CN113450561A - Traffic speed prediction method based on space-time graph convolution-generation countermeasure network - Google Patents

Traffic speed prediction method based on space-time graph convolution-generation countermeasure network Download PDF

Info

Publication number
CN113450561A
CN113450561A CN202110491442.0A CN202110491442A CN113450561A CN 113450561 A CN113450561 A CN 113450561A CN 202110491442 A CN202110491442 A CN 202110491442A CN 113450561 A CN113450561 A CN 113450561A
Authority
CN
China
Prior art keywords
data
convolution
network
traffic
time
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
CN202110491442.0A
Other languages
Chinese (zh)
Other versions
CN113450561B (en
Inventor
郭海锋
吴铨力
刘瑞
程茂恒
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.)
Zhejiang University of Technology ZJUT
Original Assignee
Zhejiang University of Technology ZJUT
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 Zhejiang University of Technology ZJUT filed Critical Zhejiang University of Technology ZJUT
Priority to CN202110491442.0A priority Critical patent/CN113450561B/en
Publication of CN113450561A publication Critical patent/CN113450561A/en
Application granted granted Critical
Publication of CN113450561B publication Critical patent/CN113450561B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

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
    • 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
    • G08G1/0125—Traffic data processing
    • G08G1/0129—Traffic data processing for creating historical data or processing based on historical data
    • 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
    • G08G1/052—Detecting movement of traffic to be counted or controlled with provision for determining speed or overspeed

Landscapes

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

Abstract

A traffic speed prediction method based on a space-time graph convolution-generation countermeasure network comprises the steps of firstly, according to an urban road network structure, taking road sections in a road network as graph nodes, and taking intersections as edges connecting the nodes to construct a traffic graph network; then, acquiring sampled and collected vehicle flow speed data through a vehicle flow speed detector in an urban road network, constructing a characteristic matrix, then, using a constructed space-time graph convolutional network (STGCN) as a generator for generating a countermeasure network and a fully-connected neural network as a discriminator to generate traffic speed data through countermeasure, and performing mutual game training and lifting together to finally obtain an optimal traffic speed prediction model. The final generator for generating the confrontation network is utilized to generate a traffic state data predicted value closest to the real data, and the purpose of predicting the road network state data is achieved. The invention can realize more accurate traffic speed prediction.

