CN112017436B

CN112017436B - Method and system for predicting urban traffic travel time

Info

Publication number: CN112017436B
Application number: CN202010940567.2A
Authority: CN
Inventors: 向世明; 张奇; 孟高峰; 霍春雷; 潘春洪
Original assignee: Institute of Automation of Chinese Academy of Science
Current assignee: Institute of Automation of Chinese Academy of Science
Priority date: 2020-09-09
Filing date: 2020-09-09
Publication date: 2021-09-28
Anticipated expiration: 2040-09-09
Also published as: CN112017436A

Abstract

The invention relates to a method and a system for predicting urban traffic travel time, wherein the method comprises the steps of dividing a city to be tested into rectangular grids; based on the rectangular grid, constructing a normalized floating point traffic flow matrix according to historical traffic flow data; training an urban traffic flow prediction network according to the normalized floating point traffic flow matrix; simplifying the track point sequence of the vehicle driving path according to the rectangular grid to obtain a gridded track of the vehicle driving path; determining the characteristic vector of each track point in the gridding track according to the normalized floating point traffic flow matrix and the gridding track; training a urban traffic travel time prediction network according to the feature vectors of the track points; based on the urban traffic flow prediction network and the urban travel time prediction network, the travel time required by the vehicle to be detected to run through the path to be detected can be accurately determined according to the running path to be detected of the vehicle to be detected, and the prediction precision under a complex scene can be improved.

Description

Method and system for predicting urban traffic travel time

Technical Field

The invention relates to the technical field of urban road condition information processing, in particular to a method and a system for predicting urban traffic travel time by combining real-time road condition information.

Background

The urban traffic vehicle travel time prediction is the basis of application of path planning, vehicle navigation, intelligent car sharing, taxi intelligent ordering and the like, and the task is to give a certain specific path and predict the time required by the vehicle to run through the path under the condition of considering possible parking delay and intersection waiting delay. When a traveler can consider a plurality of selectable paths, an accurate travel time prediction result can help to select an optimal route and avoid congested road sections, and the method plays an important role in optimizing city management and improving the traveling efficiency of citizens.

At present, the traditional urban traffic travel time prediction methods mainly fall into two categories, namely parametric methods and nonparametric methods. The parameterization method relies on the existing mathematical and statistical theories for prediction, and comprises a linear regression model, an autoregressive integrated moving average model, Kalman filtering, a Bayesian dynamic linear model, a hidden Markov model, a support vector machine and the like. These methods have been successful to varying degrees on short-distance travel times, but their accuracy on long-distance travel time predictions is difficult to meet application requirements. Non-parametric methods use a data-driven approach to capture the underlying knowledge of the data. The non-parametric method does not need to make assumptions on the applicable data distribution situation of the model and has certain data self-adaption capability. Nonparametric methods include k-nearest neighbor estimation, voting decisions, gaussian processes, and the like. But the parameter method has large calculation amount and is difficult to meet the requirement of real-time performance index.

The traditional parametric method and the non-parametric method are difficult to be applied to practical scenes at present, and mainly have the following two reasons. First, the traditional method only utilizes historical urban traffic travel time data, neglects the influence of real-time road conditions on urban traffic travel time prediction. When a user needs to predict the travel time of a certain path, the traditional method is only based on the traffic road condition on the path at the current time point, and the real-time road condition information is difficult to be dynamically fused. However, when the user travels to a certain road section, the road condition of the road section has changed, such as changing from the original congestion state to the current clear state, or changing from the original clear state to the current congestion state. Such a change in the road condition is particularly noticeable in the case of long-distance driving. Secondly, the conventional method has limited nonlinear fitting capability. In a short-distance local range and in an urban area with obvious traffic condition change rules, the traditional method can generally obtain a more ideal urban traffic travel time prediction result. However, for a long distance and a large range of areas, the traffic conditions present highly nonlinear dynamic change characteristics, and the space-time correlation relationship is quite complex. The traditional method is difficult to capture the complex relations, so that a large prediction error of urban traffic travel time is generated.

In recent years, deep learning has been developed vigorously, and deep learning methods represented by Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs) have been widely introduced to a large number of tasks related to machine learning, and have been greatly successful in a large number of practical applications such as image recognition, image semantic segmentation, visual object detection, video content understanding, machine translation, and automatic driving. Deep learning has strong nonlinear fitting capability. Meanwhile, due to the flexibility of the network structure design, the deep learning is beneficial to the fusion and processing of the multivariate information.

In conclusion, the traditional method has a difficult ideal effect in the task of predicting the traffic travel time in the urban long-distance city under the complex scene.

Disclosure of Invention

In order to solve the above problems in the prior art, i.e. to improve the accuracy of predicting the urban traffic travel time, the present invention aims to provide a method and a system for predicting the urban traffic travel time.

In order to solve the technical problems, the invention provides the following scheme:

a prediction method of urban traffic travel time, the prediction method comprising:

dividing a city to be tested into rectangular grids;

based on the rectangular grid, a normalized floating point traffic flow matrix is constructed according to historical traffic flow data;

training an urban traffic flow prediction network according to the normalized floating point traffic flow matrix;

simplifying the track point sequence of the vehicle running path according to the rectangular grid to obtain a gridded track of the vehicle running path;

determining the characteristic vector of each track point in the gridding track according to the normalized floating point traffic flow matrix and the gridding track of the vehicle running path;

training a prediction network of urban traffic travel time according to the feature vectors of all track points in the gridding track;

and determining the travel time required by the vehicle to be tested to run through the path to be tested according to the running path to be tested of the vehicle to be tested on the basis of the urban traffic flow prediction network and the urban travel time prediction network.

Optionally, the constructing a normalized floating point traffic flow matrix according to historical traffic flow data based on the rectangular grid specifically includes:

will select history n_dDays are training periods, each day of which is divided into K with fixed time intervals_dA time interval; the total number of consecutive time intervals into which the training period is divided is K: k is n_d×K_d；

Counting the track data of the vehicles entering the networking system in the rectangular grid G in the training period, wherein the longitude position, the latitude position and the time information of the vehicles are recorded in each track data;

representing the number of vehicles appearing in the ith row and jth column grid sub-area of the rectangular grid G in the tth time interval;

traversing all time intervals in the training period, counting the vehicle flow entering the networking system in the same time interval to obtain K traffic flow matrixes which are respectively marked as X₁,X₂,...,X_k,...,X_K；X_kA traffic flow matrix in a K-th time interval, K being 1, 2.... K;

according to the following formula, carrying out normalization floating point processing on each traffic flow matrix to obtain a corresponding normalization floating point traffic flow matrix:

wherein the content of the first and second substances,

for the normalized floating point traffic flow matrix Y in the t-th time interval_tRow i and column j elements of (1), x₀Represents the maximum of all elements in the K traffic flow matrices.

Optionally, the training of the urban traffic flow prediction network according to the normalized floating point traffic flow matrix specifically includes:

according to the normalized floating point traffic flow matrix, constructing a corresponding first training sample:

P_t＝(A_t；B_t)，

wherein t represents a time interval number, P_tRepresenting a first training sample; a. the_tRepresenting a first training sample P_tThe sample characteristic is a three-dimensional tensor of L rows, M columns and J layers, and is formed by stacking normalized traffic flow matrixes obtained in the first J continuous time intervals with the t-th time interval as an end point according to the time sequence: a. the_t＝[Y_t-J+1,Y_t-J+2,…,Y_t-1,Y_t]；

B_tRepresenting a first training sample P_tThe sample mark is a three-dimensional tensor of an L-row M-column Q layer, and is formed by stacking normalized traffic flow matrixes obtained in Q continuous time intervals with the t +1 th time interval as a starting point according to a time sequence: b is_t＝[Y_t+1,Y_t+2,…,Y_t+Q]；

Wherein, Y_t-J+1For the normalized floating-point traffic flow matrix, Y, obtained in the t +1 time interval_tFor the normalized floating-point traffic flow matrix, Y, obtained in the t-th time interval_t+QObtaining a normalized floating point traffic flow matrix in the t + Q time intervals;

training an urban traffic flow prediction network according to the first training sample; the urban traffic flow prediction network is used for predicting the normalized floating point traffic flow of the rectangular grid G in the future Q continuous time intervals.

