CN109035761B - A Travel Time Estimation Method Based on Assisted Supervised Learning - Google Patents

Abstract

The invention belongs to the technical field of intelligent transportation, in particular to a travel time estimation method based on auxiliary supervision learning. It looks for statistical rules from massive historical trajectory data, and estimates the overall travel time through an end-to-end deep learning model; the steps include: feature extraction and representation stage, preprocessing the trajectory data, and extracting its time and time respectively. Spatial features, driving state features, short-term and long-term traffic status features; in the training and prediction stage, these extracted features are trained and predicted with a unified bidirectional recurrent neural network; the output of the recurrent neural network passes through the current small area at each step. The time cost of these small areas is the sum of the time cost of the total path. At the same time, a bidirectional interval loss function is also introduced to constrain the intermediate time overhead. The method can efficiently and accurately estimate the travel time of vehicles in the city, and has a good effect in the actual environment.

Description

A Travel Time Estimation Method Based on Assisted Supervised Learning

technical field

The invention belongs to the technical field of intelligent transportation, and in particular relates to a travel time estimation method based on auxiliary supervision learning.

Background technique

Travel time estimation is an essential and important technology in the field of urban transportation, which can help people commute and support government planning decisions. But this is not a simple small problem, but will be affected by various dynamic factors, such as traffic dynamics, intersection conditions, changes in driver driving behavior and historical periodic data evolution, etc. These factors lead to uncertainty and difficulty in estimating travel time. With the development and popularization of GPS-enabled mobile devices, a large amount of trajectory data has been continuously generated, covering every corner of the city. With these massive historical trajectory data, we can mine the inherent laws behind the data, and learn the cycle and trend of travel time changes by building an algorithm model, so as to more accurately infer the time cost required for the current query trajectory.

At present, most of the existing methods adopt the divide-and-conquer method, which mainly decomposes the path into a series of road segments or sub-paths.

(1) Method based on a single road segment:

The single-segment-based method mainly estimates the average speed of each single-segment trajectory when it passes, and then calculates the average time cost according to the length of the segment, and finally accumulates the time of each segment to obtain the total time. But this method does not consider the intersection time overhead between road segments. Additionally, this estimation relies heavily on high-quality velocity data, which is often unavailable in trajectory data.

(2) Method based on sub-path:

The sub-path-based method mainly divides the path into a series of sub-path methods, so that the time cost of the intersection is also considered. The main idea is to splicing and mining the rich public sub-path information in historical data. Although this approach can overcome many of the shortcomings of the single-segment approach, it is still based on heuristic design rather than directly targeting travel time as an algorithmic optimization objective.

To sum up, there are two reasons why the existing methods cannot achieve satisfactory accuracy. One is that they don't think of the path as a whole, but split into individual sub-blocks. During this split, a lot of useful information is lost. Moreover, they do not take full advantage of the intermediate supervision labels specific to trajectory data, that is, the timestamp information of 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, which is more efficient than traditional heuristic models. Moreover, deep learning has a powerful representation ability, which can capture more latent features than manual models, and can handle various complex dynamics in the itinerary estimation problem.

SUMMARY OF THE INVENTION

The purpose of the present invention is to propose a travel time estimation method based on historical trajectories of auxiliary supervised learning in view of the limitations of the traditional two types of travel time estimation techniques, so as to overcome the deficiencies of the prior art.

The method of the invention searches for statistical laws from massive historical trajectory data, and makes an overall estimation of the time of the entire journey through an end-to-end deep learning model. The basic steps include: feature extraction and representation stage, preprocessing trajectory data, and extracting its various features; training and prediction stage, using these extracted features to train and predict with a unified bidirectional recurrent neural network; recurrent neural network Each step of the network outputs the time cost of passing through the current small area; the sum of the time cost of these small areas is the time cost of the total path; for more efficient training, a bidirectional interval loss function is also introduced to constrain the intermediate time cost.

The travel time estimation method based on the historical trajectory of the auxiliary supervision learning proposed by the present invention is divided into the following three stages:

(1) Feature extraction and representation stage, preprocessing the historical trajectory data and extracting its various features (including temporal and spatial features, driving state features, short- and long-term traffic conditions, etc.). The specific steps are:

In step (1), within the city limits, fine-grained division is performed on the grid according to the latitude and longitude coordinates to form adjacent small rectangular areas. Sorting in chronological order, each coordinate point in the trajectory sequence composed of GPS coordinates is mapped to the corresponding small area, forming a sequence composed of grid coordinates. For the case where the adjacent track points are far apart and fall in a discontinuous small area, the intermediate passing path can be obtained by algorithms such as map matching, and the information of this part of the discontinuous area can be completed.

