CN112466117A

CN112466117A - Road network short-term traffic flow prediction method based on deep space-time residual error network

Info

Publication number: CN112466117A
Application number: CN202011326844.7A
Authority: CN
Inventors: 施佺; 丁新宇; 施振佺; 曹阳; 曹志超; 朱森来
Original assignee: Nantong University
Current assignee: Nantong University
Priority date: 2020-11-24
Filing date: 2020-11-24
Publication date: 2021-03-09

Abstract

The invention provides a road network short-time traffic flow prediction method based on a deep space-time residual error network, which respectively designs corresponding residual error network branches aiming at the adjacency and periodicity of two unique attributes of space-time data, dynamically aggregates the outputs of the two branch networks by distributing different weights to the same roads in the two branches, thereby adjusting the space-time attributes to different branchesAnd secondly, fusing the aggregation results of the two residual error networks with external factors. By selecting RMSE and R²The DST-ResNet model has higher effectiveness and feasibility compared with a mainstream LSTM model by experimental verification of evaluation indexes of the model.

Description

Road network short-term traffic flow prediction method based on deep space-time residual error network

Technical Field

The invention relates to the technical field of road network short-term traffic flow prediction, in particular to a road network short-term traffic flow prediction method based on a deep space-time residual error network.

Background

Short-term traffic flow prediction is a popular research topic in the field of Intelligent Transportation Systems (ITS), and can provide a solid foundation and data support for an Intelligent traffic management system. As an important role in the system, real-time accurate short-time traffic flow prediction is essential for both traffic management departments and travelers. On one hand, the real-time and accurate traffic flow prediction can provide accurate road condition information for travelers, effectively avoid congested road sections and save travel time; on the other hand, the traffic management department can utilize the result of traffic flow prediction to guide traffic in advance, so that some road sections are prevented from being too congested. Therefore, short-time traffic flow prediction has become one of the research hotspots in the traffic field in recent years.

For the short-term traffic flow prediction problem, researchers in different fields at home and abroad respectively start from respective fields and establish more excellent traffic flow theories and models. Chan, K.Y et al use a kalman filtering method to introduce linear system state equations for optimal estimation of the overall state. Fusco et al fused the Bayesian network with the artificial neural network for modeling, and verified the validity of the model using floating car large sample data. YaoZhisheng et al propose a traffic status real-time prediction scheme based on a support vector regression machine. Duaphone et al established a traffic flow prediction model based on a multi-condition random field based on the inherent characteristics of time-varying property, non-linear property and relevance of crossing upstream and downstream in space of traffic data stream. The SARIMA-RF model is proposed by the people of the panda and the like by using the SARIMA model to extract the periodic variation of traffic flow data and combining the strong prediction capability of the random forest model.

With the intensive research of deep learning, more and more deep learning theories and methods are applied to traffic flow prediction. The lobe scene and the like provide a short-term traffic flow prediction method based on CNN-XGboost. And (3) predicting data input of the current road section and the adjacent road section by combining the time correlation and the space correlation of the short-time traffic flow data, and optimizing the CNN model parameters by using a drosophila algorithm. Wenheim et al use LSTM to predict highway traffic flow and optimize the step size of the data time window by genetic algorithm. Yanzhen et al excavate the spatial relevance of the traffic flow at the adjacent crossing through the convolution network (CNN), excavate the time series characteristic of the traffic flow through the LSTM model, and carry out characteristic fusion on the extracted space-time characteristic to realize short-term flow prediction. Guizhiming et al extract the spatio-temporal features of traffic flow using Convolutional Neural Network (CNN) and gated round-robin unit (GRU), and the prediction error is reduced by 9% compared with other models.

However, the existing deep learning method aiming at traffic flow prediction, such as the recurrent neural network (LSTM), has two main disadvantages when solving the problem of urban road network traffic flow prediction with mass data scale. First, input data of the LSTM must be a continuous time series, and if it is desired that the input data contain proximity and periodicity, the input data must be very long, and if only data of the last two hours or days is used as input, the periodicity cannot be embodied. However, if data from the past week or even month is used as the LSTM input, the model becomes very complex and difficult to train. Second, when the LSTM predicts the traffic flow of the road network, the spatial correlation is not considered, and it needs to use the Reshape of a frame of data as a vector, so that the spatial correlation between the road segments is lost.