Description

Traffic speed prediction method based on space-time graph convolution-generation countermeasure network
Technical Field
The invention relates to the field of intelligent traffic engineering, in particular to a method for predicting traffic states of an urban road network.
Background
With the continuous improvement of national economy, the urbanization construction is continuously promoted, and the number of urban residents is continuously increased. Along with the continuous improvement of the living standard of residents, the number of automobiles is also kept increasing at a high speed, meanwhile, the urban traffic network is also becoming more complicated and complicated, and the occurrence of traffic transportation problems becomes frequent and complicated. In order to alleviate the slow and congested traffic conditions in cities, traffic field researchers begin to predict traffic data and further improve the traffic conditions in cities based on the predicted data.
Because traffic speeds have a strong time-varying nature, short-term traffic speed predictions are often used for data analysis. The predicted data can be used for road condition state prediction analysis, support is provided for third-party services, a traffic navigation system is improved, technical support can be provided for an urban traffic management center, a traffic signal control scheme is regulated and controlled in real time, and the road resource utilization rate is improved. Therefore, traffic data prediction has become a problem of important attention and research of theoretical researchers in the traffic industry.
Many representative traffic speed prediction methods have been proposed at home and abroad nowadays. The method can be mainly divided into three types, wherein the first type is a statistical theory model, the second type is a machine learning model, and the third type is a deep learning model of the present fire and heat. An urban road traffic system is an extremely complex system, and factors such as weather conditions, traveling peaks in the morning and at the evening, weekdays, holidays, traffic accidents and the like influence the operation of the whole traffic system in different ways. Therefore, neither the early statistical theory model nor the machine learning model often has enough model complexity to support high-precision prediction, but deep learning can be easily achieved.
The existing deep learning methods analyze and predict time-series traffic data and analyze and predict traffic space structure feature extraction. The time-space graph convolutional network (STGCN) analyzes time sequence traffic data and a graph space node neighborhood simultaneously, but the prediction methods usually only focus on the second-order neighborhood of a road network node and ignore the global characteristics of a traffic network.
The invention provides a novel traffic speed prediction method by applying a generated countermeasure network technology to the field of traffic data prediction by combining the generated countermeasure network and taking a space-time diagram convolutional network as a generator for generating the countermeasure network.
Disclosure of Invention
The invention aims to overcome the defect that the existing method for predicting the traffic speed of the space-time graph convolutional network ignores the global characteristics of a road network, and provides a short-time traffic speed prediction method combining the space-time graph convolutional network and a generation countermeasure network so as to improve the accuracy of traffic speed prediction.
The method comprises the steps of firstly extracting time sequence characteristics and space characteristics of traffic speed by using a space-time graph convolution network to generate predicted traffic speed, then extracting global traffic speed characteristics by using a generated countermeasure network, scoring generated traffic speed data, and continuously optimizing traffic speed prediction accuracy.
A traffic speed prediction method based on a space-time graph convolution-generation countermeasure network comprises the following steps:
step 1, constructing an adjacency matrix; constructing a traffic network topological structure graph network according to the spatial connection relation among all road sections of an urban traffic network, taking each road section as a road network node, connecting the nodes corresponding to the road sections connected by intersections, constructing a road network graph, obtaining an adjacency matrix among the road sections, and processing to obtain a Laplace matrix; and calculating to obtain a Laplace matrix L according to the acquired adjacency matrix A, carrying out normalization processing on the Laplace matrix, and finally carrying out feature decomposition.
Step 2, obtaining traffic flow speed data of a road network; the method comprises the steps that a traffic speed time sequence of each road section is obtained through regular sampling of an urban road network traffic flow detector;
step 3, constructing a generator for generating a countermeasure network (GAN); the method comprises the steps that a space-time graph convolutional network (STGCN) is used as a generator, the obtained traffic speed time sequence and an adjacent matrix are input into a space-time convolutional neural network of a graph, and a predicted value of the traffic speed of the next period of each road section of a road network is generated;
