CN109035761B - Travel time estimation method based on auxiliary supervised learning - Google Patents

Travel time estimation method based on auxiliary supervised learning Download PDF

Info

Publication number
CN109035761B
CN109035761B CN201810658375.5A CN201810658375A CN109035761B CN 109035761 B CN109035761 B CN 109035761B CN 201810658375 A CN201810658375 A CN 201810658375A CN 109035761 B CN109035761 B CN 109035761B
Authority
CN
China
Prior art keywords
time
neural network
grid
vector
training
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Expired - Fee Related
Application number
CN201810658375.5A
Other languages
Chinese (zh)
Other versions
CN109035761A (en
Inventor
孙未未
章瀚元
吴昊
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Fudan University
Original Assignee
Fudan University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Fudan University filed Critical Fudan University
Priority to CN201810658375.5A priority Critical patent/CN109035761B/en
Publication of CN109035761A publication Critical patent/CN109035761A/en
Application granted granted Critical
Publication of CN109035761B publication Critical patent/CN109035761B/en
Expired - Fee Related legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G08—SIGNALLING
    • G08G—TRAFFIC CONTROL SYSTEMS
    • G08G1/00—Traffic control systems for road vehicles
    • G08G1/01—Detecting movement of traffic to be counted or controlled
    • G08G1/0104—Measuring and analyzing of parameters relative to traffic conditions
    • G08G1/0125—Traffic data processing
    • G08G1/0129—Traffic data processing for creating historical data or processing based on historical data
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G08—SIGNALLING
    • G08G—TRAFFIC CONTROL SYSTEMS
    • G08G1/00—Traffic control systems for road vehicles
    • G08G1/01—Detecting movement of traffic to be counted or controlled
    • G08G1/0104—Measuring and analyzing of parameters relative to traffic conditions
    • G08G1/0137—Measuring and analyzing of parameters relative to traffic conditions for specific applications

Landscapes

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

Abstract

The invention belongs to the technical field of intelligent transportation, and particularly relates to a travel time estimation method based on auxiliary supervised learning. The method comprises the steps of searching a statistical rule from mass historical track data, and integrally estimating the time of the whole travel through an end-to-end deep learning model; the method comprises the following steps: a characteristic extraction and representation stage, namely preprocessing the track data, and respectively extracting time and space characteristics, driving state characteristics and short-time and long-time traffic condition characteristics of the track data; training and predicting stage, training and predicting the extracted features with a uniform bidirectional circulation neural network; the time cost of passing through the current small area is output by each step of the recurrent neural network; the sum of the time cost of these small areas is the time cost of the total path. Meanwhile, a bidirectional interval loss function is also introduced to restrain the intermediate time overhead. The method can efficiently and accurately estimate the vehicle travel time in the city, and has a good effect in an actual environment.