Through literature analysis, the current road network short-time traffic flow prediction problem has the following difficulties:

(1) spatial dependency. The traffic flow of the road R5 is affected by the flow of vehicles on nearby roads (e.g., R1, R2, R3, R4, R6) and roads in more distant areas, and similarly, the traffic flow of R5 affects the traffic flow of other links.

(2) Time dependence. The current time of day traffic flow in a road segment may be affected by the recent traffic flow for that road segment. For example, 8 a.m. traffic congestion may affect 9 a.m. traffic flow. Traffic conditions during the morning rush hour of the weekday may be similar, repeating every 24 hours, and traffic flow on weekends and weekdays may vary in time distribution.

(3) The complexity of the scale of massive data. To reflect the periodicity of the traffic flow data, the model inputs historical data for at least one week, however, the mass data size can cause the model to be extremely complex in calculation.

(4) Uncertainty of a particular event. Certain special events, such as unusual weather and holidays, can greatly alter the flow of vehicles in a city, causing uncertainty in the predictions.

Disclosure of Invention

Aiming at the problems, the invention provides a road network short-time traffic flow prediction method based on a deep space-time residual error network aiming at urban road network traffic flow data with mass data scale on the premise of fully considering the spatiality of the road network, the proximity and the periodicity of the traffic flow data), and the method comprises the following steps:

the road network short-time traffic flow prediction method based on the deep space-time residual error network comprises the following steps:

step 1: dividing historical traffic flow data of a road network into two-dimensional data frames according to time periods, wherein each element in the data frames represents the traffic flow passing through one road section in the time period, and adjacent road sections are adjacent in the data frames in reality;

step 2: extracting h periods of traffic flow data from the data frames to form an adjacent data set; extracting traffic flow data in the same time period within d days from the data frames to form a periodic data set; extracting external factors influencing traffic flow of a road network to form an external factor data set;

and step 3: constructing a DST-ResNet (Deep spread-Temporal analytical Network) Network, wherein the DST-ResNet Network comprises a traffic flow data analysis Network and an external factor analysis Network, and the traffic flow data analysis Network comprises a proximity analysis Network and a periodic analysis Network; respectively training a proximity analysis network, a periodic analysis network and an external factor analysis network by using the proximity data set, the periodic data set and the external factor data set;

and 4, step 4: merging the output XR1 of the proximity analysis network and the output XR2 of the periodic analysis network to obtain XR, and merging the XR with the output XE of the external factor analysis network to be marked as X;

and 5: mapping X to [ -1,1] through a Tanh function, carrying out LOSS calculation with a target, and carrying out model parameter optimization by using a back propagation mode.

Furthermore, the proximity analysis network and the periodicity analysis network have the same structure and comprise convolution networks and residual error units, the convolution networks comprise a plurality of convolution layers which are directly connected, direct convolution is realized, and the input and output sizes of the convolution networks are not changed.

Further, the model of the residual unit is as follows:

X^(l+1)＝F(X^(l)；θ^(l))+X^(l),l＝1,…,L

where F is the residual function, θ^(l)Including all learnable parameters in the ith residual unit, L represents the number of network layers.

Furthermore, the external factor analysis network is composed of an input layer and two full-connection layers, the first full-connection layer receives input data and performs first-step feature fusion, and the second layer is used for expanding the output of the network to the size of a road network so as to perform subsequent fusion operation.

Further, the fusion of the proximity analysis network and the periodic analysis network is as follows:

wherein

Is Hadamard product, and utilizes Xavier method to randomly initialize two parameter matrixes W_t，W_p。

Further, the fusion mode of XR and XE is direct combination.

Has the advantages that: the invention can carry out overall prediction on the traffic flow of the road network, has high operation speed and accurate result, and only needs to know the space relative position of the road sections in the road network without knowing the connectivity between each road section.

