CN112216108A

CN112216108A - Traffic prediction method based on attribute-enhanced space-time graph convolution model

Info

Publication number: CN112216108A
Application number: CN202011082028.6A
Authority: CN
Inventors: 朱佳玮; 李海峰; 赵玲; 黄浩哲; 彭剑; 陈力; 崔振琦
Original assignee: Central South University
Current assignee: Central South University
Priority date: 2020-10-12
Filing date: 2020-10-12
Publication date: 2021-01-12
Anticipated expiration: 2040-10-12
Also published as: CN112216108B

Abstract

The invention discloses a traffic prediction method based on an attribute-enhanced space-time graph convolution model, which comprises the following steps: constructing an adjacency matrix A based on road network data; establishing an attribute enhancement matrix K at the time t based on the feature matrix X, the interest point information vector p and the weather information matrix B^t＝[X^t,p,B^t](ii) a Inputting the attribute enhancement matrix of historical n moments and the adjacent matrix of the road network into a blank graph convolution model for learning and training, and calculating intersectionAnd obtaining a traffic predicted value through the through-flow hidden state. According to the method, on the basis of modeling the space-time characteristics by using the space-time graph convolution model, the multi-source fragmented urban data are fused to capture the relation between external factors influencing traffic and traffic flow, the perception of the space-time graph convolution model on the external factors is enhanced, and therefore more efficient and accurate traffic prediction is achieved.

Description

Traffic prediction method based on attribute-enhanced space-time graph convolution model

Technical Field

The invention relates to the technical field of intelligent traffic, in particular to a traffic prediction method based on an attribute-enhanced space-time graph convolution model.

Background

In recent years, with the rapid development of urban traffic roads, the problem of imbalance among people, vehicles and roads is increasingly prominent, so that the traffic problem brings great inconvenience to life and work of people, and even seriously affects the life quality of urban residents. Intelligent transportation systems can be seen as an important approach to solve a range of urban traffic problems. As one of important components of an intelligent traffic system, traffic flow prediction can provide scientific basis for management and control and planning of an urban traffic system, and is one of key technologies for building a traffic guidance information platform and a traffic guidance service platform. According to the predicted traffic flow condition, traffic departments can deploy and guide traffic flow in advance, and traffic services such as route planning and navigation can realize more scientific route decision, so that the running efficiency of a road network is improved, and traffic jam is relieved.

The task of traffic prediction is to predict traffic conditions over a future period of time based on historical traffic information. At present, a plurality of traffic prediction methods with higher prediction precision emerge, and the methods can be mainly divided into methods based on model driving and methods based on data driving. The traffic problem is predicted by using mathematical formulas such as differential equations and the like and physical knowledge through calculation simulation, the modeling process is complicated, the parameter calibration difficulty is high, a large amount of systematic priori knowledge and calculation force support are needed, and a model based on unreasonable hypothesis can generate large prediction deviation. The latter can be classified into classical statistical models and machine learning models. Among them, the conventional statistical methods such as autoregressive moving average model (ARIMA) and seasonal autoregressive moving average model (SARIMA) are linear parameter estimation models, limited by the stable assumption of time series, and cannot capture the complex pattern and characteristics of nonlinear traffic flow data, so that it is difficult to accurately describe the change of traffic flow, and the prediction of large-scale traffic is not accurate enough, and thus they are gradually replaced by machine learning models. In the traditional machine learning algorithm, a K-nearest neighbor algorithm (KNN) is firstly applied to the prediction of traffic flow, then a Bayesian inference method is used for parameter estimation in research, and a Markov Chain Monte Carlo (MCMC) simulation is utilized to obtain an optimal model. Combined adaptive particle swarm optimization algorithms and Support Vector Machine (SVM) models have also been investigated to achieve short-term traffic flow predictions. These algorithms can model more complex data, but have limited improvement in prediction accuracy. With the rise of deep learning in recent years, many deep learning methods are applied to traffic flow prediction, such as Deep Belief Network (DBN), stacked self-coding neural network (SAE), and the like. However, these methods independently process traffic flow information at each time, and do not directly model changes in the traffic flow in time series. Therefore, studies have been made to predict traffic flow using a Recurrent Neural Network (RNN) based on sequence prediction, but the recurrent neural network has limited short-term memory, and for longer-term sequence memory, the memory fails due to gradient disappearance, so researchers have adopted improved recurrent neural networks, such as long short-term memory network (LSTM), gated cyclic unit (GRU) models, to extract temporal features of traffic flow data. In order to better represent the spatial dependence of traffic flow, a plurality of researches improve on the basis, and a Convolutional Neural Network (CNN) is introduced to extract spatial information and is combined with a long-term and short-term memory network (LSTM), so that the prediction precision is improved to a certain extent. However, because the CNN is substantially suitable for euclidean space such as images and grids, there is a limitation on traffic networks having non-european topological structures, and the spatial dependence characteristic of traffic flow cannot be substantially described. And an emerging graph convolution neural network (GCN) is specially used for processing a network structure and can better model the spatial dependence of road sections on a traffic network.

