN111260131A (en) * 2020-01-16 2020-06-09 汕头大学 Short-term traffic flow prediction method and device
CN111861027A (en) * 2020-07-29 2020-10-30 北京工商大学 Urban traffic flow prediction method based on deep learning fusion model
US20200356839A1 (en) * 2019-05-09 2020-11-12 ClimateAI, Inc. Systems and methods for selecting global climate simulation models for training neural network climate forecasting models
CN112382094A (en) * 2020-11-13 2021-02-19 北京航空航天大学 Urban traffic flow interpretable prediction method based on space-time potential energy field
CN112466117A (en) * 2020-11-24 2021-03-09 南通大学 Road network short-term traffic flow prediction method based on deep space-time residual error network
CN112749832A (en) * 2020-12-03 2021-05-04 南京邮电大学 Traffic flow prediction method based on graph convolution embedded space-time duration memory network
US11238729B1 (en) * 2020-09-11 2022-02-01 Toyota Motor Engineering & Manufacturing North America, Inc. Systems and methods for traffic flow prediction
CN114492992A (en) * 2022-01-25 2022-05-13 重庆邮电大学 Self-adaptive space-time graph neural network traffic flow prediction method and system based on Transformer

Patent Citations (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20160379486A1 (en) * 2015-03-24 2016-12-29 Donald Warren Taylor Apparatus and system to manage monitored vehicular flow rate
WO2020024319A1 (en) * 2018-08-01 2020-02-06 苏州大学张家港工业技术研究院 Convolutional neural network based multi-point regression forecasting model for traffic flow forecasting
CN110021165A (en) * 2019-03-18 2019-07-16 浙江工业大学 A kind of traffic flow forecasting method based on Autoencoder-LSTM Fusion Model
US20200356839A1 (en) * 2019-05-09 2020-11-12 ClimateAI, Inc. Systems and methods for selecting global climate simulation models for training neural network climate forecasting models
CN111260131A (en) * 2020-01-16 2020-06-09 汕头大学 Short-term traffic flow prediction method and device
CN111861027A (en) * 2020-07-29 2020-10-30 北京工商大学 Urban traffic flow prediction method based on deep learning fusion model
US11238729B1 (en) * 2020-09-11 2022-02-01 Toyota Motor Engineering & Manufacturing North America, Inc. Systems and methods for traffic flow prediction
CN112382094A (en) * 2020-11-13 2021-02-19 北京航空航天大学 Urban traffic flow interpretable prediction method based on space-time potential energy field
CN112466117A (en) * 2020-11-24 2021-03-09 南通大学 Road network short-term traffic flow prediction method based on deep space-time residual error network
CN112749832A (en) * 2020-12-03 2021-05-04 南京邮电大学 Traffic flow prediction method based on graph convolution embedded space-time duration memory network
CN114492992A (en) * 2022-01-25 2022-05-13 重庆邮电大学 Self-adaptive space-time graph neural network traffic flow prediction method and system based on Transformer

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
SHUMIN YANG,ETC.: "Spatiotemporal Adaptive Fusion Graph Network for Short-Term Traffic Flow Forecasting", 《MATHEMATICS》 *

Similar Documents

Publication Publication Date Title
US11270579B2 (en) Transportation network speed foreeasting method using deep capsule networks with nested LSTM models
CN109754605B (en) Traffic prediction method based on attention temporal graph convolution network
CN113313947B (en) Road condition evaluation method of short-term traffic prediction graph convolution network
CN109887282A (en) A kind of road network traffic flow prediction technique based on level timing diagram convolutional network
CN114818515A (en) Multidimensional time sequence prediction method based on self-attention mechanism and graph convolution network
CN111832814A (en) Air pollutant concentration prediction method based on graph attention machine mechanism
CN114299723B (en) Traffic flow prediction method
CN114802296A (en) Vehicle track prediction method based on dynamic interaction graph convolution
CN110401978B (en) Indoor positioning method based on neural network and particle filter multi-source fusion
CN112767682A (en) Multi-scale traffic flow prediction method based on graph convolution neural network
CN112905379B (en) Traffic big data restoration method of graph self-encoder based on self-attention mechanism
CN110428614B (en) Traffic jam heat degree space-time prediction method based on non-negative tensor decomposition
CN112785848B (en) Traffic data prediction method and system
CN114495500B (en) Traffic prediction method based on dual dynamic space-time diagram convolution
CN111554118B (en) Dynamic prediction method and system for bus arrival time
CN114202120A (en) Urban traffic travel time prediction method aiming at multi-source heterogeneous data
CN115862319A (en) Traffic flow prediction method for space-time diagram self-encoder
Kim et al. Urban traffic prediction using congestion diffusion model
Zhang et al. Boosted trajectory calibration for traffic state estimation
CN111815075B (en) Prediction method for transportation travel demand under major public health incident
CN112269931A (en) Data-driven group intelligent interaction relation inference and evolution calculation method
CN115019504A (en) Short-term traffic flow prediction method based on new deep space time self-adaptive fusion graph network
CN114566048B (en) Traffic control method based on multi-view self-adaptive space-time diagram network
CN115797557A (en) Self-supervision 3D scene flow estimation method based on graph attention network
Chen et al. Digital twin mobility profiling: A spatio-temporal graph learning approach

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