Description

Travel time estimation method based on auxiliary supervised learning
Technical Field
The invention belongs to the technical field of intelligent transportation, and particularly relates to a travel time estimation method based on auxiliary supervised learning.
Background
Travel time estimation is an essential important technology in the field of urban traffic, can provide help for travel and commute of people, and can also provide support for government planning decisions. However, this is not a simple and small problem, but is influenced by various dynamic factors, such as traffic dynamics, intersection conditions, changes in driver driving behavior, and historical periodic data evolution. These factors contribute to uncertainty and difficulty in the travel time estimate. With the development and popularization of mobile devices supporting GPS, a great amount of trajectory data is continuously generated and covers all corners of a city. With the massive historical track data, the internal rules behind the data can be mined, and the period and the trend of the change of the travel time are learned by constructing an algorithm model, so that the time overhead required by the current query track can be accurately inferred.
Most of the existing methods adopt a divide-and-conquer (divide-and-conquer) method, mainly by decomposing a path into a series of segments or sub-paths.
(1) Single-road-segment-based methods:
the method based on the single road section mainly estimates the average speed of each single road section when the track passes through, calculates the passing average time cost according to the length of the road section, and finally accumulates the time sum of each road section to obtain the total time. This approach does not take into account the intersection time overhead between road segments. In addition, such estimation relies heavily on high quality velocity data, which is often not available in the trajectory data.
(2) The method based on the sub-path comprises the following steps:
the sub-path based approach mainly divides the path into a series of sub-paths, so that the time overhead of the intersection is also considered. The main idea is to splice and mine rich public sub-path information in historical data. Although this approach may overcome many of the drawbacks of the single-route approach, it is still based on heuristic design rather than directly targeting travel time as an algorithm optimization.
In summary, the methods available today fail to achieve satisfactory accuracy for two reasons. One is that they do not see the path as a whole, but split into sub-blocks. During this splitting process, much useful information is lost. And they do not take full advantage of the intermediate surveillance tags that are unique to the trace data, i.e., the time stamp information for each intermediate GPS sample point. On the other hand, with the development and prosperity of deep learning technology, more problems can be solved in an end-to-end integrated manner, and the method is more efficient compared with the traditional heuristic model. In addition, deep learning has strong characterization capability, and compared with a manual model, the deep learning can capture more potential features and can process various complex dynamics in the travel estimation problem.
Disclosure of Invention
The invention aims to provide a travel time estimation method based on a history track of auxiliary supervision learning aiming at the limitations of the traditional two types of travel time estimation technologies so as to overcome the defects of the prior art.
The method searches for a statistical rule from massive historical track data, and integrally estimates the time of the whole travel through an end-to-end deep learning model. The method comprises the following basic steps: a characteristic extraction and representation stage, wherein the track data is preprocessed, and various characteristics of the track data are respectively extracted; training and predicting stage, training and predicting the extracted features with one unified bidirectional circulating neural network; the time cost of passing through the current small area is output by each step of the recurrent neural network; the sum of the time overhead of the small areas is the time overhead of the total path; for more efficient training, a two-way interval loss function is also introduced to constrain the intermediate time overhead.
The invention provides a travel time estimation method of a historical track based on auxiliary supervised learning, which comprises the following three stages:
and (I) a characteristic extraction and representation stage, namely preprocessing historical track data and extracting various characteristics (including time characteristics, space characteristics, driving state characteristics, short-time and long-time traffic condition characteristics and the like) of the historical track data. The method comprises the following specific steps:
and (1) carrying out fine-grained division on the grids according to longitude and latitude coordinates in the urban range to form adjacent rectangular small areas. And mapping each coordinate point in the track sequence consisting of the GPS coordinates into a corresponding small area according to the chronological sequence to form a sequence consisting of grid coordinates. For the condition that the adjacent track points are far away and fall into discontinuous small areas, the intermediate passing path can be obtained by algorithms such as map matching and the like, and the discontinuous area information is supplemented.
And (2) mining the characteristics of different aspects of each grid. First, latent semantic information is mined using an embedded vector technique. The embedded vector technology is widely used in the fields of natural language processing, social networks and the like, mainly utilizes low-dimensional real number vectors to represent semantic information of each word or thing, and measures the corresponding relation between real objects through the distance relation in a vector space. The invention utilizes an embedded vector technology to represent semantic information of each grid small region in different spaces and different time periods. The information includes spatial location information of different functional areas (such as residential areas, commercial areas or industrial areas, etc.) of the city, and also includes time information of morning rush hours, weekends, etc. In particular, the space vector V of each mesh is represented by a low-dimensional vectorspDividing a day into a plurality of time buckets (e.g., one bucket per hour), each track obtaining a time vector V according to a specific falling time buckettp. To VspAnd VtpA random initialization is performed followed by training with the model as it is trained.