step 4, constructing a discriminator for generating a countermeasure network (GAN); in the invention, a fully-connected neural network is used as a discriminator, the traffic speed data and the real data generated by a generator are used as input, and the global characteristics of the input data are extracted to judge the authenticity of the data;
step 5, generating traffic data by generating a countermeasure network (GAN) iterative optimization generator model; and alternately and iteratively optimizing the discriminator and the generator to form a countermeasure, calculating an error by using the generated data and the real data during the countermeasure, and finally generating the data with the minimum error with the real data.
In the step 2, the data sampling interval of the urban road network traffic flow speed detector is 5 minutes; the step 2 specifically comprises the following steps:
the method comprises the following steps of cleaning road network vehicle flow speed time sequence characteristic data, removing redundant data and error data, and completing the road section with a deletion phenomenon at a certain moment by adopting a linear interpolation method, wherein the method comprises the following steps:
Figure BDA0003052364830000021
in the step 3, the characteristic extraction is carried out on the traffic flow speed time sequence characteristic data through a space-time graph convolution network, the time relevance of the traffic data is obtained by using a gating sequence convolution layer, and the space relevance of the traffic speed data is obtained by using a space convolution layer. Generator STGCN for generating a countermeasure network, comprising two spatio-Temporal convolution modules and one output layer, wherein each spatio-Temporal convolution module consists of two Gated sequence convolution layers (Temporal gate-Conv) and one Spatial Graph convolution module in between (Spatial Graph-Conv).
In the step 4, the arbiter network for generating the countermeasure network adopts three layers of fully connected neural networks, the number of the hidden nodes in the last layer is 1, the data passes through the ReLU activation function before accessing each layer of connection layer, and the final output of the arbiter network is the true and false probability score of the data generated by the generator and the real data. The generation countermeasure network discriminator adopts three layers of fully connected neural networks, and the number of nodes is 128, 32 and 1 respectively.
In the step 5, the generation countermeasure network continuously counteracts the historical traffic speed data to generate predicted data according to the input historical traffic speed data, finally generates traffic speed data closest to the real data, and calculates and evaluates the error between the generated data and the real data by using the mean absolute error MAE, the root mean square error RMSE and the mean absolute percentage error MAPE.
Through the alternate iteration of the generators and the discriminators in the generation countermeasure network, mutual game and common promotion, a best generator is finally obtained, and a predicted value closest to real data can be generated.
The beneficial effects of the scheme of the invention are as follows:
according to the invention, by using the space-time graph convolution network (STGCN), on the basis of considering the non-European structure between road sections by using graph convolution, the historical time domain correlation of the traffic flow is obtained by one-dimensional convolution of the time domain, the capture of the global characteristics of the traffic graph structure is realized by the fully-connected neural network, the traffic data is analyzed more comprehensively, the more accurate traffic speed prediction is realized, and more accurate data support is provided for relieving urban congestion, increasing the traffic efficiency of an urban road network and ensuring the travel experience of urban residents.
Drawings
FIG. 1 is a block diagram of a generator space-time graph convolutional network for generating a countermeasure network of the present invention;
fig. 2 is a block diagram of the generation-based countermeasure network of the present invention.
Detailed Description
In order to more clearly explain the problems to be solved, the technical implementation processes and the points of the present invention, the implementation processes will be described in detail below by combining the accompanying drawings.
The invention discloses a traffic speed prediction method based on a space-time graph convolution-generation countermeasure network, which comprises the following steps:
step 1: constructing a traffic graph network adjacency matrix;
the traffic graph network is constructed by taking road sections in the road network as nodes and taking intersections as edges connecting the nodes, and the traffic network is expressed as follows:
G＝{V,E,A} (1)
wherein V ═ { V ═ V1,V2,…,VnThe node set in the traffic graph network is represented, the number of the nodes is n, E
Representing a set of connected edges of a network of traffic maps, A being an n × n symmetric adjacency matrix and being Ai,j＝Aj,iThe rules are defined as follows:
Figure BDA0003052364830000031
processing the adjacent matrix acquired in the last step:
calculating a Laplace matrix L ═ D-A of the adjacency matrix, where D is the degree matrix of A and is the diagonal matrix, Di,iFor the degree of the ith node, the degree matrix calculation formula is as follows:
Figure BDA0003052364830000032
the values of the adjacency matrix represent the weights of the adjacency matrix, and in order to avoid neglecting the characteristics of each node in the graph convolution process, an identity matrix is added to the adjacency matrix, so that the diagonal elements of the adjacency matrix become 1.
The laplacian matrix is normalized, so that the data can be limited to a range required by us. By the normalization process, the data can be made comparable while maintaining the relationship between the data. The normalized laplacian matrix is defined as:
Figure BDA0003052364830000033
wherein In∈Rn×nIs a unit array.
Decomposing the normalized Laplace matrix L to obtain L ═ UΛ UTWherein
Figure BDA0003052364830000034
Representing n mutually orthogonal featuresThe vector of the vector is then calculated,
Figure BDA0003052364830000035
is a diagonal matrix, λiRepresents uiThe corresponding characteristic value.
Step 2: acquiring and processing road network traffic flow data;
traffic data is collected through detectors in a road network, the sampling rate is 5 minutes once, and the sampling times in one day are 288. And cleaning traffic data, removing redundant data and error data, and completing the missing phenomenon of a road section at a certain moment by adopting a linear interpolation method, as shown in the following figure:
Figure BDA0003052364830000041
where x denotes the time of the data missing point, y0Last time x when x is represented0Traffic speed value of y1The next time x representing the x time1Y represents the result obtained by linear interpolation fitting.
And step 3: constructing a generator that generates a countermeasure network (GAN);
the invention adopts a space-time graph convolutional network (STGCN) as a generator for generating a countermeasure network, and extracts spatial characteristics and time domain characteristics of traffic speed data;
the method for extracting the features by using the STGCN comprises the following steps:
the framework of STGCN consists of two space-time convolution modules and an output layer.
Space-time volume block:
each space-time convolution module consists of two gating sequence convolution layers (Temporal Gated-Conv) and a space diagram convolution layer (Spatial Graph-Conv) in the middle, wherein the gating sequence convolution layers are used for acquiring the time relevance of traffic data; the space convolution layer is used for acquiring the space relevance of the traffic speed data;
the gating sequence convolution layer is composed of a one-dimensional convolution neural network (1-D Conv) and a Gating Linear Unit (GLU), and the gating sequence convolution layer conducts one-dimensional convolution on traffic characteristic data in a time dimension so as to extract traffic speed time domain characteristics.
The gating sequence convolution layer is shown at the far right of FIG. 1, with the input to each node being
Figure BDA0003052364830000042
One-dimensional convolution along the time dimension with a convolution kernel of
Figure BDA0003052364830000043
The number of convolution kernels is 2CoIn which C isiAnd CoIs the size of the input and the extracted feature map, the input contains the channel map of M frames, thus two outputs can be obtained
Figure BDA0003052364830000044
Figure BDA0003052364830000045
Then activated by gating the linear unit:
Figure BDA0003052364830000046
for a complete space-time graph convolutional network, each frame can be represented by a matrix, the dimension is n × C, n represents n nodes, and C represents a characteristic dimension: input device
Figure BDA0003052364830000047
Output of
Figure BDA0003052364830000048
The space convolution layer uses Chebyshev graph convolution, carries out high-order feature extraction on graph structure data in a space domain, and uses Chebyshev polynomial approximation, wherein the Chebyshev graph convolution formula is as follows:
Figure BDA0003052364830000049
where K is the size of the convolution kernel, TkIs a Chebyshev polynomial obtained by recursion, and the recursion expression is as follows:
Tk(x)＝2xTk-1(x)-Tk-2(x) (8)
wherein T is0(x)＝1，T1(x)＝x。
θkIs a polynomial coefficient and needs to be obtained by learning,
Figure BDA00030523648300000410
the hierarchical linear formula may be defined by superimposing multiple localized graph convolution layers using a first order approximation of the laplace map. Thus, deeper architectures can be constructed to recover spatial information deeply without the constraints of explicit parameterization given by the polynomial. Due to the scaling and normalization of the neural network, we can further assume λmaxAnd is approximately equal to 2. Therefore, equation (7) can be simplified as:
Figure BDA0003052364830000051
wherein theta is0，θ1Two shared parameters of convolution kernel, where one parameter theta is used instead for constraint and stability improvement0＝-θ1And then define respectively
Figure BDA0003052364830000052
Equation (9) can then be further simplified as:
Figure BDA0003052364830000053
the structure of the mth layer of the chebyshev network is defined as follows:
Figure BDA0003052364830000054
the final graph convolution with data for C channels can be expressed as:
Figure BDA0003052364830000055
according to the one-dimensional convolution of the time domain, the length of the data in the time dimension is reduced by 2 (K) every time a space-time convolution block passest-1), after passing through the successive space-time convolution layers, outputting
Figure BDA0003052364830000056
An output layer:
after passing through the first two time-space convolution blocks, the data finally enters an output layer, the output layer comprises a time-domain convolution layer and a full-connection layer, and the convolution kernel of the time-domain convolution layer has the size of
Figure BDA0003052364830000057
The number is CoMapping the output to
Figure BDA0003052364830000058
Full connection layer
Figure BDA0003052364830000059
Wherein
Figure BDA00030523648300000510
Final output
Figure BDA00030523648300000511
Figure BDA00030523648300000512
After the generator generates data, the loss function G of the generator is passedlossCalculating an average difference value between the generated data and the real data, and expecting GlossAs small as possible, GlossThe definition is as follows:
Figure BDA00030523648300000513
wherein WθIs all of the parameters that can be trained and,
Figure BDA00030523648300000514
is the data generated by the generator, v is the real traffic speed data, where the data of successive M times t-M +1, …, t are input, and the data of t + epsilon is compared.
Step 4, constructing a discriminator for generating a countermeasure network (GAN);
the discriminator for generating the countermeasure network is mainly used for discriminating the true and false of the data generated by the generator, in the invention, the discriminator adopts a fully connected neural network with three layers to discriminate the true and false probability of the output traffic speed data, and when the input of the discriminator is real data, the output of the discriminator is as large as possible; when the input is false data, the input is as small as possible, and the discriminator discrimination process is expressed as follows:
Figure BDA00030523648300000515
in the generation of the countermeasure network, the use of the generator is divided into two processes, firstly, real data and generated data are required to be used as input to train the generator, the characteristics of the real data can be identified, and the output loss D of the training of the discriminator is required to be causedlossAs small as possible, the loss function is defined as follows:
Figure BDA0003052364830000061
wherein D is the output score when the discriminator outputs true and false data, namely the true and false probability;
and 5: generating traffic speed data through the constructed generation countermeasure network;
and continuously competing to generate predicted data by generating a countermeasure network according to the input historical traffic speed data, finally generating traffic data which is very close to real data, and meanwhile, calculating and evaluating the error between the generated data and the real data by using the average absolute error MAE, the root mean square error RMSE and the average absolute percentage error MAPE.
The experimental process of the invention is as follows:
(1) experimental background and data selection
The model of the invention was validated by using a highway traffic data set, pemds 7, in california, collected by the Caltrans performance measurement system (PeMS) in real time every 30 seconds, with speed data aggregated from the raw data every 5 minutes. The system deploys more than 39000 detectors on the expressway of the major city of california, and the geographic information of the sensor stations is recorded in a data set. Traffic speeds of 44 days from 5 months to 6 months of 2012 were chosen as experimental data.
(2) Parameter determination
Constructing a generator for generating the countermeasure network in the step 3: inputting characteristic dimension M of gating sequence convolution layer data in the time-space convolution block to be 12, and predicting data in generator loss function calculation at moment epsilon to be 3, namely predicting traffic speed data after 15 minutes by 12 historical data in one hour; one-dimensional convolution kernel dimension Kt3; the size K of the space convolution layer Chebyshev convolution kernel is 2;
and 4, constructing a discriminator for generating the countermeasure network in the step 4: the discriminator adopts 3 layers of full-connected Bingyun networks, and the number of nodes is respectively 128, 32 and 1.
And 5, generating a process of generating the traffic data by the confrontation network training: the generator iterates for 1 time every 4 times of iteration of the discriminator; in model training, an AdamW optimizer is adopted as an optimizer, the learning rate is 8e-5, the iteration times are 1000, and the parameter weight of the fully-connected neural network of the discriminator is limited to +/-0.03 through cutting;
(3) results of the experiment
And judging the error magnitude of the data obtained by model prediction and the real data by using the average absolute error MAE, the root mean square error RMSE and the average absolute percentage error MAPE. On the California highway data set, the predicted belt traffic speed data and the real data of the present invention had RMSE values of 4.05, MAE values of 2.28, and MAPE values of 5.46.