In addition to temporal-spatial correlation, however, traffic is also susceptible to a number of external factors, such as weather factors and distribution of surrounding roadway facilities. Therefore, when the traffic characteristic space-time factors are considered, how to capture the influence of various external factors on the trip is an important step for realizing accurate prediction of traffic conditions. Although the traffic flow prediction method based on the deep neural network can solve the limitation of the traditional method in traffic flow prediction and enhance the comprehension capability of the space-time characteristics of traffic flow data, the method has limitation in comprehensively considering various influence factors.

Disclosure of Invention

In view of the above, the present invention provides a traffic prediction method based on an attribute-enhanced space-time graph convolution model, which, on the basis of using a space-time graph convolution model to model space-time features, fuses multisource fragmented urban data to capture a relationship between external factors affecting traffic and traffic flow, enhances perception of the space-time graph convolution model on the external factors, and further implements more efficient and accurate traffic prediction.

The invention aims to realize the method for predicting the traffic based on the attribute-enhanced space-time graph convolution model, which comprises the following steps:

step 1, constructing an adjacency matrix A based on road network data; modeling a road network as an unweighted graph G ═ (V, E), where V ═ V₁,v₂,...,v_NRepresents a set of nodes v representing road segments in the road, N representing the number of road segments in the road; e ═ E₁,e₂,...,e_MRepresenting a set of edges e, e representing an edge connecting two road segments, M representing the number of edges, creating an adjacency matrix A e R of the road network based on the graph G^N×NDescribing the connection relation between road sections, wherein the value 1 in the matrix A represents that two road sections represented by corresponding ranks are connected, and the value 0 represents that the two road sections are not connected;

step 2, constructing an attribute enhancement matrix K at the time t based on the feature matrix X, the interest point information vector p and the weather information matrix B^t＝[X^t,p,B^t]Size is Nx 3, X^tThe traffic speed at the current moment is used as a list of characteristics; b is^tIndicating that the weather at the current moment is selected as a list of characteristics; the interest point information belongs to static characteristics, and the same p vector is used for enhancement matrixes at different moments;

and 3, inputting the attribute enhancement matrix of the historical n moments and the adjacent matrix of the road network into an empty graph convolution model for learning and training, calculating the traffic flow hidden state and obtaining a traffic predicted value.

Specifically, step 3 comprises the following steps:

301, connecting the adjacent matrix A of road network and the enhancement matrix K at a certain time^tThe method is used as the input of a graph convolution model, the features of capturing road network topology and enhancing attributes are convoluted by means of a graph theory, and then the representation of road section nodes at the moment t is obtained, and the calculation formula is as follows:

wherein

I_NIs an N x N identity matrix, in order to add self-connection to the adjacency matrix,

is that

Degree matrix of (Relu) () a modified linear element, W₀And W₁Two weight matrixes are obtained through back propagation learning and are obtained through n graph convolution models and n enhanced matrixes K at different moments^t-n+1,...,K^t-1,K^tThe time series y with spatial information can be obtained as input^t-n+1,...,y^t-1,y^t；