And (3) when a driver drives the vehicle, the driving speed and the driving behavior can be changed in different driving states. For example, the vehicle may be more inclined to travel on a road or overhead in the middle of the travel path, where the speed may be faster. In addition, theWhen the vehicle starts or arrives at the end point, the speed tends to be slow due to the fact that the vehicle runs on a small road or an area with many people. In particular, a four-dimensional vector V is useddriTo indicate whether the current driving phase is a departure phase, a midway phase, or an end phase, and the proportion of driving that has been performed in each phase. For example, VdriThe value of (1,0,0,0.2) indicates that the driver is driving at the beginning stage, and accounts for 20% of the total travel.
And (4) the traffic condition in one area is changed periodically and regularly along with the time evolution. For example, if a road segment is very congested from 8 o ' clock to 8 o ' clock, 8 o ' clock 35 may also be very congested. That is, the traffic condition information in the past in a short time is useful for predicting the current traffic state. Defining the short-time traffic condition as Vshort. Meanwhile, the long-time periodic traffic condition change can also help to predict the current traffic condition, such as the change rule of the traffic condition on weekdays and weekends. Defining the long-term traffic condition as Vlong. In particular, the present invention relates to a method for producing,
defining:
Figure BDA0001706094160000031
indicating the current small area g in the past jth time intervaliIn which v isjRepresenting historical average speed, njIndicates the amount of historical track data, leni/vjRepresenting the roughly estimated transit time. The traffic condition features are input into a sub-cyclic neural network according to the historical time sequence, and the traffic condition features can be extracted.
In addition, since the historical data are unevenly distributed in different spatial regions, the tracks of some regions pass through a small number of regions, which may affect the accuracy of estimation. To solve this data sparseness problem, traffic condition information of neighboring small areas is also taken into account, i.e., traffic condition information of neighboring small areas is taken into account
Defining:
Figure BDA0001706094160000032
represents the distance giAnd collecting the traffic condition characteristics of the grids with the distance not exceeding d in the past short time, and inputting the collected traffic condition characteristics into the neural network together. Wherein x, y represent the coordinates of the grid, gjRepresents except giOther meshes than mesh.
(II) a training stage, inputting the features extracted from the historical track data into a uniform bidirectional recurrent Neural network (bidirectional LSTM, reference: Graves A, Schmidhuber J. frame with a phosphor classification with bidirectional LSTM and other Neural network architecture [ J ]. Neural Networks,2005,18(5-6):602 + 610) for training, and taking a bidirectional interval loss function as the constraint of training; the method comprises the following specific steps:
and (1) constructing a recurrent neural network. Defining a network hidden layer as
Figure BDA0001706094160000033
Input data as
Figure BDA0001706094160000034
Then, the input data of the t step is xtAnd the calculation result obtained in the t step is htThen, there are:
ht＝φ(xt·Wx+ht-1·Wh+b) (3)
wherein,
Figure BDA0001706094160000035
is a weight matrix (weight matrix) of the input data,
Figure BDA0001706094160000036
is a weight matrix of the hidden layer(s),
Figure BDA0001706094160000037
is the bias parameter (bias). Phi denotes a non-linear activation function, which may be a sigmoid function, a ReLU function, a tanh function, etc.
That is, the hidden state can be expressed as a function:
ht＝f(ht-1,xt) (4)
on this basis, define forgetting the door to be:
ft＝σ(Wf·[ht-1,xt]+bf) (5)
the input gates are:
it＝σ(Wi·[ht-1,xt]+bi) (6)
the output gate is:
ot＝σ(Wo[ht-1,xt]+bo) (7)
the memory cell is updated by:
Figure BDA0001706094160000041
Figure BDA0001706094160000042
the hidden layer is updated as follows:
ht＝Ot·tanh(Ct) (10)
wherein, Wf、Wi、Wo、
Figure BDA0001706094160000043
Weight matrices representing forgetting gate, input gate, output gate and memory cell, respectively, bf、bi、bo、
Figure BDA0001706094160000044
Then it is the corresponding bias parameter. σ () is a non-linear activation function, e.g.
Figure BDA0001706094160000045
Is a function of the sigmoid and is,
Figure BDA0001706094160000046
is a hyperbolic tangent function, and f () represents an abstract neural network function containing parameters of each layer. Defining the corresponding parameter of the recurrent neural network as WN(ii) a From [ - α, α]Wherein α is a hyper-parameter, and is set to be in a range of 0.01 to 1.
The bidirectional recurrent neural network performs calculation by using a forward recurrent neural network and a reverse recurrent neural network simultaneously. The forward circulation neural network inputs the grid features extracted in the previous step according to the sequence of the sequence, and the reverse circulation neural network inputs the grid features after the sequence is in the reverse order. This has the advantage that the neural network can be made to observe the position distances of the current mesh from the start point and the end point simultaneously, thereby having an overall characteristic. Defining its hidden variables as a concatenation of forward and reverse networks
Figure BDA0001706094160000047
Wherein
Figure BDA0001706094160000048