Step (2), for each grid, excavate its features in different aspects. First, the embedding vector technique is used to mine latent semantic information. Embedding vector technology has been widely used in natural language processing and social networks. It mainly uses low-dimensional real-number vectors to represent the semantic information of each word or thing, and measures the distance between objects through the distance relationship in the vector space. Correspondence. The present invention utilizes the embedded vector technology to represent the semantic information of each small grid area in different spaces and different time periods. This information includes the spatial location information of different functional areas of the city (such as residential areas, commercial areas or industrial areas, etc.), as well as time information such as morning peak hours and weekends. Specifically, a low-dimensional vector is used to represent the space vector V _sp of each grid, and a day is divided into multiple time buckets (for example, one bucket per hour), and each trajectory obtains a time vector V according to the time bucket it falls into. _tp . Randomly initialize V _sp and V _tp , and then train with the model during model training.

In step (3), when the driver is driving, in different driving states, the driving speed and driving behavior will change. For example, when the vehicle is in the middle part of the driving path, it will be more inclined to drive on the road or on the elevated road, and then the speed will be faster. When you just start or approach the end point, the speed tends to slow down due to driving on a small road or in a crowded area. Specifically, the four-dimensional vector V _dri is used to represent whether the current travel stage is a departure stage, an intermediate stage, or an end stage, as well as the proportions of which have been traveled in each stage. For example, V _dri =(1,0,0,0.2) means that the driver is driving at the beginning, accounting for 20% of the total trip.

In step (4), the traffic conditions in an area often have periodic and regular changes over time. For example, if a road segment is congested between 8:00 and 8:30, it may also be congested by 8:35. That is to say, the traffic situation information in a short period of time in the past is very helpful for predicting the current traffic state. The traffic condition characteristic of this short time is defined as V _short . At the same time, long-term periodic changes in traffic conditions can also help predict current traffic conditions, such as the regularity of traffic conditions on weekdays and weekends. The traffic condition characteristic for this long time is defined as V _long . Specifically,

definition:

Represents the current traffic condition of the small area g _i in the jth time interval in the past, where v _j represents the historical average speed, n _j represents the number of historical trajectory data, and len _i /v _j represents the rough estimated transit time. These traffic condition features are input into a sub-recurrent neural network according to the historical time sequence, and the traffic condition features can be extracted.

In addition, due to the uneven distribution of historical data in different spatial regions, the number of trajectories in some regions is small, which may affect the accuracy of the estimation. In order to solve this data sparse problem, the traffic condition information of adjacent small areas is also taken into account, namely

definition:

Represents a set of grids whose distance from _gi does not exceed d, and collects their short-term traffic conditions in the past and inputs them into the neural network together. Among them, x, y represent the coordinates of the grid, and g _j represents other grids except _gi .

(2) In the training phase, the features extracted from the historical trajectory data are input into a unified bidirectional recurrent neural network (bidirectional LSTM, reference: Graves A, Schmidhuber J. Framewise phonemeclassification with bidirectional LSTM and other neural network architectures [J]. Neural Networks, 2005, 18(5-6): 602-610.) for training, and the two-way interval loss function is used as the training constraint; the specific steps are:

输入数据为

那么，第t步的输入数据为x_t，第t步得到的计算结果为h_t，则有：Step (1), build a recurrent neural network. Define the hidden layer of the network as

The input data is

Then, the input data of the t-th step is x _t , and the calculation result obtained in the t-th step is h _t , then there are:

h _t =φ(x _t ·W _x +h _t-1 ·W _h +b) (3)

是输入数据的权重矩阵(weight matrix),

是隐层的权重矩阵,

是偏置参数(bias)。φ表示一个非线性激活函数，可以是sigmoid函数，ReLU函数，tanh函数等等。in,

is the weight matrix of the input data,

is the weight matrix of the hidden layer,

is the bias parameter (bias). φ represents a nonlinear activation function, which can be a sigmoid function, a ReLU function, a tanh function, and so on.