Drawings

FIG. 1 is a logical mapping of an urban road network;

FIG. 2 is a drawing process of a data frame;

FIG. 3 is an overall frame diagram of the model;

FIG. 4 is a schematic diagram of a convolutional network;

FIG. 5 is a schematic diagram of a residual unit;

FIG. 6 is a comparison of holiday and workday traffic flows for a section of road;

FIG. 7 is a comparison of traffic flow for a certain road segment in rainy and sunny days;

FIG. 8 is a comparison between the real value and the predicted value of the traffic flow in 7 days of a certain road section;

FIG. 9 is a comparison between the real and predicted traffic flow values at the high and low peak periods of the road network.

Detailed Description

The invention is further illustrated by the following examples and figures.

0 data preprocessing

0.1 road network transitions

The values of R1, R2.. in the matrix on the right side of fig. 1 represent the traffic flow of the corresponding road segments in a city over a certain period of time.

The spatial dependency between the urban road networks can be reserved by logically mapping the urban road networks, so that the correlation between the road sections can be more easily captured when the convolution operation is carried out. One data frame represents the traffic flow of the entire road network for one time period.

0.2 data frame extraction

The time interval size predicted by the short-time traffic flow is usually 5min, 10min and 15min, wherein 15min is selected as the time interval size, and the traffic flow predicted in the ith time interval can be converted into a value of a prediction matrix roadnetwork (i) (marked as R (i)).

There are two temporal attributes of traffic flow data: proximity and periodicity, assuming a target period of i, data frame extraction is performed from the above two angles, respectively:

the method comprises the following steps that (1) a neighboring data set, wherein the count (i) is { R (i-h) }, R (i-3), R (i-2), R (i-1) }, and h values represent the number of data frames, and represent the number of channels of input data frames in a network, and the practical meaning is that traffic flow data of the previous h time periods are extracted to serve as input data, and the h values can be freely determined according to the requirement;

a periodic data set, period (i) { R (i-d × p),.., R (i-3 × p), R (i-2 × p), R (i-1 × p) }, where p represents the span of a day, since 15min is chosen as the period size, when p takes a fixed value 96; d represents the size of the cycle, typically one cycle in length, where d takes the value of 7. The method has the specific meaning that traffic flow data consistent with the time of the period i in the previous week are extracted from yesterday, the previous day, the previous year and the previous week;

secondly, extracting external factors influencing traffic flow of a road network, abnormal weather and holidays: ex (i) ═ weather (i), holiday (i).

The process of predicting the traffic flow of the road network is the process of predicting the value of R (i) by using the data frames of Recent (i), period (i) and Ex (i).

1 model analysis and design

1.1 model structural analysis and design

Fig. 3 shows the structure of a DST-ResNet network, which mainly consists of 2 networks: traffic flow data analysis network, external factor analysis network. Wherein the traffic flow data analysis network consists of a proximity analysis network and a periodicity analysis network.

Firstly, the traffic flow passing through each time interval of the urban road network is converted into a data frame form.

Secondly, two time characteristics are obtained from the data frame: proximity and periodicity, data extraction; and inputting the two characteristic data frames into respective residual error networks, capturing the spatial dependency between urban road networks through convolution, and recording the outputs of the two networks as XR1 and XR2 respectively.

Thirdly, some features such as abnormal weather and holidays are extracted from the external data set, input into the fully-connected neural network, and output XE.

The outputs XR1 and XR2 of the two residual networks are then aggregated once in combination with the parameter matrix, and the result is denoted XR. And XR is also fused with the output XE of the external factor analysis network and denoted X.

And finally, mapping X to [ -1,1] through a Tanh function, carrying out LOSS calculation with the target, and carrying out model parameter optimization by using a back propagation mode.

1.2 traffic flow data analysis network

The network mainly analyzes historical data of traffic flow from the aspects of proximity and periodicity, and the two analysis sub-networks of the proximity and the periodicity have the same structure. Wherein the analysis subnetwork is mainly composed of two parts: convolution network and residual unit, which will be analytically designed from two aspects below.