Claims (1)

1. A traffic speed prediction method based on a space-time graph convolution-generation countermeasure network comprises the following steps:
step 1: constructing a traffic graph network adjacency matrix;
the traffic graph network is constructed by taking road sections in the road network as nodes and taking intersections as edges connecting the nodes, and the traffic network is expressed as follows:
G＝{V,E,A} (1)
wherein V ═ { V ═ V1,V2,…,VnThe node set in the traffic graph network is represented, the number of the nodes is n, E represents the connection edge set of the traffic graph network, A is an n multiplied by n symmetric adjacent matrix and is Ai,j＝Aj,iThe rules are defined as follows:
Figure FDA0003052364820000011
processing the adjacent matrix obtained in the previous step:
calculating a Laplace matrix L ═ D-A of the adjacency matrix, wherein D is a degree matrix of A and is a diagonal matrix, Di,iFor the degree of the ith node, the degree matrix calculation formula is as follows:
Figure FDA0003052364820000012
the value of the adjacency matrix represents the weight of the adjacency matrix, and an identity matrix is added to the adjacency matrix so that the diagonal element of the adjacency matrix becomes 1.
Normalizing the Laplace matrix, wherein the normalized Laplace matrix is defined as:
Figure FDA0003052364820000013
wherein In∈Rn×nIs a unit array.
Decomposing the normalized Laplace matrix L to obtain L ═ UΛ UTWherein
Figure FDA0003052364820000014
Representing n mutually orthogonal eigenvectors,
Figure FDA0003052364820000015
is a diagonal matrix, λiRepresents uiThe corresponding characteristic value.
Step 2: acquiring and processing road network traffic flow data;
collecting traffic data, cleaning the traffic data, removing redundant data and error data, and completing the missing phenomenon of a road section at a certain moment by adopting a linear interpolation method, wherein the method comprises the following steps:
Figure FDA0003052364820000016
where x denotes the time of the data missing point, y0Last time x when x is represented0Traffic speed value of y1The next time x representing the x time1Y represents the result obtained by linear interpolation fitting.
And step 3: constructing a generator that generates a countermeasure network (GAN);
adopting a space-time graph convolutional network (STGCN) as a generator for generating a countermeasure network, and extracting spatial features and time domain features of traffic speed data;
the method for extracting the features by using the STGCN comprises the following steps:
the framework of STGCN consists of two space-time convolution modules and an output layer.
Space-time volume block:
each space-time convolution module consists of two gating sequence convolution layers (Temporal Gated-Conv) and a space diagram convolution layer (Spatial Graph-Conv) in the middle, wherein the gating sequence convolution layers are used for acquiring the time relevance of traffic data; the space convolution layer is used for acquiring the space relevance of the traffic speed data;
the gating sequence convolution layer is composed of a one-dimensional convolution neural network (1-D Conv) and a Gating Linear Unit (GLU), and the gating sequence convolution layer conducts one-dimensional convolution on traffic characteristic data in a time dimension so as to extract traffic speed time domain characteristics.
The gating sequence convolution layer is shown at the far right of FIG. 1, with the input to each node being
Figure FDA0003052364820000021
One-dimensional convolution along the time dimension with a convolution kernel of
Figure FDA0003052364820000022
The number of convolution kernels is 2CoIn which C isiAnd CoIs the size of the input and the extracted feature map, the input contains the channel map of M frames, thus two outputs can be obtained
Figure FDA00030523648200000214
Figure FDA0003052364820000023
Then activated by gating the linear unit:
Figure FDA0003052364820000024
for a complete space-time graph convolutional network, each frame can be represented by a matrix, the dimension is n × C, n represents n nodes, and C represents a characteristic dimension: input device
Figure FDA0003052364820000025
Output of
Figure FDA0003052364820000026
The space convolution layer uses Chebyshev graph convolution, carries out high-order feature extraction on graph structure data in a space domain, and uses Chebyshev polynomial approximation, wherein the Chebyshev graph convolution formula is as follows:
Figure FDA0003052364820000027
where K is the size of the convolution kernel, TkIs a Chebyshev polynomial obtained by recursion, and the recursion expression is as follows:
Tk(x)＝2xTk-1(x)-Tk-2(x) (8)
wherein T is0(x)＝1，T1(x)＝x。
θkIs a polynomial coefficient and needs to be obtained by learning,
Figure FDA0003052364820000028
the hierarchical linear formula may be defined by superimposing multiple localized graph convolution layers using a first order approximation of the laplace map. Thus, deeper architectures can be constructed to recover spatial information deeply without the constraints of explicit parameterization given by the polynomial. Due to the scaling and normalization of the neural network, we can further assume λmaxAnd is approximately equal to 2. Therefore, equation (7) can be simplified as:
Figure FDA0003052364820000029
wherein theta is0，θ1Two shared parameters of convolution kernel, where one parameter theta is used instead for constraint and stability improvement0＝-θ1And then define respectively
Figure FDA00030523648200000210
Equation (9) can then be further simplified as:
Figure FDA00030523648200000211
the structure of the mth layer of the chebyshev network is defined as follows:
Figure FDA00030523648200000212
the final graph convolution with data for C channels can be expressed as:
Figure FDA00030523648200000213
according to the one-dimensional convolution of the time domain, the length of the data in the time dimension is reduced by 2 (K) every time a space-time convolution block passest-1), after passing through the successive space-time convolution layers, outputting
Figure FDA0003052364820000031
An output layer:
after passing through the first two time-space convolution blocks, the data finally enters an output layer, the output layer comprises a time-domain convolution layer and a full-connection layer, and the convolution kernel of the time-domain convolution layer has the size of
Figure FDA0003052364820000032
The number is CoMapping the output to
Figure FDA0003052364820000033
Full connection layer
Figure FDA0003052364820000034
Wherein
Figure FDA0003052364820000035
Final output
Figure FDA0003052364820000036
Figure FDA0003052364820000037
After the generator generates data, the loss function G of the generator is passedlossCalculating an average difference value between the generated data and the real data, and expecting GlossAs small as possible, GlossThe definition is as follows:
Figure FDA0003052364820000038
wherein WθIs all of the parameters that can be trained and,
Figure FDA0003052364820000039
is the data generated by the generator, v is the real traffic speed data, where the data of successive M times t-M +1, …, t are input, and the data of t + epsilon is compared.
Step 4, constructing a discriminator for generating a countermeasure network (GAN);
the discriminator for generating the countermeasure network is used for discriminating the true and false of the data generated by the generator, the discriminator adopts a fully-connected neural network with three layers to discriminate the true and false probability of the output traffic speed data, and when the input of the discriminator is real data, the output of the discriminator is as large as possible; when the input is false data, the input is as small as possible, and the discriminator discrimination process is expressed as follows:
Figure FDA00030523648200000310
in the generation of the countermeasure network, the use of the generator is divided into two processes, firstly, the real data and the generated data are used as input to train the generator, the characteristics of the real data can be identified, and the output loss D of the training of the discriminator is requiredlossAs small as possible, the loss function is defined as follows:
Figure FDA00030523648200000311
CN202110491442.0A 2021-05-06 2021-05-06 Traffic speed prediction method based on space-time graph convolution-generation countermeasure network Active CN113450561B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110491442.0A CN113450561B (en) 2021-05-06 2021-05-06 Traffic speed prediction method based on space-time graph convolution-generation countermeasure network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110491442.0A CN113450561B (en) 2021-05-06 2021-05-06 Traffic speed prediction method based on space-time graph convolution-generation countermeasure network