That is, the hidden state can be represented as a function:

h _t =f(h _t-1 ,x _t ) (4)

On this basis, the forget gate is defined as:

f _t =σ(W _f ·[h _t-1 ,x _t ]+b _f ) (5)

The input gate is:

i _t =σ(W _i ·[h _t-1 ,x _t ]+b _i ) (6)

The output gate is:

o _t =σ(W _o [h _t-1 ,x _t ]+b _o ) (7)

The update of the memory unit is:

The update of the hidden layer is:

h _t =O _t ·tanh(C _t ) (10)

分别表示遗忘门、输入门、输出门和记忆单元的权重矩阵，b_f、b_i、b_o、

则是对应的偏置参数。σ()为一个非线性的激活函数，例如

是一个sigmoid函数，

是一个双曲正切函数，f( )表示一个包含各层参数的抽象神经网络函数。定义循环神经网络对应的参数为W_N；从[-α,α]的均匀分布中对循环神经网络中的每个权重参数进行初始化，其中，α是为一个超参数，设定范围为0.01到1。Among them, W _f , _{Wi , W o ,}

respectively represent the weight matrix of forgetting gate, input gate, output gate and memory unit, b _f , b _i , b _o ,

is the corresponding bias parameter. σ() is a nonlinear activation function, such as

is a sigmoid function,

is a hyperbolic tangent function, and f( ) represents an abstract neural network function containing parameters of each layer. Define the parameter corresponding to the recurrent neural network as W _N ; initialize each weight parameter in the recurrent neural network from the uniform distribution of [-α,α], where α is a hyperparameter, and the setting range is 0.01 to 1.

其中

表示正向循环神经网络的隐层，

表示反向循环神经网络的隐层。Bidirectional RNNs use both a forward RNN and a reverse RNN for computation. The forward recurrent neural network sequentially inputs the grid features extracted in the previous steps according to the sequence of the sequence, while the reverse recurrent neural network inputs the grid features in reverse order of the sequence. The advantage of this is that the neural network can simultaneously observe the position distance of the current grid from the starting point and the ending point, so as to have an overall feature. Define its latent variables as the concatenation of forward and reverse networks

in

represents the hidden layer of the forward recurrent neural network,

Represents the hidden layer of an inverse recurrent neural network.

时间特征

驾驶状态特征

历史上短时间和长时间的交通状态特征

和

拼接成一个统一的特征向量：Step (2), the features extracted from the historical trajectory data, that is, the spatial features

temporal features

Driving state characteristics

Historical short- and long-term traffic state characteristics

and

Concatenated into a unified feature vector:

为：Input to the bidirectional recurrent neural network at each passing small grid to get the transit time of passing through the grid, namely W ^T · h _i +b. total travel time

for:

分别为计算总时间开销的权重矩阵和偏置参数。W^T表示W矩阵的转置。definition

are the weight matrix and bias parameters for calculating the total time cost, respectively. W ^T represents the transpose of the W matrix.

In step (3), the real time cost vector of the trajectory passing through each grid sequence is defined as T. The sequential real time cost vector is T ^f , and the reversed real time cost vector is T ^b . Then the time cost vector estimated by the neural network is:

The model is assisted with supervised learning using a bidirectional interval loss function, so that it not only learns the time cost of the entire path, but also learns the transit time of each intermediate stage. The two-way interval loss function is defined as:

Among them, M represents whether the trajectory passes through the mask of the small area, and [] represents the operation between each element of the vector.

Step (4), the goal of training is to minimize the loss function L, that is:

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 parameters of the model are updated and optimized using a time-order-based backpropagation algorithm. Backpropagation Algorithm References: Chauvin Y, Rumelhart D E. Backpropagation:theory,architectures,andapplications[M].Psychology Press,2013.

(3) In the prediction stage, the bidirectional recurrent neural network is used to infer the features extracted from the query path and estimate the travel time; the specific steps are:

In step (1), a real itinerary without timestamp is given as a query path, and its mapped grid sequence is obtained according to the actual path passed. For each passing small grid, using the spatiotemporal features _Vsp and _Vtp extracted from the feature extraction and representation stage, the driving state feature _Vdri , and the historical short and long time traffic state features _Vshort and _Vlong , Denote V as the total feature of this mesh. Among them, the embedding vector of the spatiotemporal feature uses the vector information updated by the parameters of the training process. Short- and long-term traffic state features are mined using a trained sub-recurrent neural network.

Step (2), in each passing grid, input the extracted features into the bidirectional recurrent neural network that has been trained to obtain the current hidden variable h _t , then the estimated time passing through the current area is W ^T · _ht +b. And the total time cost estimate is:

为经过训练得到的，计算总时间开销的权重矩阵和偏置参数。W^T表示W矩阵的转置。where n is the total number of grids passed,

For the weight matrix and bias parameters obtained by training, the total time overhead is calculated. W ^T represents the transpose of the W matrix.