Optionally, the track point sequence of the vehicle driving path is simplified according to the rectangular grid, so as to obtain a gridded track of the vehicle driving path, and the method specifically includes:

each vehicle driving path is described by a track point sequence, and each track point comprises a longitude coordinate and a latitude coordinate of the track point;

for a given vehicle travel path, the corresponding trajectory sequence is T, which is described by N trajectory points: t { (x)₁,y₁),(x₂,y₂),…,(x_N,y_N)}；

Wherein x is₁And y₁Respectively representing the longitude and latitude, x, of the first trace point_NAnd y_NRespectively representing the longitude and the latitude of the Nth track point;

according to the rectangular grid G, dividing N track points of the track sequence T into different segments in sequence, and enabling the track points in each segment to be located in the same grid area of the rectangular grid G:

wherein, let the trajectory T be divided into n segments, the first segment containing the point (x) from the first trajectory₁,y₁) To d₁One track point

The second fragment comprises the second fragment from₁+1 track points

To d₂One track point

By analogy, the nth segment contains the segment from d_n-1+1 track points

To the Nth track point (x)_N,y_N) And | represents a fragment space symbol; d₁、d₂、d_n-1N is a natural number;

calculating the average value of the longitude and the latitude of the track points in the same segment to obtain a gridded track T_G：

Wherein the content of the first and second substances,

represents the average of the longitudes of track points belonging to the ith segment,

mean value, x, representing the latitude of a tracing point belonging to the ith segment_jAnd y_jLongitude coordinate and latitude coordinate, when i is 1, d₀1 is ═ 1; when i is n, d_n＝N；

Optionally, determining a feature vector of each track point in the gridding track according to the normalized floating point traffic flow matrix and the gridding track of the vehicle driving path, specifically including:

taking out continuous Q normalized floating point traffic flow matrixes Y from the normalized floating point traffic flow matrix_p+1,Y_p+2,…,Y_p+Q(ii) a Wherein p is the first track point (x) of the vehicle driving path T₁,y₁) The sequence number of the time interval to which the time belongs in the training time interval;

according to the gridding track T_GThe ith track point of

R is located at rectangular grid G_iRow and c_iTaking out the normalized floating point traffic flow matrix Y from the grid area where the row is located_p+1,Y_p+2,…,Y_p+QMiddle r_iRow and c_iElements of the column:

wherein the content of the first and second substances,

for normalizing floating-point traffic flow matrix Y_p+QR of_iRow and c_iElements of a column; r is_iAnd c_iRespectively, a gridded track T_GThe ith track point of

The row serial numbers and the column serial numbers of the rectangular grids;

constructing a gridded trajectory T_GCharacteristic vector f of ith track point_i：

Wherein n is a gridding track T_GThe number of the track points; the superscript T denotes vector transposition.

Optionally, training the urban traffic travel time prediction network according to the feature vector of each track point in the gridding track specifically includes:

acquiring the recording time o of each track point in the track sequence T of the vehicle running path₁,o₂,…,o_N(ii) a Wherein o is_NRecording time of the Nth track point;

according to the rectangular grid G, dividing each recording time into different segments in sequence, and enabling the recording time in each segment to be located in the same grid area of the rectangular grid G:

wherein the recording time is divided into n segments, the first oneThe segments containing the time from the first recording o₁To d₁A recording time

The second fragment comprises the second fragment from₁+1 recording time

To d₂A recording time

By analogy, the nth segment contains the segment from d_n-1+1 recording time

To the Nth recording time o_NAnd | represents a fragment space symbol; d₁、d₂、d_n-1N is a natural number | | | represents a fragment space character, d₁、d₂、d_n-1N is a natural number;

calculating the driving time required by driving the current section i:

wherein the content of the first and second substances,

for the time of the i-th segment, corresponding to the feature vector f_iThe marking time of (1); i. j, d_i、d_i-1Are all natural numbers, and when i is equal to 1, d₀1 is ═ 1; when i is n, d_n+1＝N；

And constructing a second training sample according to the feature vector and the driving time of each segment after driving:

wherein F represents the second trainingTraining samples, feature vector sequence f₁,f₂…,f_nFor the sample characteristics of the second training sample F, time sequence

Labeling the sample of the second training sample F; f. of_nThe characteristic vector of the nth track point of the gridding track of the vehicle running path is obtained;

to correspond to the feature vector f_nThe marking time of (1);

and training an urban traffic travel time prediction network according to the second training sample, wherein the urban traffic travel time prediction network is used for predicting the travel time of the whole travel path.

Optionally, determining the travel time required by the vehicle to be tested to finish running the path to be tested according to the to-be-tested running path of the vehicle to be tested based on the urban traffic flow prediction network and the urban travel time prediction network, specifically including:

calculating normalized floating point traffic flow matrixes in the first J continuous time intervals by taking the current running time of the vehicle to be detected as a terminal time interval, and stacking the matrixes into a three-dimensional tensor according to a time sequence:

A＝[Y₁,Y₂,…,Y_J]；

a is a three-dimensional tensor formed by three-dimensional stereo data of L rows, M columns and J layers, and is formed by stacking normalized floating point traffic flow matrixes obtained in the first J continuous time intervals with the current time interval as an end point according to a time sequence, and Y is₁For the normalized floating point traffic flow matrix, Y, obtained in the first J-1 time interval₂For the normalized floating point traffic flow matrix, Y, obtained in the first J-2 time intervals_JObtaining a normalized floating point traffic flow matrix in the current time; l is the number of rows of the rectangular grid G in the latitude direction, and M is the number of columns of the rectangular grid G in the longitude direction;

according to the three-dimensional tensor A and urban traffic flowAnd the prediction network predicts the normalized floating point traffic flow in the future continuous Q time intervals: b is₁,B₂,…,B_Q；

Wherein, B_QPredicting the normalized floating point traffic flow in the future Q-th time interval predicted by the network for the urban traffic flow;

determining a corresponding gridding track R according to a track sequence R of a running path to be tested of a vehicle to be tested_GFeature vectors of the track points;

sample feature B of trace sequence R: b ═ B₁,b₂,…,b_n)；

Wherein, the sample characteristic B of the running path R to be measured is composed of a characteristic vector sequence B₁,b₂,…,b_nComposition is carried out; b₁Is a gridded track R_GThe feature vector of the first trace point of (b)_nIs a gridded track R_GThe feature vector of the nth trace point; n is a gridded locus R_GThe number of the contained track points;

gridding track R of vehicle running path R to be tested_GThe feature vector of the ith trace point of (1):

wherein, b_iRepresenting a gridded footprint R_GThe feature vector of the ith trace point of (a),

and

respectively represent gridded traces R_GLongitude and latitude of the ith trace point,

floating point normalized traffic flow moment B in the first time interval predicted by urban traffic flow prediction network at present moment₁R of_iRow and c_iThe elements of the column are,

a floating point normalized traffic flow matrix B in a second time interval predicted by the urban traffic flow prediction network at the current moment₂R of_iRow and c_iThe elements of the column are,

floating point normalization traffic flow matrix B in Q time interval predicted by urban traffic flow prediction network at present moment_QR of_iRow and c_iElements of a column; r is_iAnd c_iRespectively, a gridded trace R_GThe ith track point of

The row serial numbers and the column serial numbers of the rectangular grids; q is the number of time intervals predicted by the urban traffic flow prediction network, and superscript T represents vector transposition;

according to the gridding track R_GAnd determining the time of the running path R to be tested of the vehicle to be tested after the vehicle to be tested runs.