Publications (2)

Publication Number Publication Date
CN113450561A true CN113450561A (en) 2021-09-28
CN113450561B CN113450561B (en) 2022-05-24

Family

ID=77809667

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110491442.0A Active CN113450561B (en) 2021-05-06 2021-05-06 Traffic speed prediction method based on space-time graph convolution-generation countermeasure network

Country Status (1)

Country Link
CN (1) CN113450561B (en)

Cited By (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114492886A (en) * 2022-01-25 2022-05-13 四川大学 Network taxi appointment demand prediction method and system considering departure place and destination
CN114511644A (en) * 2022-01-21 2022-05-17 电子科技大学 Self-adaptive digital camouflage method based on deep learning
CN114519932A (en) * 2022-01-10 2022-05-20 中国科学院深圳先进技术研究院 Regional traffic condition integrated prediction method based on space-time relation extraction
CN114566047A (en) * 2022-03-02 2022-05-31 中远海运科技股份有限公司 Early warning method and system based on short-time circuit condition prediction algorithm
CN114626585A (en) * 2022-02-28 2022-06-14 北京交通大学 Urban rail transit short-time passenger flow prediction method based on generation of countermeasure network
CN114664091A (en) * 2022-04-26 2022-06-24 中远海运科技股份有限公司 Early warning method and system based on holiday traffic prediction algorithm
CN114944053A (en) * 2022-03-16 2022-08-26 浙江工业大学 Traffic flow prediction method based on spatio-temporal hypergraph neural network
CN115018073A (en) * 2022-08-09 2022-09-06 之江实验室 Time-space perception information prediction method and system based on graph neural network
CN117935561A (en) * 2024-03-20 2024-04-26 山东万博科技股份有限公司 Intelligent traffic flow analysis method based on Beidou data

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110827546A (en) * 2019-11-21 2020-02-21 银江股份有限公司 Signalized intersection road section short-term speed prediction method
CN112330952A (en) * 2020-09-14 2021-02-05 浙江工业大学 Traffic flow prediction method based on generating type countermeasure network

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110827546A (en) * 2019-11-21 2020-02-21 银江股份有限公司 Signalized intersection road section short-term speed prediction method
CN112330952A (en) * 2020-09-14 2021-02-05 浙江工业大学 Traffic flow prediction method based on generating type countermeasure network

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
王一鸣: "面向交通流预测的深度神经网络建模与应用研究", 《中国优秀博硕士学位论文全文数据库(硕士) 工程科技Ⅱ辑》 *
郭海锋 等: "基于改进ACGAN算法的车道排队车辆估计及其分类", 《高技术通讯》 *