Step 302, modeling the time dependence of traffic flow, and comparing y^t-n+1,...,y^t-1,y^tInputting n gate control circulation units connected in series, learning dynamic change of traffic flow through transfer among the gate control circulation units, outputting a traffic flow hidden state at a corresponding moment by each gate control circulation unit, wherein each gate control circulation unit consists of a reset gate and an update gate, and the specific calculation is as follows; r is_t＝σ(W_r·[y^t,h_t-1]+b_r)，c_t＝tanh(W_c·[y^t,(r_t,h_t-1)]+b_c)，u_t＝σ(W_u·[y^t,h_t-1]+b_u)，h_t＝u_t*h_t-1+(1-u_t)*c_t，h_t-1Representing a hidden state of traffic flow at the previous moment, σ being an activation function, W and b including W_r，b_r，W_c，b_c，W_uAnd b_uAll are parameters of the model available for training, r_tFor resetting gate, for deciding how much past hidden state information of traffic flow is ignored, based on the reset gate, hidden state c of traffic flow at current time can be calculated_t；u_tFor updating the gate, the hidden state of the traffic flow at the previous moment can be selected and output to the current moment, then the hidden state of the traffic flow at the current moment is updated based on the updating gate, and the final hidden state h of the traffic flow at the current moment is output_t，y^t-n+1,...,y^t-1,y^tH can be obtained after inputting n gate control circulation units connected in series_t-n+1,…,h_t-1,h_t；

Step 303, mixing h_tAnd inputting the linear regression model to obtain a traffic predicted value.

Specifically, the feature matrix X in step 2 is constructed based on a taxi track data set; taking the traffic speed as the inherent characteristic attribute of each road section on the road network, and using a characteristic matrix X to belong to R^N×SExpressing that S represents the number of divided time segments, N represents the number of road segments in the road, each row represents the traffic speed change of the road segments in different time segments, and X represents the traffic speed change of the road segments in different time segments^sIs the s-th column in the matrix X, which represents the traffic speed situation on each road section at the s-th moment;

the interest point information vector p in the step 2 is constructed based on the interest point data set; the interest point data set is used as auxiliary feature information of the road section, if the interest point categories share L categories, the categories are correspondingly numbered 0, 1, 2, … and L-1, the number of various interest points contained in each road section is counted, the number of the interest point category with the largest proportion is used as the feature of the road section, and an interest point information vector p with the size of Nx 1 can be constructed;

the weather information matrix B in the step 2 is constructed based on a weather data set; s represents the number of divided time periods, N represents the number of road sections in a road, the size of a constructed weather information matrix B is NxS, the change of weather on the road sections along with time is described, and the weather information matrix belongs to dynamic external factors and serves as auxiliary characteristic information of the road sections; wherein the weather conditions can be classified into J types, which are respectively represented by the values 0, 1, 2, … and J-1 in the matrix, B^sIs the s-th column in the matrix B, which represents the weather conditions on each link at time s.

Compared with the prior art, the method has the advantages that: the invention provides a traffic prediction method of an attribute-enhanced space-time diagram convolution model, aiming at the problem that external factors influencing traffic conditions cannot be comprehensively considered in the traditional urban traffic prediction model. On one hand, the method provided by the invention overcomes the defect that the existing model can not fuse multi-source data; on the other hand, the method provided by the invention can not only fuse static external factor data, but also fuse dynamic external factor data, so that traffic prediction is more accurate.

Drawings

FIG. 1 shows a schematic flow diagram of an embodiment of the invention;

fig. 2 shows a schematic diagram of the prediction model proposed by the present invention.

Detailed Description

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

It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

As shown in fig. 1, a traffic prediction method based on attribute-enhanced space-time graph convolution model includes the following steps:

The traffic prediction problem can be understood as learning a function f based on the basic topology a and the enhanced matrix K of the urban road network to obtain traffic information in a future period of time T, as shown in the following formula: [ x ] of_t+1,x_t+2,...,x_t+T]＝f(A,(K^t-n+1,…,K^t-1,K^t) That is, an attribute enhancement matrix based on n historical time instants, and an urban road networkThe adjacency matrix a predicts traffic at T moments in the future.