Represents a hidden layer of the forward recurrent neural network,
Figure BDA0001706094160000049
representing a hidden layer of the inverse recurrent neural network.
Step (2), extracting the characteristics, namely the spatial characteristics, in the historical track data
Figure BDA00017060941600000410
Temporal characteristics
Figure BDA00017060941600000411
Driving state characteristics
Figure BDA00017060941600000412
Historically short and long term traffic status characteristics
Figure BDA00017060941600000413
And
Figure BDA00017060941600000414
splicing into a unified feature vector:
Figure BDA00017060941600000415
at each small grid passing through, the two-way recurrent neural network is input to obtain the passing time of the grid, namely WT·hi+ b. Time overhead of total trip
Figure BDA00017060941600000416
Comprises the following steps:
Figure BDA00017060941600000417
definition of
Figure BDA0001706094160000051
Respectively, a weight matrix and bias parameters for calculating the total time overhead. WTRepresenting the transpose of the W matrix.
And (3) defining the real time overhead vector of the track passing through each grid sequence as T. The sequential real time overhead vector is TfThe real time overhead vector of the reverse order is Tb. The time overhead vector estimated by the neural network is:
Figure BDA0001706094160000052
Figure BDA0001706094160000053
and the model is subjected to auxiliary supervised learning by using the bidirectional interval loss function, so that the time overhead of the whole path can be learned, and the transit time of each intermediate stage can be learned. The two-way interval loss function is defined as:
Figure BDA0001706094160000054
where M denotes whether a trajectory passes through a mask of a small region, [ ] denotes an operation between each element of the vector.
Step (4), the training goal is to minimize the loss function L, i.e.:
Figure BDA0001706094160000055
where θ represents the training parameters of the model, ε represents the embedding vector in time and space, and S is the size of the training set. Finally, the model is subjected to parameter updating and optimization by using a time sequence-based back propagation algorithm. Back propagation algorithm references: chauvin Y, Rumelhart D E.Backproperation the term, architecture, and applications [ M ] Psychology Press,2013.
A prediction stage, deducing the features extracted from the query path by using a bidirectional cyclic neural network and estimating the travel time; the method comprises the following specific steps:
and (1) giving a real travel without a timestamp mark as a query path, and obtaining a grid sequence mapped by the real travel according to a passed actual path. For each small grid passing through, using space-time characteristics V obtained by characteristic extraction and expression stage extractionspAnd VtpDriving state characteristic VdriAnd historical short and long term traffic condition characteristics VshortAnd VlongV is indicated as a total feature of the mesh. Wherein the embedded vector of the spatio-temporal features uses vector information updated by parameters of a training process. Short-term and long-term traffic state features are feature mined using a trained sub-cyclic neural network.
Step (2), inputting the extracted characteristics of all aspects into the trained bidirectional cyclic neural network in each passing grid to obtain the current hidden variable htThen the estimated time to pass through the current region is WT·ht+ b. The total time overhead estimate is:
Figure BDA0001706094160000056
where n represents the total number of grids passed,
Figure BDA0001706094160000057
for the trained results, a weight matrix and bias parameters for the total time overhead are calculated. WTRepresenting the transpose of the W matrix.
Overall, the process of the invention has several advantages. Firstly, an end-to-end (end-to-end) deep learning method based on historical data training is utilized to directly learn the characteristics of the whole path and estimate the whole transit time. A bidirectional interval loss function is defined, and the time overhead of passing through an intermediate road section can be monitored in an auxiliary mode on the basis of monitoring the whole path time. The method for introducing the auxiliary supervision enriches the sample information of the path and can ensure that the propagation signal can be more accurate when the back propagation algorithm updates the parameters. Secondly, a feature extraction structure is provided, and the passing time of the path can be effectively estimated by extracting dynamic features of different dimensions such as space-time embedded vectors, driving states, short-time and long-time traffic conditions and the like. Finally, the experimental result is better than that of the existing method through experimental verification in the actual environment.
As shown in table 1, we performed experiments with real historical trajectory data, including two cities, boehr diagram and shanghai. The existing methods such as a road section average time method, a sub-path dynamic programming method, a grid full-connection network method, a grid convolution network method and the like are used for comparison. The road section average time method obtains a result by directly accumulating the average passing time of each road section. Dynamic programming of subpaths [ Yilun Wang, Yu Zheng, and Yexing Xue. travel time estimation of a path using mapping project-ries. in Proceedings of the 20th International Conference on Knowledge Discovery and Data Mining (SIGKDD), pages 25-34,2014 ] uses dynamic programming to find the optimal splicing method of subpaths. The grid full-connection Network method and the grid convolution Network method take an NxN whole grid as input, and respectively carry out optimization and estimation by using a full-connection Network (Multi-Layer Perceptin) and a convolution Neural Network (Convolutional Neural Network). We use three error measurement indexes of MAE, RMSE and MAPE to measure the quality of the method.
Figure BDA0001706094160000061
Figure BDA0001706094160000062
Figure BDA0001706094160000063
Wherein, y represents the true value,
Figure BDA0001706094160000064