3.2.1 design of convolutional networks

Convolutional Neural Networks (CNN) are a type of feed forward Neural Networks (fed forward Neural Networks) that include convolution computation and have a Deep structure, and are one of the representative algorithms of Deep Learning (Deep Learning), and have a strong ability to hierarchically capture spatial structure information. In the urban road network, the traffic flows of adjacent road sections can influence each other in a short time, namely, the traffic flows of the adjacent road sections have potential correlation, and the convolutional neural network can just mine the potential regularity of the traffic flows, so that the convolution is used for capturing the dependency relationship of the traffic flows of the adjacent road sections. In addition, since car speeds are typically fast, the physical locations of the same car in adjacent time periods may be far apart, so that there may also be some correlation in traffic flow between two road segments that are far apart. Therefore, it is necessary to design a convolutional neural network with multiple layers (the specific number of layers is determined by the problem, and generally the larger the number of layers required for the road network, mainly depending on the size of the road network), for capturing the spatial correlation of the long distance road segments. Multiple convolutions can further capture dependencies between segments over greater distances, even the entire city.

The size of the data frame represents the size of the urban road network, the size of the data finally output by the model needs to be consistent with the size of the input data, and the output of the general convolution network is one-dimensional, so that the output structure of the network needs to be improved.

There may be two solutions to ensure that the input and output sizes of the convolutional network are unchanged:

(1) the input and output of each layer of convolution keep the same size, and do not carry on the downsampling at the same time, so the net final output can keep the same size as the initial input;

(2) adding a deconvolution (transposed convolution) layer at the end of the network, the convolution and downsampling will result in the image size becoming smaller, while the deconvolution can make the image size larger, and setting appropriate parameters can adjust the size of the final output image to the size of the input, so that the output and the input can be kept consistent, as shown in the following figure:

using downsampling + deconvolution loses a portion of the data content, resulting in a high error rate for the model. Direct convolution without downsampling increases the amount of computation, but has the advantage that multiple convolutions can be performed. In order to make the model have higher accuracy, downsampling and deconvolution are not used, and a direct convolution scheme is adopted.

Data size change formula before and after convolution:

O＝(I+2*P-K)/S+1 (1)

where I represents the input data size, O represents the output data size, K is the convolution kernel size, P is the fill size, and S is the step size.

From the above equation, if P, S is set to 1 and K is set to 3, the condition of I ═ O is satisfied.

3.2.2 design of residual Unit

When the direct convolution scheme is adopted, the size of the data frame is kept unchanged after each convolution layer, and the network can be extended infinitely in theory. The method aims to predict the traffic flow of the whole urban road network, so that only one network with a deeper level is needed to capture the dependency relationship in the whole urban road network range, and the larger the road network scale is, the more the number of required network layers is.

The training set LOSS generally decreases gradually as the number of network layers increases, but when the number of network layers is greater than a certain value, if the network depth is increased, the training set LOSS increases, which is a phenomenon that a gradient in a convolutional network disappears (explodes).

The problem of network accuracy rate reduction (error rise) caused by the fact that the network is too deep can be effectively solved by adding the residual error unit in the convolutional network. The principle is that if a convolutional network increases the number of layers in an identity mapping mode, the training error of the network after the number of layers is increased is not larger than the training error of the network without the identity mapping layer. That is, after the network adds the residual unit, the error will not become large and will most likely decrease.

A residual unit can be represented by the following diagram:

to avoid the network degradation problem due to too many network layers, residual error units are stacked behind the convolutional network of fig. 4 as follows:

X^(l+1)＝F(X^(l)；θ^(l))+X^(l),l＝1,…,L (2)

where F is the residual function (i.e., the residual unit of FIG. 5), and θ^(l)Including all learnable parameters in the ith residual unit. The gradient vanishing (explosion) problem can be effectively solved by adding a residual error unit in the convolution network to change the network into a residual error network.

1.3 overlay external factor analysis network

It is known from daily life experience that the size of urban road traffic flow may be affected by many complex external factors, such as holidays, weather and public emergencies.

By analyzing urban road network traffic flow data of the United states Borland metropolitan area during working days and holidays, the major influence of holidays on traffic flow is verified. As shown in fig. 6, the solid line represents a traffic flow curve during weekdays (12 months, 16 days to 20 days in 2019), and the dotted line represents a traffic flow curve during holidays (12 months, 23 days to 27 days in 2019, christmas on the maxmost statutory holidays in the united states). It can be seen by analyzing the change trend of the traffic flow of two adjacent weeks in the graph that holidays have an important influence on the size of the traffic flow.