In order to solve the technical problems, the invention also provides the following scheme:

a system for predicting urban traffic travel time, the prediction system comprising:

the grid dividing unit is used for dividing the city to be detected into rectangular grids;

the normalization processing unit is used for constructing a normalization floating point traffic flow matrix according to historical traffic flow data based on the rectangular grid;

the first training unit is used for training an urban traffic flow prediction network according to the normalized floating point traffic flow matrix;

the simplification processing unit is used for simplifying the track point sequence of the vehicle driving path according to the rectangular grid to obtain a gridded track of the vehicle driving path;

the characteristic determining unit is used for determining the characteristic vector of each track point in the gridding track according to the normalized floating point traffic flow matrix and the gridding track of the vehicle running path;

the second training unit is used for training the urban traffic travel time prediction network according to the feature vectors of all track points in the gridding track;

and the time prediction unit is used for determining the travel time required by the vehicle to be detected to run through the path to be detected according to the to-be-detected running path of the vehicle to be detected based on the urban traffic flow prediction network and the urban travel time prediction network.

a system for predicting urban traffic travel time, comprising:

a processor; and

a memory arranged to store computer executable instructions that, when executed, cause the processor to:

dividing a city to be tested into rectangular grids;

a computer-readable storage medium storing one or more programs that, when executed by an electronic device including a plurality of application programs, cause the electronic device to:

dividing a city to be tested into rectangular grids;

According to the embodiment of the invention, the invention discloses the following technical effects:

according to the invention, the urban traffic flow prediction network and the urban traffic travel time prediction network are trained by dividing the to-be-detected city into rectangular grids and referring to historical traffic flow data, and the running time of the to-be-detected running path of the to-be-detected vehicle can be determined in real time through the urban traffic flow prediction network and the urban traffic travel time prediction network, so that the prediction precision under a complex scene can be improved.

Drawings

FIG. 1 is a flow chart of a method for predicting urban traffic travel time according to the present invention;

fig. 2 is a structural diagram of an urban traffic flow prediction network;

FIG. 3 is a block diagram of a city traffic travel time prediction network;

FIG. 4 is a schematic block diagram of the urban traffic travel time prediction system according to the present invention.

Description of the symbols:

the system comprises a grid division unit-1, a normalization processing unit-2, a first training unit-3, a simplification processing unit-4, a characteristic determination unit-5, a second training unit-6 and a time prediction unit-7.

Detailed Description

Preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood by those skilled in the art that these embodiments are only for explaining the technical principle of the present invention, and are not intended to limit the scope of the present invention.

The invention aims to provide a method for predicting urban traffic travel time, which can determine the travel time of a to-be-detected vehicle to be detected on a real-time basis and improve the prediction precision in a complex scene by dividing a to-be-detected city into rectangular grids, referring to historical traffic flow data, training an urban traffic flow prediction network and an urban traffic travel time prediction network and using the urban traffic flow prediction network and the urban traffic travel time prediction network.

In order to make the aforementioned objects, features and advantages of the present invention comprehensible, embodiments accompanied with figures are described in further detail below.

As shown in fig. 1, the method for predicting urban traffic travel time of the present invention comprises:

step 100: and dividing the city to be tested into rectangular grids.

Step 200: and constructing a normalized floating point traffic flow matrix according to historical traffic flow data based on the rectangular grid.

Step 300: and training an urban traffic flow prediction network according to the normalized floating point traffic flow matrix.

Step 400: and simplifying the track point sequence of the vehicle running path according to the rectangular grid to obtain a gridded track of the vehicle running path.

Step 500: and determining the characteristic vector of each track point in the gridding track according to the normalized floating point traffic flow matrix and the gridding track of the vehicle running path.

Step 600: and training the urban traffic travel time prediction network according to the characteristic vectors of all track points in the gridding track.

Step 700: and determining the travel time required by the vehicle to be tested to run through the path to be tested according to the running path to be tested of the vehicle to be tested on the basis of the urban traffic flow prediction network and the urban travel time prediction network.

In step 100, the city to be tested is divided into a rectangular grid G according to the road grade, the road network density, the traffic driving state and other factors of each region in the city, and the sizes of each row and each column of the rectangular grid G may not be consistent. And the rectangular grid G is divided into L rows in the latitudinal direction and M columns in the longitudinal direction. Here, L and M are both natural numbers. Thus, the rectangular grid G divides the urban area of interest into L × M grid regions, each grid region corresponding to a geographically different rectangular region.

In step 200, the constructing a normalized floating point traffic flow matrix according to historical traffic flow data based on the rectangular grid specifically includes:

step 210: will select history n_dDays are training periods, each day of which is divided into K with fixed time intervals_dA time interval; the total number of consecutive time intervals into which the training period is divided is K: k is n_d×K_d。

It is noted that this reference is made to aThe problem of individual geographical area division and time interval selection. If the grid area is too small, there may be some time intervals when there is no vehicle flow in some grid areas; if the grid area is too large, it is difficult to accurately describe the traffic flow characteristics. In the embodiment of the present invention, the number L of rows in the latitude direction of the rectangular grid G is set to 120, the number M of columns in the longitude direction of the rectangular grid G is set to 120, and the number K of time intervals divided per day is set to_dIs 360. That is, the time interval was 4 minutes. The choice of a time interval of 4 minutes is primarily to take into account that for most cities, the traffic conditions will not normally change frequently during this time interval.

Step 220: and counting the track data of the vehicles entering the networking system in the training period in the rectangular grid G, wherein each track data records the longitude position, the latitude position and the time information of the vehicle.

In the implementation of the present invention, the networking system may be a combination of one or more systems, such as a taxi operation management system, a public transportation vehicle bus operation management system, a trip service system, a car networking system, and the like.

Indicating the number of vehicles appearing in the ith row and jth column grid sub-region of the rectangular grid G during the t time interval. Here, the

And then, traversing the sequence number i through an integer from 1 to L and traversing the sequence number j through an integer from 1 to M to obtain the traffic matrix with the size of L rows and M columns. Recording the flow matrix as X_t. Flow matrix X_tEach element of (a) records the vehicle flow in the corresponding grid area in the t-th time interval. t is an integer ranging from 1 to K.

Step 230: traversing all time intervals in the training period, counting the vehicle flow entering the networking system in the same time interval to obtain K traffic flow matrixes which are respectively marked as X₁,X₂,...,X_k,...,X_K；X_kA traffic flow matrix in the kth time interval is represented, K being 1, 2.

Step 240: according to the following formula, carrying out normalization floating point processing on each traffic flow matrix to obtain a corresponding normalization floating point traffic flow matrix:

wherein the content of the first and second substances,

for the normalized floating point traffic flow matrix Y in the t-th time interval_tRow i and column j elements of (1), x₀The maximum value of all elements in the K traffic flow matrices (the maximum value of the number of vehicles appearing in all time intervals and all grid intervals) is represented.