In general, the method of the present invention has the following advantages. First, an end-to-end deep learning method based on historical data training is used to directly learn the characteristics of the entire route and estimate the overall transit time. We define a bidirectional interval loss function that can supervise the overall path time while assisting in supervising the time cost of passing intermediate road segments. This method of introducing auxiliary supervision not only enriches the sample information of the path, but also enables the backpropagation algorithm to propagate the signal more accurately when the parameters are updated. Secondly, a feature extraction structure is proposed, which can effectively estimate the travel time of the path by extracting the dynamic features of different dimensions such as the spatiotemporal embedding vector, the driving state, and the short-time and long-time traffic conditions. Finally, it has been experimentally verified in the actual environment, and has better experimental results than existing methods.

As shown in Table 1, we conduct experiments with real historical trajectory data, including the cities of Porto and Shanghai. We compare the existing methods such as the road segment average time method, the sub-path dynamic programming method, the grid fully connected network method, and the grid convolution network method. Among them, the average time of the road segment is obtained by directly accumulating the average passing time of each road segment. Subpath Dynamic Programming [Yilun Wang, Yu Zheng, and Yexiang Xue. Travel timeestimation of a path using sparse trajecto-ries. In Proceedings of the 20th International Conference on Knowledge Discovery and Data Mining (SIGKDD), pages25–34, 2014.] Use dynamic programming to find the optimal splicing method of subpaths. The grid fully connected network method and the grid convolutional network method take the N×N overall grid as input, and use the fully connected network (Multi-Layer Perceptron) and the convolutional neural network (Convolutional Neural Network) respectively for optimization and estimation. We use MAE, RMSE, MAPE three error metrics to measure the quality of the method.

表示估计值，n表示样本总数。由表1结果可知，本发明方法在各项指标都要远好于已有的对比方法。例如，在上海数据集上，本发明方法估计MAE误差只有126秒，MAPE误差是13.3％，而已有最好方法的MAE误差在168秒，MAPE误差是19.1％。where y represents the true value,

represents the estimated value, and n represents the total number of samples. It can be seen from the results in Table 1 that the method of the present invention is far better than the existing comparative method in each index. For example, on the Shanghai dataset, the method of the present invention estimates that the MAE error is only 126 seconds, and the MAPE error is 13.3%, while the MAE error of the best method is 168 seconds, and the MAPE error is 19.1%.

Table 1

Description of drawings

Figure 1 shows a real trajectory sample, including the timestamp information of each intermediate GPS trajectory point, after a total of 720s.

Figure 2 shows the path sample to be queried, which only contains the specific path information passed, and does not contain any timestamp information.

Detailed ways

The specific implementation process of the present invention is described below in conjunction with specific examples:

Historical trajectories in Figure 1 are used for training, and travel times in Figure 2 are estimated.

1. The preprocessing stage, the feature extraction and representation stage, preprocesses the trajectory data and extracts all aspects of its features. Taking Figure 1 as an example, the specific steps are:

(1) Within the city limits, fine-grained grid division is performed and divided into adjacent small areas. As shown in Figure 1, the map is divided into 5×6 grids. Each coordinate point in the trajectory sequence is mapped to the corresponding small area to form a grid sequence, ie g={g ₁ , g ₂ ,...,g ₁₀ }.

和

来表示时空语义信息。即：(2) For each grid, excavate its features in different aspects. For example, for g ₁ , use a random vector

and

to represent spatiotemporal semantic information. which is:

来表示当前行驶阶段是出发阶段，中途阶段，还是结束阶段，以及在各个阶段已经行驶的比例,即：Second, use a four-dimensional vector

To indicate whether the current driving stage is a departure stage, a midway stage, or an end stage, and the proportion of travel in each stage, namely:

和

具体来说，定义

为过去的第1到6个时间区间(5min)内，当前区域g₁的交通状况。例如

表示历史平均速度是10m/s，共有8条历史轨迹，平均通过时间估计为20m/s。将

将这些交通状况特征按照历史时间顺序输入到一个子循环神经网络中，将最后输出的隐层向量h₆作为交通状况特征。Finally, use the short and long time traffic condition information in the past to predict the current traffic condition characteristics

and

Specifically, define

is the traffic condition of the current area g ₁ in the past 1st to 6th time interval (5min). E.g

It means that the historical average speed is 10m/s, there are 8 historical trajectories, and the average transit time is estimated to be 20m/s. Will

These traffic condition features are input into a sub-recurrent neural network according to the historical time sequence, and the last output hidden layer vector _h6 is used as the traffic condition feature.