representing the estimated value, and n represents the total number of samples. As can be seen from the results in Table 1, the method of the present invention is far better than the prior comparative methods in each index. For example, on the Shanghai data set, the method of the present invention estimates MAE error to be only 126 seconds, with MAPE error of 13.3%, whereas the best method has MAE error of 168 seconds and MAPE error of 19.1%.
TABLE 1
Figure BDA0001706094160000065
Drawings
Fig. 1 shows a real trace sample containing time stamp information for each intermediate GPS trace point, for a total of 720 s.
Fig. 2 shows a path sample that needs to be queried, and only contains specifically-passed path information, and does not contain any timestamp information.
Detailed Description
The following describes the specific implementation process of the present invention with reference to specific examples:
the historical trajectories as in fig. 1 are used for training and the travel times in fig. 2 are estimated.
The method comprises a preprocessing stage and a feature extraction and representation stage, wherein the track data are preprocessed to extract various features of the track data. Taking fig. 1 as an example, the specific steps are as follows:
(1) in the city range, fine-grained grid division is carried out and divided into small areas which are adjacent to each other. As in fig. 1, the map is divided into 5 × 6 meshes. Mapping each coordinate point in the track sequence into a corresponding small area to form a grid sequence, namely g ═ g { (g } {1,g2,…,g10}。
(2) For each mesh, features of different aspects of it are mined. For example, for g1Using random vectors
Figure BDA0001706094160000071
And
Figure BDA0001706094160000072
to represent spatiotemporal semantic information. Namely:
Figure BDA0001706094160000073
Figure BDA00017060941600000711
second, using a four-dimensional vector
Figure BDA0001706094160000074
To indicate whether the current driving phase is a departure phase, a midway phase, or an end phase, and the proportion of the driving that has been performed in each phase, namely:
Figure BDA0001706094160000075
finally, the past short and long time traffic condition information is used to predict the current traffic condition characteristics
Figure BDA0001706094160000076
And
Figure BDA0001706094160000077
in particular, define
Figure BDA0001706094160000078
For the past 1 st to 6 th time interval (5min), the current region g1The traffic situation of (1). For example
Figure BDA0001706094160000079
It is shown that the historical average speed is 10m/s, 8 historical tracks are in total, and the average transit time is estimated to be 20 m/s. Will be provided with
Figure BDA00017060941600000710
Inputting the traffic condition characteristics into a subcircular neural network according to historical time sequence, and finally outputting hidden vector h6As a characteristic of traffic conditions.
II, a training stage, which comprises the following specific steps:
(1) and establishing a Bi-directional circulating neural network (Bi-directional LSTM) model. And randomly initializing various parameters of the model, including matrix parameters and bias parameters of a forgetting gate, an input gate and an output gate.
(2) Extracting the feature in the historical track data, namely the spatial feature VspTime characteristic VspDriving state characteristic VdriHistorically short and long time traffic status features VshortAnd VlongAnd splicing the two into a uniform feature vector. With a grid g1For example, as
Figure BDA0001706094160000081
(3) At each small grid passing through, the two-way recurrent neural network is input to obtain the passing time of the grid, namely WT·hi+ b. The total travel time overhead is:
Figure BDA0001706094160000082
for example, W is defined as (0.1, 0.3., 0.7), and 10 mesh hidden layer variable values are defined as h1＝(0.8,0.3,…,0.2),…,h10When the offset value b is equal to 0.7, (0.7, 0.4.. 0.5) then:
Figure BDA0001706094160000083
(4) the real time cost vector of the trajectory through each grid sequence is defined as T. The sequential real time overhead vector is TfTrue time cost vector in reverse order of (70,120, …,720) is Tb(720,640, …, 50). The time overhead vector estimated by the neural network is:
Figure BDA0001706094160000084
Figure BDA0001706094160000085
and the model is subjected to auxiliary supervised learning by using the bidirectional interval loss function, so that the time overhead of the whole path can be learned, and the transit time of each intermediate stage can be learned. The two-way interval loss function is defined as:
Figure BDA0001706094160000086
where M denotes whether a trajectory passes through a mask of a small region, [ ] denotes an operation between each element of the vector.
(5) Minimize the loss function L, i.e.:
Figure BDA0001706094160000087
where θ represents the training parameters of the model, ε represents the embedding vector in time and space, and S is the size of the training set. Finally, the model is subjected to parameter updating and optimization by using a time sequence-based back propagation algorithm.
The prediction stage comprises the following specific steps (taking fig. 2 as an example):
(1) giving a real trip without timestamp marking as query path g ═ g1,g2,…,g8And obtaining a grid sequence mapped by the actual path according to the actual path. For each small grid g of passes1～g8Using spatio-temporal features V extracted in the feature extraction and representation stagesspAnd VtpDriving state characteristic VdriAnd historical short and long term traffic condition characteristics VshortAnd VlongV is indicated as a total feature of the mesh. Wherein the embedded vector of the spatio-temporal features uses vector information updated by parameters of a training process. Short-term and long-term traffic state features are feature mined using a trained sub-cyclic neural network.
(2) Inputting the extracted various aspects of characteristics into the trained bidirectional cyclic neural network in each passing grid to obtain the current hidden variable htThen the estimated time to pass through the current region is WT·ht+ b. The total time overhead estimated value is;
Figure BDA0001706094160000091
where both W and b are parameters obtained from previous training procedures.