The relative value of the traffic flow of each rectangular grid can be provided by introducing a floating point normalization operation, so that the deviation caused by insufficient statistics of the traffic flow is reduced.

In step 300, a training sample set required for constructing an urban traffic flow prediction network is constructed. The urban traffic flow prediction network is used for predicting the traffic flow in a plurality of future continuous time intervals by utilizing time sequence data formed by the normalized floating point traffic flow matrix. In order to train the urban traffic flow prediction network, a training sample needs to be constructed from a floating point normalized traffic flow matrix obtained in a training period. Each training sample is composed of two parts of a sample characteristic and an output mark.

Specifically, the training of the urban traffic flow prediction network according to the normalized floating point traffic flow matrix includes:

step 310: according to the normalized floating point traffic flow matrix, constructing a corresponding first training sample:

P_t＝(A_t；B_t)，

wherein t represents timeSection number, P_tRepresenting a first training sample; a. the_tRepresenting a first training sample P_tThe sample characteristic is a three-dimensional tensor of L rows, M columns and J layers, and is formed by stacking normalized traffic flow matrixes obtained in the first J continuous time intervals with the t-th time interval as an end point according to the time sequence: a. the_t＝[Y_t-J+1,Y_t-J+2,…,Y_t-1,Y_t](ii) a Wherein, Y_t-J+1For the normalized floating-point traffic flow matrix, Y, obtained in the t +1 time interval_t-J+2For the normalized floating-point traffic flow matrix, Y, obtained in the t +2 time interval_t-1For the normalized floating point traffic flow matrix, Y, obtained in the t-1 time interval_tFor the normalized floating point traffic flow matrix obtained in the t-th time interval, a natural number L is the number of rows of a rectangular grid G in the latitude direction, a natural number M is the number of columns of the rectangular grid G in the longitude direction, J is the number of continuous time intervals of sample characteristics, and Q is the number of continuous time intervals for constructing sample marks (namely the number of the normalized floating point traffic flow matrix in future continuous time intervals required to be predicted by the urban traffic flow prediction network). Here, J and Q are both natural numbers;

Wherein, Y is_t+1For the normalized floating-point traffic flow matrix, Y, obtained in the t +1 time interval_t+2For the normalized floating-point traffic flow matrix, Y, obtained in the t +2 time interval_t+QAnd for the normalized floating point traffic flow matrix obtained in the t + Q time intervals, the natural number L is the number of rows of the rectangular grid G in the latitude direction, the natural number M is the number of columns of the rectangular grid G in the longitude direction, the natural number Q is the number of continuous time intervals for constructing the sample markers, and t is a natural number. Sample characteristics A_tAnd sample marker B_tAre all known.

Considering the need to predict the flow data in the future Q consecutive time intervals, t may take an integer between 1 and K-J-Q +1 for a normalized floating point traffic flow matrix with a total length of K. Thus, a total of K-J-Q training samples can be obtained in the manner described above. In the present embodiment, the natural numbers J and Q are both set to 10.

Step 320: training an urban traffic flow prediction network according to the first training sample; the urban traffic flow prediction network is used for predicting the normalized floating point traffic flow of the rectangular grid G in the future Q continuous time intervals.

The backbone structure of the urban traffic flow prediction network of the present embodiment is shown in table 1. In Table 1, conv represents the convolutional layer. The urban traffic flow prediction network employed in this embodiment has 7 convolutional layers in total. The convolutional layer conv1, convolutional layer conv2 and convolutional layer conv3 form an encoder for learning high-level semantic features; the convolutional layer conv4, convolutional layer conv5, convolutional layer conv6 and convolutional layer conv7 constitute a decoder, and the decoder is used for decoding the high-level semantic features step by step to obtain a traffic flow prediction result in each grid area. The convolution kernel sizes of the urban traffic flow prediction network in the embodiment are all 3 × 3.

In table 1, the number of input channels of the convolutional layer conv1 is 10 because the urban traffic flow prediction network employed in the present embodiment takes a normalized floating-point traffic flow matrix obtained in 10 consecutive time regions as input data. The number of input channels of convolutional layer conv7 is 10, because the urban traffic flow prediction network employed in the present embodiment outputs normalized floating point traffic flow data for predicting future 10 consecutive time intervals.

In table 1, the numbers of output channels of the convolution layer conv1, convolution layer conv2, convolution layer conv3, convolution layer conv4, convolution layer conv5 and convolution layer conv6 of the urban traffic flow prediction network employed in the present embodiment are 20, 40, 80, 40, 20 and 10, respectively. Thus, the parameters in table 1 determine the structure of the urban traffic flow prediction network employed in the present embodiment.

Each convolution operation is followed by an excitation operation and is performed by a commonly used Linear rectification function (ReLU). The linear rectification function is the operation of taking the maximum value of the convolution result and zero, namely if the convolution result is larger than zero, the result is kept, otherwise, zero is set.

It is noted that, in the present invention, the urban traffic prediction network does not perform the pooling (posing) operation, i.e., the rectangular mesh area down-sampling operation, which is usually employed in the encoder. This is because urban traffic conditions exhibit local regional relevance more than global (entire cities), and therefore it is not necessary to enlarge the region of feature learning by down-sampling.

As shown in fig. 2, the output of convolutional layer conv1 (after the excitation operation) is the input of convolutional layer conv2, the output of convolutional layer conv2 (after the excitation operation) is the input of convolutional layer conv3, and the output of convolutional layer conv3 (after the excitation operation) is the input of convolutional layer conv 4.

As shown in fig. 2, the input of convolutional layer conv5 is formed by concatenating (contistate) the output of convolutional layer conv2 and the output of convolutional layer conv4, i.e., "conc" operation shown in fig. 2. Thus, the number of input channels of convolutional layer conv5 was 80, as shown in table 1. The input to convolutional layer conv6 is the concatenation of the output of convolutional layer conv1 and convolutional layer conv5 output, the "concatenation" operation shown in fig. 2. Thus, the number of input channels of convolutional layer conv6 was 40, as shown in table 1. The input to convolutional layer conv7 is the concatenation of the input to convolutional layer conv1 and the output to convolutional layer conv6, the "concatenation" operation shown in fig. 2. Thus, the number of input channels of convolutional layer conv7 was 20, as shown in table 1. In the invention, the 'concatenation' operation is introduced to increase the precision of the normalized floating point traffic flow in each grid region.

Table 1: urban traffic flow prediction network convolution kernel size and number of input and output channels of each layer

Name (R)	Convolution kernel size	Number of input channels	Number of output channels
				conv1	3×3	10	20
conv2	3×3	20	40
				conv3	3×3	40	80
conv4	3×3	80	40
				conv5	3×3	80	20
conv6	3×3	40	20
				conv7	3×3	30	10

And finally, training the urban traffic flow prediction network adopted by the embodiment. In the present invention, the Mean Squared Error (MSE) is used as a loss function for the network. The optimization algorithm employs an error back-propagation algorithm that is classical in the art. In the present invention, the learning rate is 0.001, the size of the batch data set is 16, and each batch data set has 20 training rounds.

After the network is trained, when the network is applied, a three-dimensional tensor formed by stacking normalized floating point traffic flow matrixes obtained in J continuous time regions in a time sequence is given, the network outputs a three-dimensional tensor formed by stacking Q matrixes, and each matrix is a normalized floating point traffic flow matrix in a time interval predicted by the network in a time sequence.