And then analyzing the influence of the overlapped abnormal weather on the traffic flow, and selecting two sections of data from 19 days to 21 days in month 2 and 26 days to 28 days in month 2 in 2019. During days 19-21 of month 2, city weather was good, while during days 26-28 of month 2, only the first day was sunny, and the remaining two days were rainy. As shown in fig. 7, rain significantly reduced the traffic flow on the day compared to the same day of the previous week.

In implementation, the external factors considered by the model are mainly abnormal weather and holidays, because public emergencies have great uncertainty and are difficult to quantitatively analyze. The holiday data can be directly obtained, but the weather in the future period t is unknown, and the weather data in the previous period can only be used for replacing the future weather condition.

The external factor analysis network consists of an input layer and two fully connected layers. The first fully connected layer receives input data and performs a first step of feature fusion. The second layer is used to expand the output of the network to the size of the road network for subsequent fusion operations.

1.4 network convergence design

The model requires merging the outputs of the three sub-networks as shown in fig. 4. The proximity analysis network output XR1 is first fused with the periodic analysis network output XR 2. However, for different road segments, the proximity and periodicity do not affect the traffic flow itself to the same extent, and for some road segments the periodicity is important, while for other road segments the proximity may be more important, e.g. traffic flows near attractions and parks are more susceptible to periodicity and holidays than to proximity.

In general, different roads are affected by proximity and periodicity, but the extent to which each road is affected by these two factors varies. Therefore, the method designs a fusion method based on a parameter matrix, and fuses traffic flow data analysis networks (namely adjacent sub-networks and periodic sub-networks) of the model as follows:

wherein

Is a Hadamard product (i.e., a matrix element-by-element product). For each road section, two learnable parameters are used for adjusting the influence of the proximity and periodicity on the road section, and the learnable parameters of the whole road network are combined in a matrix to form a parameter matrix W_t，W_p。

And secondly, fusing external components, and directly combining the output XR of the first two components with the output XE of the external components.

Finally, the predicted road network traffic flow value of the t-th time period can be expressed as:

wherein the function of the tanh activation function is to ensure that the output value is between-1 and 1.

2 experiment and analysis of results

2.1 data Source

The experimental data come from official data of a Botland-Wingpenhua metropolitan area, http:// new. port. its. pdx. edu:8080/downloads/, wherein 80 main roads are selected to form an urban road network, the data sampling time interval is 15min, and traffic flow data of 2019, 4, month, 30 days to 6, month and 2 days are selected as training set data, and the total number is 3264; selecting traffic flow data from 3 days in 6 months to 9 days in 6 months in 2019 from the test set data, wherein 672 pieces are counted; and counting all weather data and holidays of the city during the period.

2.2 evaluation index

1) Root mean square error

The Root Mean Square Error (RMSE) can well reflect the deviation degree of the prediction value and the true value of the regression model, and the smaller the value is, the better the fitting effect is. The definition is as follows:

wherein, y_iThe actual value is represented by the value of,

representing the predicted value.

2) Determining coefficients

Determining the Coefficient (R)²) Is defined as the ratio of the regression sum of squares to the total sum of squares,

wherein, the regression Sum of Squares (SSR), which is the Sum of squares of the difference between the predicted data and the original data mean, is as follows:

the mean of the true values is indicated.

The Total Sum of Squares (SST), which is the sum of squares of the differences between the raw data and the mean, is given by the following formula:

therefore, determine the coefficient R²：

Namely:

R²normal value range ofIs enclosed as [0, 1]]Closer to 1 indicates a better fit of the model to the data.

2.3 analysis of results

The method is characterized in that a DST-ResNet model is built based on a PyTorch deep learning framework, and a data set is preprocessed and then placed into the model for training. After multiple times of experimental simulation debugging, model training parameters are selected as shown in table 1.

TABLE 1 model principal parameter Preset values

Where ResNet1 represents the proximity analysis subnetwork and ResNet2 represents the periodic analysis subnetwork. And fusing the two sub-networks by using a parameter matrix, calculating the loss degree of the current network by using a loss function, and finally optimizing model parameters by using an optimizer. Set batch size 100 and iterate 200 times. And after the model training is finished, putting the test data set into the model to operate to obtain a final prediction result.