Cited By (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114519932A (en) * 2022-01-10 2022-05-20 中国科学院深圳先进技术研究院 Regional traffic condition integrated prediction method based on space-time relation extraction
CN114519932B (en) * 2022-01-10 2023-06-20 中国科学院深圳先进技术研究院 Regional traffic condition integrated prediction method based on space-time relation extraction
CN114511644A (en) * 2022-01-21 2022-05-17 电子科技大学 Self-adaptive digital camouflage method based on deep learning
CN114492886B (en) * 2022-01-25 2022-09-13 四川大学 Network taxi appointment demand prediction method and system considering departure place and destination
CN114492886A (en) * 2022-01-25 2022-05-13 四川大学 Network taxi appointment demand prediction method and system considering departure place and destination
CN114626585A (en) * 2022-02-28 2022-06-14 北京交通大学 Urban rail transit short-time passenger flow prediction method based on generation of countermeasure network
CN114626585B (en) * 2022-02-28 2023-09-08 北京交通大学 Urban rail transit short-time passenger flow prediction method based on generation countermeasure network
CN114566047A (en) * 2022-03-02 2022-05-31 中远海运科技股份有限公司 Early warning method and system based on short-time circuit condition prediction algorithm
CN114944053A (en) * 2022-03-16 2022-08-26 浙江工业大学 Traffic flow prediction method based on spatio-temporal hypergraph neural network
CN114664091A (en) * 2022-04-26 2022-06-24 中远海运科技股份有限公司 Early warning method and system based on holiday traffic prediction algorithm
CN115018073A (en) * 2022-08-09 2022-09-06 之江实验室 Time-space perception information prediction method and system based on graph neural network
CN117935561A (en) * 2024-03-20 2024-04-26 山东万博科技股份有限公司 Intelligent traffic flow analysis method based on Beidou data
CN117935561B (en) * 2024-03-20 2024-05-31 山东万博科技股份有限公司 Intelligent traffic flow analysis method based on Beidou data

Also Published As

Publication number Publication date
CN113450561B (en) 2022-05-24

Similar Documents

Publication Publication Date Title
CN113450561B (en) Traffic speed prediction method based on space-time graph convolution-generation countermeasure network
Chen et al. Traffic flow prediction based on deep learning in internet of vehicles
CN111612206B (en) Neighborhood people stream prediction method and system based on space-time diagram convolution neural network
Qin et al. A novel combined prediction scheme based on CNN and LSTM for urban PM 2.5 concentration
CN111223301B (en) Traffic flow prediction method based on graph attention convolution network
CN113313947B (en) Road condition evaluation method of short-term traffic prediction graph convolution network
CN106251625B (en) Three-dimensional urban road network global state prediction technique under big data environment
Kim et al. Urban traffic flow prediction system using a multifactor pattern recognition model
CN112863182B (en) Cross-modal data prediction method based on transfer learning
CN114529081B (en) Space-time combined traffic flow prediction method and device
CN114202120A (en) Urban traffic travel time prediction method aiming at multi-source heterogeneous data
CN115565369B (en) Space-time hypergraph convolution traffic flow prediction method and system based on hypergraph
CN113435658B (en) Traffic flow prediction method based on space-time fusion correlation and attention mechanism
CN113808396B (en) Traffic speed prediction method and system based on traffic flow data fusion
WO2022142418A1 (en) Traffic performance index prediction method and device based on gis map information
CN113823090A (en) Traffic flow prediction method based on multi-space-time graph convolution network
WO2024068527A1 (en) A computer-implemented method for forecasting traffic indicators in a road network
CN112101132A (en) Traffic condition prediction method based on graph embedding model and metric learning
Pan et al. Traffic speed prediction based on time classification in combination with spatial graph convolutional network
CN116740959A (en) Expressway flow prediction method and system based on multi-source data fusion
CN116386020A (en) Method and system for predicting exit flow of highway toll station by multi-source data fusion
CN115456238A (en) Urban trip demand prediction method based on dynamic multi-view coupling graph convolution
CN117037461A (en) Road network traffic jam prediction method based on multi-weight graph three-dimensional convolution
CN115204445A (en) Travel demand prediction method and system for interaction between urban functional areas
CN111966770A (en) Urban street function identification method and system based on geographic semantic word embedding

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