The invention divides the urban region of interest into a rectangular grid G comprising L rows and M columns. The spatial resolution of the rectangular grid G is less than the spatial resolution of the vehicle travel path that can be recorded by existing positioning systems, such as the north navigation system. In order to characterize the travel path, it is necessary to keep the sequence of trajectory points of the travel path consistent with the spatial resolution of the rectangular grid G. Therefore, it is necessary to simplify the meshing of the original trajectory sequence of the vehicle travel path by the rectangular mesh G.

Specifically, in step 400, the simplifying processing is performed on the track point sequence of the vehicle driving path according to the rectangular grid to obtain a gridded track of the vehicle driving path, including:

step 410: each vehicle driving path is described by a track point sequence, and each track point comprises a longitude coordinate and a latitude coordinate of the track point;

for a given vehicle travel path, the corresponding trajectory sequence is T, railThe trace sequence T is described by N trace points: t { (x)₁,y₁),(x₂,y₂),…,(x_N,y_N)}；

Wherein x is₁And y₁Respectively representing the longitude and latitude, x, of the first trace point_NAnd y_NRespectively representing the longitude and latitude of the nth trace point.

Step 420: according to the rectangular grid G, dividing N track points of the track sequence T into different segments in sequence, and enabling the track points in each segment to be located in the same grid area of the rectangular grid G:

The second fragment comprises the second fragment from₁+1 track points

To d₂One track point

By analogy, the nth segment contains the segment from d_n-1+1 track points

To the Nth track point (x)_N,y_N) And | represents a fragment space symbol; d₁、d₂、d_n-1And n is a natural number.

Step 430: calculating the average value of the longitude and the latitude of the track points in the same segment to obtain a gridded track T_G：

Wherein the content of the first and second substances,

Wherein the symbol "x" represents the longitude coordinate of the corresponding track point, the symbol "y" represents the latitude coordinate of the corresponding track point, d₁、d₂、d_n-1And n is a natural number.

For a given vehicle travel path, it is not sufficient to predict the time to travel through the path using only the position coordinates of the track points in the path. To achieve the purpose, the invention adds the normalized floating point traffic flow in the corresponding grid area to each track point. The normalized floating point traffic flow is dynamic information, closely related to the vehicle travel time, and thus can be used to estimate the travel time.

Setting a gridded track T obtained by gridding and simplifying the vehicle running path track T in the step S3_GThe ith track point of

R is located at rectangular grid G_iRow and c_iThe grid area where the columns are located. Here, the natural number i is an integer ranging from 1 to n, and the natural number r_iIs an integer ranging from 1 to L, a natural number c_iThe value range of (1) to (M) is an integer, the natural number L is the number of lines of the rectangular grid G in the latitude direction, and the natural number M is the number of lines of the rectangular grid G in the latitude directionNumber of columns in the longitudinal direction.

Specifically, in step 500, determining a feature vector of each track point in the gridded track according to the normalized floating point traffic flow matrix and the gridded track of the vehicle driving path includes:

step 510: taking out continuous Q normalized floating point traffic flow matrixes Y from the normalized floating point traffic flow matrix_p+1,Y_p+2,…,Y_p+Q(ii) a Wherein p is the first track point (x) of the vehicle driving path T₁,y₁) The sequence number of the time interval to which the training period belongs.

In the training period, the vehicle driving path track T belongs to historical data, so that the first track point (x) of the vehicle driving path track T₁,y₁) Is known. Marking the first track point (x)₁,y₁) Time of (d) the index value of the belonging time interval of the training period in step S1 is p. Where p is a natural number.

Step 520: according to the gridding track T_GThe ith track point of

wherein the content of the first and second substances,

The row sequence number and the column sequence number are located in the rectangular grid.

Step 530: constructing a gridded trajectory T_GCharacteristic vector f of ith track point_i：

Wherein f is_iRepresenting a gridded track T_GThe feature vector of the ith trace point of (a),

and

respectively represent gridded tracks T_GLongitude and latitude of the ith trace point,

normalizing traffic flow matrix Y for floating point_p+1R of_iRow and c_iThe elements of the column are,

normalizing traffic flow matrix Y for floating point_p+2R of_iRow and c_iThe elements of the column are,

normalizing traffic flow matrix Y for floating point_p+QR of_iRow and c_iElements of a column; r is_iAnd c_iRespectively, a gridded track T_GThe ith track point of

A row index value and a column index value located in a rectangular grid; p is an index value of the time of the first track point of the vehicle driving path T in the time interval of the training time period; the natural number Q is the number of time intervals predicted by the urban traffic flow prediction network; natural number n is gridding track T_GThe number of the track points; the natural number i having a value in the range of 1 to nAn integer number; the superscript T denotes vector transposition.

Gridding track T_GEach track point has both position coordinate information and dynamic information related to the traffic state, so that the accuracy of the estimation of the running time is improved.

In step 600, training a prediction network of urban traffic travel time according to the feature vectors of the track points in the gridded track specifically includes:

step 610: acquiring the recording time o of each track point in the track sequence T of the vehicle running path₁,o₂,…,o_N(ii) a Wherein o is_NThe recording time of the Nth track point is recorded.

Specifically, for a track T of a vehicle traveling path within the rectangular grid G, the track T is simplified into a grid according to the method of step 400 to obtain a grid track T_G. Further, a gridding track T is obtained according to the method of step 500_GOf each trace point, i.e. f₁,f₂…,f_n. Wherein f is₁For gridding the track T_GOf the first trace point of (a), f₂For gridding the track T_GOf the second trace point, f_nFor gridding the track T_GThe feature vector of the nth trace point. Here, the natural number n is a gridded locus T_GThe number of contained trace points.

Correspondingly, the number of track points contained in the track T of the vehicle running path is recorded as N. The recording time of each track point of the track path is taken out from the database and is arranged as o in sequence₁,o₂,…,o_N. Wherein o is₁Recording time, o, for the first track point of track T₂For recording the time of the first trace point, o_NThe recording time of the Nth track point is recorded. Thus, the total time to complete the trajectory T is o_N-o₁。

Step 620: according to the rectangular grid G, dividing each recording time into different segments in sequence, and enabling the recording time in each segment to be located in the same grid area of the rectangular grid G:

wherein it is provided that the recording time is divided into n segments, the first segment containing the time from the first recording time o₁To d₁A recording time

The second fragment comprises the second fragment from₁+1 recording time

To d₂A recording time

By analogy, the nth segment contains the segment from d_n-1+1 recording time

To the Nth recording time o_NAnd | represents a fragment space symbol; d₁、d₂、d_n-1N is a natural number | | | represents a fragment space character, d₁、d₂、d_n-1And n is a natural number.

The number of the track points is different because the driving paths of the vehicles are different. That is, to perform the scanning, the number N of trace points is different for different traces T. Correspondingly, gridding the trajectory T_GThe number n of the tracing points is different. In order to adapt to the situation of different track lengths, the invention will follow the gridding track T_GOne travel time is output in each grid area.

Step 630: calculating the driving time required by driving the current section i:

wherein the content of the first and second substances,

for the time of the i-th segment, corresponding to the feature vector f_iThe marking time of (1); i. j, d_i、d_i-1Are all natural numbers, and when i is equal to 1, d₀1 is ═ 1; when i is n, d_n+1＝N。

Step 640: and constructing a second training sample according to the feature vector and the driving time of each segment after driving:

where F represents the second training sample, the sequence of feature vectors F₁,f₂…,f_nFor the sample characteristics of the second training sample F, time sequence

Labeling the sample of the second training sample F; f. of₁Feature vector f of the first track point of the gridded track of the vehicle driving path₂Characteristic vector f of a second track point of the gridded track of the vehicle driving path_nThe characteristic vector of the nth track point of the gridding track of the vehicle running path is obtained;

to correspond to the feature vector f₁The time of the marking of (a) is,

to correspond to the feature vector f₂The time of the marking of (a) is,

to correspond to the feature vector f_nThe marking time of (1); the natural number n is the number of track points of the gridded track of the vehicle running path.