Specifically, step 3 comprises the following steps:

wherein

is that

Step 302, modeling the time dependence of traffic flow, and comparing y^t-n+1,...,y^t-1,y^tInputting n gate control circulation units connected in series, learning dynamic change of traffic flow through transfer among the gate control circulation units, outputting a traffic flow hidden state at a corresponding moment by each gate control circulation unit, wherein each gate control circulation unit consists of a reset gate and an update gate, and the specific calculation is as follows; r is_t＝σ(W_r·[y^t,h_t-1]+b_r)，c_t＝tanh(W_c·[y^t,(r_t,h_t-1)]+b_c)，u_t＝σ(W_u·[y^t,h_t-1]+b_u)，h_t＝u_t*h_t-1+(1-u_t)*c_t，h_t-1Representing the hidden state of the traffic flow at the previous moment, σ is the activation function, W and b are parameters of the model available for training, r_tFor resetting gate, for deciding how much past hidden state information of traffic flow is ignored, based on the reset gate, hidden state c of traffic flow at current time can be calculated_t；u_tFor updating the gate, the hidden state of the traffic flow at the previous moment can be selected and output to the current moment, then the hidden state of the traffic flow at the current moment is updated based on the updating gate, and the final hidden state h of the traffic flow at the current moment is output_t，y^t-n+1,...,y^t-1,y^tH can be obtained after inputting n gate control circulation units connected in series_t-n+1,...,h_t-1,h_t；

the weather information matrix B in the step 2 is based on the weather numberConstructing a data set; s represents the number of divided time periods, N represents the number of road sections in a road, the size of a constructed weather information matrix B is NxS, the change of weather on the road sections along with time is described, and the weather information matrix belongs to dynamic external factors and serves as auxiliary characteristic information of the road sections; wherein the weather conditions can be classified into J types, which are respectively represented by the values 0, 1, 2, … and J-1 in the matrix, B^sIs the s-th column in the matrix B, which represents the weather conditions on each link at time s.

In the embodiment of the invention, the lake region of Shenzhen city is taken as a research region, and the data set is selected to include taxi track data, road network data, interest point data and weather data in the research region.

Road network data: selecting 156 main road sections in a research area as research objects, constructing an adjacent matrix G, wherein the size of the adjacent matrix G is 156 × 156, values in the matrix are used for describing the communication relation between roads, wherein 1 represents that two road sections represented by corresponding lines are communicated, and 0 represents that the two road sections are not communicated;

taxi track data set: this data set contains taxi track data for days 1 month 1 through 31 months 2015. And describing speed change on the road section according to the data set feature matrix, wherein columns of the matrix represent the road sections, rows represent traffic speeds on the road sections in different time periods, 1 day is divided into 15 minutes, 31 days can be divided into 31 times (24 times 60/15) and 2976 time periods, and therefore the size of the final feature matrix X is 156 times 2976.

The point of interest data set: in this embodiment, the interest points are 9 categories, which are catering services, company enterprises, shopping services, transportation facility services, education services, living services, medical services, accommodation services, and others, respectively, the number of each type of interest point included in each road segment is counted, the interest point category with the largest occupation ratio is used as the feature of the road segment, and then an interest point information vector p with a size of 156 × 1 is constructed, and the category to which the main interest point of the road segment belongs is represented by using a value of 0-8 in the vector.

Weather data set: this data set contains the weather conditions of every street every 15 minutes in month 1 of 2015, so the size of the constructed weather information matrix W is 156 × 2976, and the weather information matrix describes the change of the weather on the road section with time. In the embodiment, the weather conditions are divided into five major categories, namely 5 categories including sunny days, cloudy days, fog days, light rain days and heavy rain days, and the values in the matrix are respectively 0-4.