Claims (1)

1. A travel time estimation method based on auxiliary supervised learning is characterized by comprising three stages:
the method comprises the steps of (I) feature extraction and representation, wherein historical track data are preprocessed, and various features of the historical track data are extracted;
in the training stage, the features extracted from the historical track data are input into a uniform bidirectional cyclic neural network for training, and a bidirectional interval loss function is used as the constraint of training;
a prediction stage, deducing the features extracted from the query path by using a bidirectional cyclic neural network and estimating the travel time;
the specific steps of the characteristic extraction and representation stage are as follows:
step (1), in an urban area, carrying out fine-grained division on grids according to longitude and latitude coordinates to form adjacent rectangular small areas; mapping each coordinate point in a track sequence consisting of historical GPS coordinates sequenced according to a time sequence into a corresponding small region to form a sequence consisting of grid coordinates;
step (2), for each grid, excavating the characteristics of different aspects of the grid; firstly, representing semantic information of each grid small region in different spaces and different time periods by using an embedded vector technology; the information comprises spatial location information of different functional areas of a city, and also comprises information of early peak, weekend time; in particular, the space vector V of each mesh is represented by a low-dimensional vectorspDividing a day into a plurality of time buckets, and obtaining a time vector V according to the time bucket in which each track fallstp(ii) a To VspAnd VtpCarrying out random initialization, and then training the model along with the model during model training;
step (3) of using the four-dimensional vector VdriTo indicate whether the current driving phase is a starting phase, a midway phase or an ending phase, and the proportion of driving in each phase;
step (4), defining the short-time traffic condition characteristic as VshortDefining a long-term traffic condition characteristic as VlongIn the case of a liquid crystal display device, in particular,
definition of：
Figure FDA0002929081460000011
Indicating the current small area g in the past jth time intervaliIn which v isjRepresenting historical average speed, njIndicates the amount of historical track data, leni/vjRepresenting a coarse estimated transit time; inputting the traffic condition characteristics into a sub-cyclic neural network according to a historical time sequence so as to extract the traffic condition characteristics;
in addition, traffic condition information of adjacent small areas is taken into account, i.e.
Defining:
Figure FDA0002929081460000012
represents the distance giCollecting the traffic condition characteristics of the grid set with the distance not exceeding d in the past short time, and inputting the traffic condition characteristics into the neural network together; wherein x, y represent the coordinates of the grid, gjRepresents except giOther meshes than the mesh;
the second training stage comprises the following specific steps:
step (1), constructing a recurrent neural network; defining a network hidden layer as
Figure FDA0002929081460000021
Input data as
Figure FDA0002929081460000022
Then, the input data of the t step is xtAnd the calculation result obtained in the t step is htThen, there are:
ht＝φ(xt·Wx+ht-1·Wh+b) (3)
wherein,
Figure FDA0002929081460000023
is a weight matrix of the input data,
Figure FDA0002929081460000024
is a weight matrix of the hidden layer(s),
Figure FDA0002929081460000025
is a bias parameter;
i.e. the hidden state is represented as a function:
ht＝f(ht-1，xt) (4)
on this basis, define forgetting the door to be:
ft＝σ(Wf·[ht-1，xt]+bf) (5)
the input gates are:
it＝σ(Wi·[ht-1，xt]+bi) (6)
the output gate is:
ot＝σ(Wo[ht-1，xt]+bo) (7)
the memory cell is updated by:
Figure FDA0002929081460000026
Figure FDA0002929081460000027
the hidden layer is updated as follows:
ht＝Ot·tanh(Ct) (10)
wherein,
Figure FDA0002929081460000028
weight matrices respectively representing the forgetting gate, the input gate, the output gate and the memory cell,
Figure FDA0002929081460000029
then it is the corresponding bias parameter; σ () is a nonlinear activation function; f () represents an abstract neural network function containing parameters of each layer, and the corresponding parameter of the recurrent neural network is defined as WNFrom [ - α, α]Initializing each element in the uniform distribution, wherein alpha is a hyper-parameter and is set to be in a range of 0.01 to 1;
the bidirectional cyclic neural network simultaneously uses the forward cyclic neural network and the reverse cyclic neural network for calculation; the forward circulation neural network inputs the grid features extracted in the previous step according to the sequence of the sequence in sequence, and the reverse circulation neural network inputs the grid features after the sequence is in reverse order; its hidden variables are defined as the concatenation of the forward and reverse networks, i.e.:
Figure FDA00029290814600000210
step (2), extracting the feature in the historical track data, namely the spatial feature VspTime characteristic VtpDriving state characteristic VdriHistorically short and long time traffic status features VshortAnd VlongAnd splicing into a unified feature vector:
V＝(Vsp，Vtp，Vdri，Vshort，Vlong) (11)
at each small grid passing through, the two-way recurrent neural network is input to obtain the passing time of the grid, namely WT·hi+ b, total travel time overhead
Figure FDA0002929081460000031
Comprises the following steps:
Figure FDA0002929081460000032
Figure FDA0002929081460000033
to calculate the total timeWeight matrix and bias parameter of the overhead, WTRepresents the transpose of the W matrix;
step (3), defining the real time overhead vector of the track passing through each grid sequence as T; the sequential real time overhead vector is TfThe real time overhead vector of the reverse order is Tb(ii) a The time overhead vector estimated by the neural network is:
Figure FDA0002929081460000034
Figure FDA0002929081460000035
the model is subjected to auxiliary supervised learning by using a bidirectional interval loss function, so that the time overhead of the whole path can be learned, and the transit time of each intermediate stage can be learned; the two-way interval loss function is defined as:
Figure FDA0002929081460000036
wherein, M represents whether the track passes through the mask of a small region, and [ ] represents the operation among each element of the vector;
step (4), the goal of training is to minimize the loss function L, i.e.:
Figure FDA0002929081460000037
wherein, theta represents the training parameter of the model, epsilon represents the embedding vector on time and space, and S is the size of the training set; finally, updating and optimizing parameters of the model by using a time sequence-based back propagation algorithm;
the third step of the prediction stage comprises the following steps:
step (1), a real journey without a time stamp mark is givenAs a query path, obtaining a grid sequence mapped by the actual path according to the actual path; for each small grid passing through, using space-time characteristics V obtained by characteristic extraction and expression stage extractionspAnd VtpDriving state characteristic VdriAnd historical short and long term traffic condition characteristics VshortAnd VlongV is represented as a total feature of the mesh; the embedded vector of the space-time characteristics uses vector information updated by parameters in a training process; carrying out feature mining on the traffic state features of short time and long time by using the trained sub-cycle neural network;
step (2), inputting the extracted characteristics of all aspects into the trained bidirectional cyclic neural network in each passing grid to obtain the current hidden variable htThen the estimated time to pass through the current region is WT·ht+ b, and the total time overhead estimate is:
Figure FDA0002929081460000038
where n represents the total number of grids passed,
Figure FDA0002929081460000039
for the trained weight matrix and bias parameters, W, of the total time overheadTRepresenting the transpose of the W matrix.
CN201810658375.5A 2018-06-25 2018-06-25 Travel time estimation method based on auxiliary supervised learning Expired - Fee Related CN109035761B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201810658375.5A CN109035761B (en) 2018-06-25 2018-06-25 Travel time estimation method based on auxiliary supervised learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201810658375.5A CN109035761B (en) 2018-06-25 2018-06-25 Travel time estimation method based on auxiliary supervised learning