Step 650: and training an urban traffic travel time prediction network according to the second training sample, wherein the urban traffic travel time prediction network is used for predicting the travel time of the whole travel path.

As shown in fig. 3, the present invention uses a Long-Short term Memory (LSTM) unit as a standard information processing unit to design the main structure of the network.

For a training sample F, first, the feature vector F is₁Sending the signal into an LSTM unit corresponding to the first track point, and outputting a hidden state vector v₁Further, the hidden state vector v₁Input to a Fully-Connected neural Network (FCN). Note that this fully-connected neural network is an FCN unit. FCN unit with hidden state vector v₁For the input layer, there is no hidden layer, the output layer contains a node, and the excitation function of the output layer is a linear rectification function, i.e. the ReLU used in the step. At this time, the FCN unit outputs the predicted time of the first track segment, and the predicted time is recorded

Then, the feature vector f is processed₂And hidden state vector v₁Simultaneously sent to an LSTM unit corresponding to the second track point, and the unit outputs a hidden state vector v₂Further, the hidden state vector v₂The total running time of the first two segments after running is input into the FCN unit and output as the predicted time

By analogy, and finally, the feature vector f_nSending the data to an LSTM unit corresponding to the nth track point, outputting the total running time of the n segments before the running is finished, namely the running time of the whole track through an FCN unit, and recording the running time as the predicted time

It should be noted that the above process can be applied to training samples containing any number of trace points. The embodiment of the invention discloses a prediction network for urban traffic travel timeThe trace points which can be processed by the network are collected into all gridding traces T according to the training samples_GThe maximum number of track points involved.

In the above processing procedure, according to the calculation flow of the standard LSTM unit, the structure of the whole network can be determined only by determining the dimension of the hidden state vector of the LSTM unit. In the present invention, the dimension of the hidden state vector is set as the feature vector f₁,f₂…,f_nIs half the dimension of (i.e. the

Wherein]Representing a rounding operation. Here, Q is the number of continuous-time-interval traffic flow matrices output by the urban traffic flow prediction network in step 300.

In the whole urban traffic travel time prediction network, the structures of all LSTM units are the same, and the parameters are the same. In addition, the structure of all FCN units is the same, while the parameters are the same. The advantage of this design is that when the network is applied, it can be extended by copying the LSTM and FCN elements, and thus can be adapted to any length of the driving track.

Next, a loss function is defined. For gridded tracks T_GFor each trace point of (a), the fully-connected neural network P outputs a predicted time, and therefore a loss function needs to be defined. For the loss caused by sample F, it is calculated by the following loss function:

wherein loss (F) represents the loss caused by the sample F,

gridded trace T output for FCN unit_GThe predicted time of the first i segments of (c),

is the known time of marking of the sample F. Here, the natural number i takes on an integer ranging from 1 to n for the sample F.

And finally, training a urban traffic travel time prediction network. In the invention, a standard error back propagation algorithm is adopted for training the urban travel time prediction network. In this process, in the embodiment of the present invention, the learning rate is set to 0.001, and the batch data sets are set to 64, each for 20 rounds of training.

In step 700, determining the travel time required for the vehicle to run through the path to be measured according to the path to be measured of the vehicle to be measured based on the urban traffic flow prediction network and the urban travel time prediction network, specifically including:

step 710: calculating normalized floating point traffic flow matrixes in the first J continuous time intervals by taking the current running time of the vehicle to be detected as a terminal time interval, and stacking the matrixes into a three-dimensional tensor according to a time sequence:

A＝[Y₁,Y₂,…,Y_J]；

a is a three-dimensional tensor formed by three-dimensional stereo data of L rows, M columns and J layers, and is formed by stacking normalized floating point traffic flow matrixes obtained in the first J continuous time intervals with the current time interval as an end point according to a time sequence, and Y is₁For the normalized floating point traffic flow matrix, Y, obtained in the first J-1 time interval₂For the normalized floating point traffic flow matrix, Y, obtained in the first J-2 time intervals_JObtaining a normalized floating point traffic flow matrix in the current time; l is the number of rows in the latitude direction of the rectangular grid G, and M is the number of columns in the longitude direction of the rectangular grid G.

Step 720: according to the three-dimensional tensor A and an urban traffic flow prediction network, predicting normalized floating point traffic flow in Q time intervals in future: b is₁,B₂,…,B_Q；

Wherein, B₁Predicting future first time zone for urban traffic flow prediction networkNormalized floating point traffic flow within the interval; b is₂Normalizing the floating point traffic flow for a second future time interval predicted by the network; b is_QNormalized floating point traffic flow for the predicted future qth time interval for the network.

Step 730: determining a corresponding gridding track R according to a track sequence R of a running path to be tested of a vehicle to be tested_GFeature vectors of the track points;

sample feature B of trace sequence R: b ═ B₁,b₂,…,b_n)；

Wherein, the sample characteristic B of the running path R to be measured is composed of a characteristic vector sequence B₁,b₂,…,b_nComposition is carried out; b₁Is a gridded track R_GThe feature vector of the first trace point of (b)₂Is a gridded track R_GThe feature vector of the second trace point of (b)_nIs a gridded track R_GThe feature vector of the nth trace point; natural number n is gridding track R_GThe number of the contained track points;

and

respectively represent gridded traces R_GThe longitude and latitude of the ith track point of (1) are set as a gridding track R_GThe ith track point of

R is located at rectangular grid G_iRow and c_iThe grid area where the columns are located. Here, the natural number i takes on an integer ranging from 1 to n; natural number r_iIs an integer ranging from 1 to L, a natural number c_iIs an integer ranging from 1 to M;

The row serial numbers and the column serial numbers of the rectangular grids; the natural number Q is the number of time intervals predicted by the urban traffic flow prediction network, and the superscript T represents vector transposition;

step 740: according to the gridding track R_GAnd determining the time of the running path R to be tested of the vehicle to be tested after the vehicle to be tested runs.

Specifically, gridding the track R_GCharacteristic vector b of each track point₁,b₂,…,b_nAnd sequentially inputting the data into corresponding LSTM units, and operating the urban travel time prediction network. By the last fully-connected nerveThe network FCN unit outputs the time of the vehicle running path R to be predicted, namely the time predicted by the urban traffic flow prediction network.

The urban traffic travel time prediction method can predict the urban traffic travel time of a given path, and is mainly embodied in the following aspects:

1) the invention needs to train two networks, one network carries out real-time road condition information prediction, and the other network carries out urban traffic travel time prediction of a given path by utilizing the prediction result of the previous network. The frame design can consider real-time road conditions influencing urban traffic travel time.

2) A deep learning approach is used. Its excellent non-linear fitting ability has been demonstrated in many machine learning fields. Compared with the traditional method, the method has better effect on predicting the urban traffic travel time;

3) although the real-time traffic flow information is taken as a representative of additional factors influencing the urban traffic travel time, the framework of the invention can be easily applied to the situation that other factors are considered, and only partial modification of the network structure is needed. If the influence of dynamically changed weather information on urban traffic travel time is considered, the urban traffic flow prediction network is only required to be changed into a real-time weather prediction network. This increases the flexibility of the invention.