Constructing attribute enhancement matrix K of size 156 x 3 for each time instant^t＝[X^t,p,B^t]And obtaining the attribute enhancement matrix sequence.

The adjacency matrix and the attribute enhancement matrix are input into the traffic prediction model shown in fig. 2, and a prediction result is obtained.

Finally, the predicted results are compared to conventional methods, where y_tFor the real data at the time t,

the prediction result of the model at the time t, n represents the number of predictions, and the evaluation indexes include the following 4.

Root Mean Square Error (RMSE):

accuracy (Accuracy):

wherein | | | purple hair_FIs the Frobenius norm of the matrix (Frobenius norm).

Coefficient of determinability (R)²)：

Interpretation variance (Var):

where var represents the variance.

The four indexes can be used for evaluating the performance of the model. Wherein, the Root Mean Square Error (RMSE) is used for evaluating the prediction error, and the smaller the value is, the smaller the prediction error is, the better the model performance is; the Accuracy (Accuracy) is used for evaluating the Accuracy of model prediction, and the larger the value is, the higher the prediction Accuracy is and the more accurate the prediction is; coefficient of determinability (R)²) And interpreting the variance (Var) to measure the capability of the result obtained by model prediction to represent actual data, wherein the larger the values of the two evaluation indexes are, the better the model prediction result is.

The attribute enhancement based spatio-temporal graph convolution model (AST-GCN) proposed by the present invention is compared to the following baseline method.

(1) Autoregressive integrated moving average model (ARIMA): the method mainly utilizes the time sequence data to simplify a relation model, researches the development rule of the object according to the time sequence characteristics, and further predicts the future change of the object.

(2) Support Vector Regression (SVR): the method mainly obtains a prediction result in a linear or nonlinear mode by learning the relation between input features and labels.

(3) Graph convolution model (GCN): the method mainly assists the neural network to obtain a prediction result by learning the adjacency relation and the characteristics among the networks and further capturing the topological structure of the network.

(4) Gated cycle unit model (GRU): the method belongs to a variant of a Recurrent Neural Network (RNN), and mainly obtains a current state through input features, and obtains a prediction result by capturing the current features and simultaneously keeping the change trend of historical features.

(5) Space-time graph convolution model (T-GCN): the method combines GCN and GRU, captures the spatial relationship of the network and the time variation trend of the characteristics, has the capability of capturing the characteristics of time and space, and further obtains a prediction result.

(6) Diffusion Convolutional Recurrent Neural Network (DCRNN): the spatial data of the graph is modeled by diffusion convolution, and the time dependence of the traffic information is modeled by using the RNN, so that the final result is predicted.

The comparative results are given in the table below, indicating that the values are negative, indicating poor results:

from the results, it can be found that the model performs better than other methods in each index. The RMSE reduction of the present method (AST-GCN) was about 44.19% and 40.49% compared to the SVR and ARIMA models of the conventional methods, respectively. Compared with GCN and GRU models, the RMSE of the method (AST-GCN) is reduced by about 28.58 percent and 20.44 percent respectively from the perspective of space-time comparative analysis, and the overall evaluation effect is obviously improved. From the aspect of attribute auxiliary comparison and analysis, the method (AST-GCN) compares the T-GCN and the DCRNN models which only capture space-time characteristics after adding attribute auxiliary information, the prediction error RMSE is respectively reduced by about 0.98 percent and 10.46 percent, and the prediction result is also improved. From the above results, it can be verified that the model provided by the invention uses the auxiliary information to enhance the effectiveness of the space-time diagram convolution model in sensing the external influence factors, and accurate traffic prediction can be realized.

As will be appreciated by one skilled in the art, embodiments of the present application may be provided as a method, system, or computer program product. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, and the like) having computer-usable program code embodied therein.