Fig. 8 shows a graph comparing a predicted value and a true value of traffic flow for 7 days for a road segment randomly selected from 80 road segments.

It can be seen from fig. 8 that the prediction result of the model for the traffic flow of the road section better fits the real traffic flow situation, and accurately reflects the change from the peak-of-day period to the peak-of-day period, and particularly accurately predicts the periodic change within 7 days of the traffic flow.

Then, road network traffic flow data of a certain noon time period (representing a peak time period) and night time period (representing a peak time period) are randomly selected from the test set, and are predicted, and the prediction result is shown in fig. 9. It can be seen that the model can better fit the traffic flow of the whole road network no matter in the peak period or the low peak period.

To better analyze the merits of the model, the methodMethod a set of control experiments was added-road network traffic flow was predicted using LSTM. Respectively predicting all the periods to be predicted by using a DST-ResNet model and an LSTM model, and using two regression evaluation indexes RMSE and R²The predicted results were calculated, and the calculation results are shown in table 2.

TABLE 2 comparison of DST-ResNet and LSTM evaluation index partial results

Wherein a larger value of RMSE indicates a more inaccurate traffic flow prediction for the road network, and the RMSE value at low peak is generally smaller than the RMSE value at high peak. R²Values of (a) are between 0 and 1, with closer to 1 indicating better fit of the predicted road network traffic flow and the true traffic flow for this model, the more excellent the model. Statistics results for 7 days total 672 test set periods are compared as shown in table 3.

TABLE 3 DST-ResNet vs. LSTM statistics

From the analysis, the road network short-time traffic flow prediction model DST-ResNet based on the deep space-time residual error network can accurately predict the traffic flow in the next time period no matter whether the single road section or the whole road network is in the peak time period or the low peak time period. Meanwhile, compared with the LSTM model, the DST-ResNet model provided by the method has the advantages that the number of the time periods superior to the LSTM model accounts for more than 90%, and the model has obvious advantages in performance.

3 concluding sentence

The method analyzes traffic flow characteristics in detail on the theory and data, and fully grasps the internal relation of the traffic flow time-space characteristics. Traffic flow characteristics and complexity of road networks are fully considered when using data and building models. Experiments prove that the road network short-time traffic flow prediction DST-ResNet model based on the deep space-time residual error network, which is provided by the method, is an excellent solution for the urban road network short-time traffic flow prediction problem. In addition, the influence of more complex emergencies on the model is not considered in the method, and the robustness of the model is enhanced in the future, so that the method can be suitable for more complex application scenes.

Claims

1. The road network short-time traffic flow prediction method based on the deep space-time residual error network is characterized by comprising the following steps of:

and step 3: constructing a DST-ResNet network, wherein the DST-ResNet network comprises a traffic flow data analysis network and an external factor analysis network, and the traffic flow data analysis network comprises a proximity analysis network and a periodic analysis network; respectively training a proximity analysis network, a periodic analysis network and an external factor analysis network by using the proximity data set, the periodic data set and the external factor data set;

2. The road network short-time traffic flow prediction method based on the deep space-time residual error network as claimed in claim 1, wherein the proximity analysis network and the periodic analysis network have the same structure and comprise convolution networks and residual error units, the convolution networks comprise a plurality of convolution layers directly connected to each other, direct convolution is realized, and the input and output sizes of the convolution networks are not changed.

3. The road network short-time traffic flow prediction method based on the deep space-time residual error network according to claim 2, characterized in that the model of the residual error unit is as follows:

X^(l+1)＝F(X^(l)；θ^(l))+X^(l),l＝1,…,L

4. The method for predicting the short-term traffic flow of the road network based on the deep space-time residual error network as claimed in claim 1, wherein the external factor analysis network is composed of an input layer and two fully connected layers, the first fully connected layer receives input data and performs first-step feature fusion, and the second layer is used for expanding the output of the network to the size of the road network so as to perform subsequent fusion operation.

5. The road network short-time traffic flow prediction method based on the deep space-time residual error network according to claim 1, wherein the fusion of the proximity analysis network and the periodicity analysis network is as follows:

wherein

6. The road network short-time traffic flow prediction method based on the deep space-time residual error network as claimed in claim 1, wherein the XR and XE are directly merged.