Experiments were conducted on taxi track data in collusion. To verify the effectiveness of the present invention, the figure shows two traces. Wherein the travel path R₁The sequence of traces of (a):

([104.048087,30.67257],[104.048688,30.667996],[104.046058,30.663864],[104.045486,30.663977],[104.041591,30.665894],[104.041185,30.665958],[104.045532,30.663658],[104.045747,30.660006],[104.046027,30.656351],[104.049302,30.653285],[104.052818,30.652634],[104.05679,30.649528],[104.060861,30.647632],[104.065253,30.647387],[104.068546,30.646833],[104.070377,30.646512],[104.074429,30.646343],[104.07896,30.645559],[104.082964,30.643429],[104.086483,30.642236],[104.088603,30.639872],[104.087352,30.63816],[104.08751,30.638336],[104.088682,30.639902],[104.088278,30.640154],[104.087855,30.639184],[104.092022,30.641449],[104.094963,30.638586],[104.094659,30.637947],[104.09227,30.637359],[104.089681,30.640921],[104.086951,30.638889],[104.08346,30.637202],[104.07876,30.634991],[104.074695,30.633712],[104.069852,30.633636],[104.064814,30.633495],[104.060504,30.633411],[104.055255,30.63328],[104.050522,30.63545],[104.04863,30.63635],[104.046449,30.637371],[104.042775,30.636046],[104.040166,30.631991],[104.037331,30.630241],[104.033926,30.633511],[104.030826,30.636792],[104.028128,30.639455],[104.025346,30.642233],[104.02234,30.641511],[104.018598,30.638672],[104.013982,30.63651],[104.01379,30.636413],[104.009396,30.634587],[104.006292,30.633066],[104.002995,30.631547],[103.999547,30.629926],[103.994709,30.627643],[103.990154,30.625509],[103.985626,30.623449],[103.981084,30.621406],[103.977224,30.61967],[103.975757,30.619052],[103.972676,30.617613],[103.970359,30.616433],[103.975089,30.618571],[103.979211,30.620433])。

travel route R₂The sequence of traces of (a) is as follows:

([104.107954,30.694667],[104.107533,30.692622],[104.106701,30.6916],[104.105737,30.69041],[104.104835,30.689299],[104.103889,30.688151],[104.102697,30.686618],[104.102828,30.68506],[104.103921,30.683158],[104.104684,30.681414],[104.10577,30.679865],[104.107494,30.677033],[104.108455,30.675276],[104.109234,30.673234],[104.110213,30.671347],[104.110638,30.670777],[104.111095,30.668693],[104.111861,30.666929],[104.112352,30.66585],[104.113085,30.663795],[104.113307,30.663361],[104.114126,30.661514],[104.114649,30.659982],[104.1149,30.657735],[104.114941,30.656178],[104.114834,30.654379],[104.114826,30.652352],[104.115693,30.650939],[104.117447,30.650225],[104.116893,30.648046],[104.118718,30.648134],[104.118882,30.64808],[104.118639,30.647728],[104.117693,30.647957],[104.116512,30.647634],[104.116169,30.646546],[104.115794,30.645291],[104.11504,30.643075],[104.11424,30.641665],[104.111912,30.642674],[104.110404,30.64393],[104.108867,30.641481],[104.107777,30.63992],[104.106731,30.63847],[104.105625,30.636528],[104.104823,30.635327],[104.104013,30.634062],[104.102934,30.632328],[104.101653,30.630342],[104.100493,30.628484],[104.100305,30.62732],[104.099455,30.626544],[104.098251,30.624969],[104.097108,30.623899],[104.095813,30.622056],[104.094057,30.621055],[104.091939,30.620839],[104.089401,30.620788],[104.087057,30.62065],[104.085334,30.620683],[104.083797,30.620698],[104.081851,30.620675],[104.080276,30.620622],[104.077718,30.620695],[104.076107,30.620588],[104.074552,30.620557],[104.07202,30.620715],[104.071711,30.622295],[104.071833,30.624357],[104.072982,30.625105],[104.073218,30.625109],[104.074141,30.625175],[104.076509,30.62515],[104.076778,30.622888],[104.076874,30.621773],[104.076716,30.624068],[104.0766,30.625956],[104.076441,30.628046],[104.076338,30.630266],[104.076184,30.632202],[104.076129,30.633659],[104.076051,30.635315],[104.075996,30.637499])。

travel route R₁Track data starting at 12 points 01 minutes on a certain day, travel route R₂For the trajectory data starting 13 minutes at 18 o' clock on a certain day, the real time of 3481 seconds after the red trajectory is traveled, and the real time of 2550 seconds after the blue trajectory is traveled. By using the method, the predicted time for driving the red track is 2975 seconds, and the predicted time for driving the blue track is 2861 seconds. Considering the complexity, uncertainty and traffic light of the road condition, the accuracy of the result is acceptable with reference to the current technical level in the industry.

Experiments show that the method can effectively predict the travel time of a given path. The method considers the influence of real-time road conditions on the travel time, and utilizes a deep learning method to construct a prediction model to obtain a prediction result with higher precision.

Preferably, the invention also provides a system for predicting urban traffic travel time, which can improve the accuracy of urban traffic travel time prediction.

As shown in fig. 4, the urban traffic travel time prediction system of the present invention includes a mesh division unit 1, a normalization processing unit 2, a first training unit 3, a simplification processing unit 4, a feature determination unit 5, a second training unit 6, and a time prediction unit 7.

Specifically, the grid dividing unit 1 is configured to divide a city to be measured into rectangular grids;

the normalization processing unit 2 is used for constructing a normalization floating point traffic flow matrix according to historical traffic flow data based on the rectangular grid;

the first training unit 3 is used for training an urban traffic flow prediction network according to the normalized floating point traffic flow matrix;

the simplification processing unit 4 is used for simplifying the track point sequence of the vehicle driving path according to the rectangular grid to obtain a gridded track of the vehicle driving path;

the characteristic determining unit 5 is used for determining characteristic vectors of all track points in the gridding track according to the normalized floating point traffic flow matrix and the gridding track of the vehicle running path;

the second training unit 6 is configured to train a prediction network of urban traffic travel time according to the feature vectors of the track points in the gridding track;

the time prediction unit 7 is configured to determine, according to a to-be-detected travel path of a to-be-detected vehicle, travel time required for the to-be-detected vehicle to travel through the to-be-detected path, based on the urban traffic flow prediction network and the urban travel time prediction network.

In addition, the invention also provides a system for predicting the urban traffic travel time, which comprises the following components:

a processor; and

dividing a city to be tested into rectangular grids;

Further, the present invention also provides a computer-readable storage medium storing one or more programs that, when executed by an electronic device including a plurality of application programs, cause the electronic device to perform operations of:

dividing a city to be tested into rectangular grids;

Compared with the prior art, the urban traffic travel time prediction system and the computer-readable storage medium have the same beneficial effects as the urban traffic travel time prediction method, and are not described again here.

So far, the technical solutions of the present invention have been described in connection with the preferred embodiments shown in the drawings, but it is easily understood by those skilled in the art that the scope of the present invention is obviously not limited to these specific embodiments. Equivalent changes or substitutions of related technical features can be made by those skilled in the art without departing from the principle of the invention, and the technical scheme after the changes or substitutions can fall into the protection scope of the invention.