Claims

1. A traffic prediction method based on an attribute-enhanced space-time graph convolution model is characterized by comprising the following steps:

step 1, constructing an adjacency matrix A based on road network data; modeling a road network as an unweighted graph G ═ (V, E), where V ═ V₁,v₂,...,v_NRepresents a set of nodes v representing road segments in the road, and N represents the number of road segments in the road; e ═ E₁,e₂,...,e_MRepresenting a set of edges e, e representing an edge connecting two road segments, M representing the number of edges, creating an adjacency matrix A e R of the road network based on the graph G^N×NDescribing the connection relation between road sections, wherein the value 1 in the matrix A represents that two road sections represented by corresponding ranks are connected, and the value 0 represents that the two road sections are not connected;

step 2, constructing an attribute enhancement matrix K at the time t based on the feature matrix X, the interest point information vector p and the weather information matrix B^t＝[X^t,p,B^t]Size is Nx 3, X^tIndicating that the traffic speed at the current moment is selected as a list of characteristics, B^tThe weather at the current moment is used as a column of characteristics, the interest point information belongs to static characteristics, and the same p vector is used for the enhancement matrixes at different moments;

2. The traffic prediction method of claim 1, wherein step 3 comprises the steps of:

step 301, enhancing the adjacent matrix A of the road network and the attribute of the t moment^tThe method is used as the input of a graph convolution model, the features of capturing road network topology and enhancing attributes are convoluted by means of a graph theory, and then the representation of road section nodes at the moment t is obtained, and the calculation formula is as follows:

wherein

is that

Degree matrix of (Relu) () a modified linear element, W₀And W₁Is two weight matrixes, which are obtained by learning through back propagationBy means of n graph convolution models and n enhancement matrices K at different time instants^t-n+1,...,K^t-1,K^tThe time series y with spatial information can be obtained as input^t-n+1,...,y^t-1,y^t；

Step 302, modeling the time dependence of traffic flow, and comparing y^t-n+1,...,y^t-1,y^tInputting n gate control circulation units connected in series, learning dynamic change of traffic flow through transfer among the gate control circulation units, outputting a traffic flow hidden state at a corresponding moment by each gate control circulation unit, wherein each gate control circulation unit consists of a reset gate and an update gate, and the specific calculation is as follows; r is_t＝σ(W_r·[y^t,h_t-1]+b_r)，c_t＝tanh(W_c·[y^t,(r_t,h_t-1)]+b_c)，u_t＝σ(W_u·[y^t,h_t-1]+b_u)，h_t＝u_t*h_t-1+(1-u_t)*c_t，h_t-1Representing the hidden state of the traffic flow at the previous moment, sigma is an activation function, W and b are parameters of a model available for training, r_tFor resetting gate, for deciding how much past hidden state information of traffic flow is ignored, based on the reset gate, hidden state c of traffic flow at current time can be calculated_t；u_tFor updating the gate, the hidden state of the traffic flow at the previous moment can be selected and output to the current moment, then the hidden state of the traffic flow at the current moment is updated based on the updating gate, and the final hidden state h of the traffic flow at the current moment is output_t，y^t-n+1,...,y^t-1,y^tH can be obtained after inputting n gate control circulation units connected in series_t-n+1,...,h_t-1,h_t；

3. The traffic prediction method according to claim 2, wherein the feature matrix X in step 2 is constructed based on a taxi track data set;taking the traffic speed as the inherent characteristic attribute of each road section on the road network, and using a characteristic matrix X to belong to R^N×SExpressing that S represents the number of divided time segments, N represents the number of road segments in the road, each row represents the traffic speed change of the road segments in different time segments, and X represents the traffic speed change of the road segments in different time segments^sIs the s-th column in the matrix X, which represents the traffic speed situation on each road section at the s-th moment;

the weather information matrix B in the step 2 is constructed based on a weather data set; s represents the number of divided time periods, N represents the number of road sections in a road, the size of a constructed weather information matrix B is NxS, the change of weather on the road sections along with time is described, wherein the weather condition can be divided into J types, the weather condition is represented by values 0, 1, 2, … and J-1 in the matrix, and B is represented by values B, 1, 2, … and J-1 respectively^sIs the s-th column in the matrix B, which represents the weather conditions on each link at time s.