Publications (2)

Publication Number Publication Date
CN109035761A CN109035761A (en) 2018-12-18
CN109035761B true CN109035761B (en) 2021-06-04

Family

ID=64611043

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201810658375.5A Expired - Fee Related CN109035761B (en) 2018-06-25 2018-06-25 Travel time estimation method based on auxiliary supervised learning

Country Status (1)

Country Link
CN (1) CN109035761B (en)

Families Citing this family (15)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109799533B (en) * 2018-12-28 2021-07-27 中国石油化工股份有限公司 Reservoir prediction method based on bidirectional circulation neural network
CN109783771B (en) * 2019-01-22 2021-01-29 清华大学 Processing method and device for converting track sequence into image matrix and storage medium
CN111721306B (en) * 2019-03-20 2022-07-05 北京嘀嘀无限科技发展有限公司 Road matching method and device, electronic equipment and readable storage medium
CN110232169B (en) * 2019-05-09 2022-01-04 北京航空航天大学 Track denoising method based on bidirectional long-time and short-time memory model and Kalman filtering
CN110488835B (en) * 2019-08-28 2020-12-08 北京航空航天大学 Unmanned system intelligent local path planning method based on double back propagation neural networks
CN110675632B (en) * 2019-11-11 2021-11-30 重庆邮电大学 Vehicle short-time trajectory prediction control method aiming at multi-feature space and data sparseness
CN110942211A (en) * 2019-12-12 2020-03-31 武汉中海庭数据技术有限公司 Prediction arrival time prediction method and device based on deep neural network
CN111047107B (en) * 2019-12-23 2022-05-10 北京百度网讯科技有限公司 Road traffic time prediction method, device, electronic equipment and storage medium
CN111339156B (en) * 2020-02-07 2023-09-26 京东城市(北京)数字科技有限公司 Method, apparatus and computer readable storage medium for long-term determination of business data
CN111738500B (en) * 2020-06-11 2024-01-12 大连海事大学 Navigation time prediction method and device based on deep learning
CN112101132B (en) * 2020-08-24 2022-04-19 西北工业大学 Traffic condition prediction method based on graph embedding model and metric learning
CN112017436B (en) * 2020-09-09 2021-09-28 中国科学院自动化研究所 Method and system for predicting urban traffic travel time
CN112859898B (en) * 2021-01-18 2022-03-22 中山大学 Aircraft trajectory prediction method based on two-channel bidirectional neural network
CN115394088B (en) * 2022-10-31 2023-04-07 江苏博宇鑫信息科技股份有限公司 Crossing traffic time prediction method based on space-time attention mechanism
CN116542438B (en) * 2023-03-28 2024-01-30 大连海事大学 Bus passenger starting and stopping point estimation and repair method based on non-reference real phase

Citations (17)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103280110A (en) * 2013-06-08 2013-09-04 北京云星宇交通工程有限公司 Method and device for predicting expressway travel time
CN104157142A (en) * 2014-08-27 2014-11-19 河海大学 Urban path travel time forecasting method based on floating vehicle data
CN104269059A (en) * 2014-10-15 2015-01-07 河海大学 City path travel time forecasting method based on multi-source data fusion
CN104299442A (en) * 2014-10-15 2015-01-21 河海大学 Urban route travel time forecasting method based on pattern matching
CN104408958A (en) * 2014-11-11 2015-03-11 河海大学 Urban dynamic route travel time predication method
CN104615897A (en) * 2015-02-13 2015-05-13 北方工业大学 Road section travel time estimation method based on low-frequency GPS data
CN105006147A (en) * 2015-06-19 2015-10-28 武汉大学 Road segment travel time deducing method based on road space-time incidence relation
CN105679021A (en) * 2016-02-02 2016-06-15 重庆云途交通科技有限公司 Travel time fusion prediction and query method based on traffic big data
CN106023588A (en) * 2016-06-15 2016-10-12 重庆云途交通科技有限公司 Traffic big data-based travel time extraction, prediction and query method
CN106096767A (en) * 2016-06-07 2016-11-09 中国科学院自动化研究所 A kind of link travel time prediction method based on LSTM
CN106940930A (en) * 2016-11-01 2017-07-11 深圳市城市交通规划设计研究中心有限公司 Motorway journeys time prediction system and Forecasting Methodology
CN106981198A (en) * 2017-05-24 2017-07-25 北京航空航天大学 Deep learning network model and its method for building up for predicting travel time
CN107085943A (en) * 2015-12-23 2017-08-22 青岛海信网络科技股份有限公司 A kind of road travel time short term prediction method and system
CN107688556A (en) * 2017-07-24 2018-02-13 中山大学 A kind of real-time travel time computation method based on function type principal component analysis
CN107945507A (en) * 2016-10-13 2018-04-20 腾讯科技（深圳）有限公司 Travel Time Estimation Method and device
KR20180064692A (en) * 2016-12-06 2018-06-15 현대자동차주식회사 A navigator, a vehicle the navigator is installed in, a system for determining a route, a method for controlling the same
WO2018109516A1 (en) * 2016-12-13 2018-06-21 日産自動車株式会社 Automatic driving assistance method and device

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106156531B (en) * 2016-08-04 2020-09-01 复旦大学 Travel time estimation method based on low-sampling historical track
CN108073941A (en) * 2016-11-17 2018-05-25 江南大学 A kind of image, semantic generation method based on deep learning
CN107480786B (en) * 2017-08-07 2021-04-30 复旦大学 Output state limitation-based recurrent neural network track likelihood probability calculation method