Claims

1. A prediction method for urban traffic travel time is characterized by comprising the following steps:

dividing a city to be tested into rectangular grids;

training an urban traffic flow prediction network according to the normalized floating point traffic flow matrix; the method specifically comprises the following steps:

P_t＝(A_t；B_t)，

wherein t represents a time interval number, P_tRepresenting a first training sample; a. the_tRepresenting a first training sample P_tThe sample characteristic is a three-dimensional tensor of L rows, M columns and J layers, and a normalized traffic flow matrix obtained in the first J continuous time intervals with the t-th time interval as an end point is piled in a time sequenceAnd (3) stacking: a. the_t＝[Y_t-J+1,Y_t-J+2,…,Y_t-1,Y_t]；

Wherein, Y_t-J+1Is a normalized floating point traffic flow matrix, Y, obtained in the t-J +1 time interval_tFor the normalized floating-point traffic flow matrix, Y, obtained in the t-th time interval_t+QObtaining a normalized floating point traffic flow matrix in the t + Q time intervals;

training an urban traffic flow prediction network according to the first training sample; the urban traffic flow prediction network is used for predicting the normalized floating point traffic flow of a rectangular grid G in Q continuous time intervals;

simplifying the track point sequence of the vehicle running path according to the rectangular grid to obtain a gridded track of the vehicle running path; the method specifically comprises the following steps:

The second fragment comprises the second fragment from₁+1 track points

To d₂One track point

By analogy, the nth segment contains the segment from d_n-1+1 track points

Wherein the content of the first and second substances,

mean value, x, representing the latitude of a tracing point belonging to the ith segment_jAnd y_jAre respectively asLongitude and latitude coordinates, when i is 1, d₀1 is ═ 1; when i is n, d_n＝N；

Determining the characteristic vector of each track point in the gridding track according to the normalized floating point traffic flow matrix and the gridding track of the vehicle running path; the method specifically comprises the following steps:

according to the gridding track T_GThe ith track point of

wherein the content of the first and second substances,

The row serial numbers and the column serial numbers of the rectangular grids;

Wherein n is a gridding track T_GThe number of the track points; superscript T represents vector transposition;

training a prediction network of urban traffic travel time according to the feature vectors of all track points in the gridding track; the method specifically comprises the following steps:

The second fragment comprises the second fragment from₁+1 recording time

To d₂A recording time

By analogy, the nth segment contains the segment from d_n-1+1 recording time

calculating the driving time required by driving the current section i:

wherein the content of the first and second substances,

to correspond to the feature vector f_nThe marking time of (1);

training an urban traffic travel time prediction network according to the second training sample, wherein the urban traffic travel time prediction network is used for predicting the travel time of the whole travel path;

determining the travel time required by the vehicle to be tested to finish running the path to be tested according to the running path to be tested of the vehicle to be tested on the basis of the urban traffic flow prediction network and the urban travel time prediction network; the method specifically comprises the following steps:

A＝[Y₁,Y₂,…,Y_J]；

according to the three-dimensional tensor A and an urban traffic flow prediction network, predicting normalized floating point traffic flow in Q time intervals in future: b is₁,B₂,…,B_Q；

sample feature B of trace sequence R: b ═ B₁,b₂,…,b_n)；

and

from said-gridded footprint R_GAnd determining the time of the running path R to be tested of the vehicle to be tested after the vehicle to be tested runs.

2. The method for predicting urban traffic travel time according to claim 1, wherein the constructing a normalized floating-point traffic flow matrix according to historical traffic flow data based on the rectangular grid specifically comprises:

wherein the content of the first and second substances,

3. A system for predicting urban traffic travel time, the system comprising:

the first training unit is used for training an urban traffic flow prediction network according to the normalized floating point traffic flow matrix; the method specifically comprises the following steps:

P_t＝(A_t；B_t)，

B_tRepresenting a first training sample P_tIs a three-dimensional tensor of L rows, M columns and Q layers, and is marked by the t +1 thThe normalized traffic flow matrixes obtained in Q continuous time intervals with the time intervals as the starting points are stacked according to the time sequence: b is_t＝[Y_t+1,Y_t+2,…,Y_t+Q]；

the simplification processing unit is used for simplifying the track point sequence of the vehicle driving path according to the rectangular grid to obtain a gridded track of the vehicle driving path; the method specifically comprises the following steps:

wherein it is assumed that the trajectory T will be divided into n segments, the firstA segment contains a point (x) from the first track₁,y₁) To d₁One track point

The second fragment comprises the second fragment from₁+1 track points

To d₂One track point

By analogy, the nth segment contains the segment from d_n-1+1 track points

Wherein the content of the first and second substances,

The characteristic determining unit is used for determining the characteristic vector of each track point in the gridding track according to the normalized floating point traffic flow matrix and the gridding track of the vehicle running path; the method specifically comprises the following steps:

according to the gridding track T_GThe ith track point of

wherein the content of the first and second substances,

The row serial numbers and the column serial numbers of the rectangular grids;

the second training unit is used for training the urban traffic travel time prediction network according to the feature vectors of all track points in the gridding track; the method specifically comprises the following steps:

The second fragment comprises the second fragment from₁+1 recording time

To d₂A recording time

By analogy, the nth segment contains the segment from d_n-1+1 recording time

To the Nth recording time o_NAnd | represents a fragment space symbol; d₁、d₂、d_n-1N is a natural number | | | represents a fragment space character, d₁、d₂、d_n-1N is naturalCounting;

calculating the driving time required by driving the current section i:

wherein the content of the first and second substances,

to correspond to the feature vector f_nThe marking time of (1);

the time prediction unit is used for determining the travel time required by the vehicle to be detected to run through the path to be detected according to the to-be-detected running path of the vehicle to be detected based on the urban traffic flow prediction network and the urban travel time prediction network; the method specifically comprises the following steps:

A＝[Y₁,Y₂,…,Y_J]；

sample feature B of trace sequence R: b ═ B₁,b₂,…,b_n)；

and

4. A system for predicting urban traffic travel time, comprising:

a processor; and

dividing a city to be tested into rectangular grids;

P_t＝(A_t；B_t)，

The second fragment comprises the second fragment from₁+1 track points

To d₂One track point

By analogy, the nth segment contains the segment from d_n-1+1 track points

Wherein the content of the first and second substances,

according to the gridding track T_GThe ith track point of

wherein the content of the first and second substances,

The row serial numbers and the column serial numbers of the rectangular grids;

The second fragment comprises the second fragment from₁+1 recording time

To d₂A recording time

By analogy, the nth segment contains the segment from d_n-1+1 recording time

calculating the driving time required by driving the current section i:

wherein the content of the first and second substances,

to correspond to the feature vector f_nThe marking time of (1);

A＝[Y₁,Y₂,…,Y_J]；

sample feature B of trace sequence R: b ═ B₁,b₂,…,b_n)；

and

5. A computer-readable storage medium storing one or more programs that, when executed by an electronic device including a plurality of application programs, cause the electronic device to:

dividing a city to be tested into rectangular grids;

P_t＝(A_t；B_t)，

Wherein, Y_t-J+1Is a normalized floating point traffic flow matrix, Y, obtained in the t-J +1 time interval_tFor normalized floating point obtained in the t-th time intervalTraffic flow matrix, Y_t+QObtaining a normalized floating point traffic flow matrix in the t + Q time intervals;

The second fragment comprises the second fragment from₁+1 track points

To d₂One track point

By analogy, the nth segment contains the segment from d_n-1+1 track points

Wherein the content of the first and second substances,

taking out continuous Q normalized floating point traffic flow matrixes Y from the normalized floating point traffic flow matrix_p+1,Y_p+2,…,Y_p+Q(ii) a Wherein p is a vehicleFirst track point (x) of travel path T₁,y₁) The sequence number of the time interval to which the time belongs in the training time interval;

according to the gridding track T_GThe ith track point of