Patent Citations (17)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN103280110A (en) * 2013-06-08 2013-09-04 北京云星宇交通工程有限公司 Method and device for predicting expressway travel time
CN104157142A (en) * 2014-08-27 2014-11-19 河海大学 Urban path travel time forecasting method based on floating vehicle data
CN104269059A (en) * 2014-10-15 2015-01-07 河海大学 City path travel time forecasting method based on multi-source data fusion
CN104299442A (en) * 2014-10-15 2015-01-21 河海大学 Urban route travel time forecasting method based on pattern matching
CN104408958A (en) * 2014-11-11 2015-03-11 河海大学 Urban dynamic route travel time predication method
CN104615897A (en) * 2015-02-13 2015-05-13 北方工业大学 Road section travel time estimation method based on low-frequency GPS data
CN105006147A (en) * 2015-06-19 2015-10-28 武汉大学 Road segment travel time deducing method based on road space-time incidence relation
CN107085943A (en) * 2015-12-23 2017-08-22 青岛海信网络科技股份有限公司 A kind of road travel time short term prediction method and system
CN105679021A (en) * 2016-02-02 2016-06-15 重庆云途交通科技有限公司 Travel time fusion prediction and query method based on traffic big data
CN106096767A (en) * 2016-06-07 2016-11-09 中国科学院自动化研究所 A kind of link travel time prediction method based on LSTM
CN106023588A (en) * 2016-06-15 2016-10-12 重庆云途交通科技有限公司 Traffic big data-based travel time extraction, prediction and query method
CN107945507A (en) * 2016-10-13 2018-04-20 腾讯科技（深圳）有限公司 Travel Time Estimation Method and device
CN106940930A (en) * 2016-11-01 2017-07-11 深圳市城市交通规划设计研究中心有限公司 Motorway journeys time prediction system and Forecasting Methodology
KR20180064692A (en) * 2016-12-06 2018-06-15 현대자동차주식회사 A navigator, a vehicle the navigator is installed in, a system for determining a route, a method for controlling the same
WO2018109516A1 (en) * 2016-12-13 2018-06-21 日産自動車株式会社 Automatic driving assistance method and device
CN106981198A (en) * 2017-05-24 2017-07-25 北京航空航天大学 Deep learning network model and its method for building up for predicting travel time
CN107688556A (en) * 2017-07-24 2018-02-13 中山大学 A kind of real-time travel time computation method based on function type principal component analysis

Non-Patent Citations (5)

* Cited by examiner, † Cited by third party
Title
Long short-term memory neural network for traffic speed;Xiaolei Ma，Zhimin Tao, Yinhai Wang, Haiyang Yu;《Transportation Research Part C》;20150430(第54期);全文 *
Short-term travel time prediction by deep learning: A comparison of different LSTM-DNN models;Yangdong Liu;《2017 IEEE 20th International Conference on Intelligent Transportation Systems (ITSC)》;20171031;全文 *
WhenWill You Arrive? Estimating Travel Time Based on Deep Neural Networks;Dong Wang, Junbo Zhang, Wei Cao, Jian Li, Yu Zheng;《Microsoft》;20180131;全文 *
利用浮动车大数据进行稀疏路段行程时间推断;张发明，朱欣焰，呙维，胡涛;《武汉大学学报· 信息科学版》;20170131;第42卷(第1期);全文 *
基于时空贝叶斯模型的行程时间可靠性预测;杨庆芳，韦学武，林赐云，李志林，刘翔宇;《华南理工大学学报( 自然科学版)》;20160430;第44卷(第4期);全文 *

Also Published As

Publication number Publication date
CN109035761A (en) 2018-12-18

Similar Documents

Publication Publication Date Title
CN109035761B (en) Travel time estimation method based on auxiliary supervised learning
Zhang et al. Deeptravel: a neural network based travel time estimation model with auxiliary supervision
Qiu et al. Nei-TTE: Intelligent traffic time estimation based on fine-grained time derivation of road segments for smart city
CN111653088B (en) Vehicle driving quantity prediction model construction method, prediction method and system
CN109658695B (en) Multi-factor short-term traffic flow prediction method
Yuan et al. Driving with knowledge from the physical world
Chien et al. Dynamic travel time prediction with real-time and historic data
Thomas et al. Predictions of urban volumes in single time series
CN106971547A (en) A kind of Short-time Traffic Flow Forecasting Methods for considering temporal correlation
Olayode et al. Prediction and modeling of traffic flow of human-driven vehicles at a signalized road intersection using artificial neural network model: A South African road transportation system scenario
CN103413443A (en) Short-term traffic flow forecasting method based on hidden Markov model
CN112785077B (en) Travel demand prediction method and system based on space-time data
CN116052427B (en) Inter-city inter-regional mobility prediction method and device based on private car travel track data
CN112863182A (en) Cross-modal data prediction method based on transfer learning
CN111815098A (en) Traffic information processing method and device based on extreme weather, storage medium and electronic equipment
CN111427932B (en) Travel prediction method, travel prediction device, travel prediction equipment and storage medium
CN115206092A (en) Traffic prediction method of BiLSTM and LightGBM model based on attention mechanism
Doğan Short-term traffic flow prediction using artificial intelligence with periodic clustering and elected set
Zhang et al. A hybrid deep learning approach for urban expressway travel time prediction considering spatial-temporal features
Buroni et al. A tutorial on network-wide multi-horizon traffic forecasting with deep learning.
Turki et al. A Markova-Chain Approach to Model Vehicles Traffic Behavior
Othman et al. Enhancing realism in simulation through deep learning
CN117077843A (en) Space-time attention fine granularity PM2.5 concentration prediction method based on CBAM-CNN-converter
CN116258284A (en) Traffic and construction organization scheme optimization decision method and system
Zhang et al. Finding paths with least expected time in stochastic time-varying networks considering uncertainty of prediction information

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant
CF01 Termination of patent right due to non-payment of annual fee

Granted publication date: 20210604

CF01 Termination of patent right due to non-payment of annual fee