Second, the training phase, the specific steps are:

(1) Establish a bi-directional recurrent neural network (Bi-directional LSTM) model. Randomly initialize various parameters of the model, including forget gate, input gate, output gate matrix parameters and bias parameters.

(2) The features extracted from the historical trajectory data, that is, the spatial feature V _sp , the temporal feature V _sp , the driving state feature V _dri , the historical short-time and long-time traffic state features V _short and V _long , are spliced into a unified eigenvectors of . Take grid g ₁ as an example

(3) Input to the bidirectional recurrent neural network at each passing small grid to obtain the transit time of passing through the grid, namely W ^T · h _i +b. The total travel time cost is:

For example, define W=(0.1,0.3,..,0.7), 10 grid hidden layer variable values h ₁ =(0.8,0.3,...,0.2),...,h ₁₀ =(0.7,0.4,.. 0.5), the offset value b=0.7, then:

(4) Define the real time cost vector of the trajectory passing through each grid sequence as T. The sequential real time cost vector is T ^f =(70,120,...,720), and the reversed real time cost vector is T ^b =(720,640,...,50). Then the time cost vector estimated by the neural network is:

The model is assisted with supervised learning using a bidirectional interval loss function, so that it not only learns the time cost of the entire path, but also learns the transit time of each intermediate stage. The two-way interval loss function is defined as:

Among them, M represents whether the trajectory passes through the mask of the small area, and [] represents the operation between each element of the vector.

(5) Minimize the loss function L, namely:

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 parameters of the model are updated and optimized using a time-order-based backpropagation algorithm.

3. Prediction stage, the specific steps are (take Figure 2 as an example):

(1) Given a real itinerary without time stamp as the query path g={g ₁ , g ₂ , ..., g ₈ }, according to the actual path passed, the grid sequence of its mapping is obtained. For each passing small grid g ₁ -g ₈ , use the spatiotemporal features V _sp and V _tp extracted in the feature extraction and representation stages, the driving state feature V _dri , and the historical short- and long-term traffic state features V _short and V _long , denote V as the overall feature of the grid. Among them, the embedding vector of the spatiotemporal feature uses the vector information updated by the parameters of the training process. Short- and long-term traffic state features are mined using a trained sub-recurrent neural network.

(2) In each passing grid, input the extracted features into the bidirectional recurrent neural network that has been trained to obtain the current hidden variable h _t , then the estimated time passing through the current area is W ^T · h _t +b. And the total time cost estimate is;

Among them, W and b are the parameters obtained in the previous training process.

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 vector_spDividing a day into a plurality of time buckets, and obtaining a time vector V according to the time bucket in which each track falls_tp(ii) a To V_spAnd V_tpCarrying out random initialization, and then training the model along with the model during model training;

step (3) of using the four-dimensional vector V_driTo 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 V_shortDefining a long-term traffic condition characteristic as V_longIn the case of a liquid crystal display device, in particular,

definition of：

Indicating the current small area g in the past jth time interval_iIn which v is_jRepresenting historical average speed, n_jIndicates the amount of historical track data, len_i/v_jRepresenting 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:

represents the distance g_iCollecting 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, g_jRepresents except g_iOther 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

Input data as

Then, the input data of the t step is x_tAnd the calculation result obtained in the t step is h_tThen, there are:

h_t＝φ(x_t·W_x+h_t-1·W_h+b) (3)

wherein,

is a weight matrix of the input data,

is a weight matrix of the hidden layer(s),

is a bias parameter;

i.e. the hidden state is represented as a function:

h_t＝f(h_t-1，x_t) (4)

on this basis, define forgetting the door to be:

f_t＝σ(W_f·[h_t-1，x_t]+b_f) (5)

the input gates are:

i_t＝σ(W_i·[h_t-1，x_t]+b_i) (6)

the output gate is:

o_t＝σ(W_o[h_t-1，x_t]+b_o) (7)

the memory cell is updated by:

the hidden layer is updated as follows:

h_t＝O_t·tanh(C_t) (10)

wherein,

weight matrices respectively representing the forgetting gate, the input gate, the output gate and the memory cell,

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 W_NFrom [ - 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.:

step (2), extracting the feature in the historical track data, namely the spatial feature V_spTime characteristic V_tpDriving state characteristic V_driHistorically short and long time traffic status features V_shortAnd V_longAnd splicing into a unified feature vector:

V＝(V_sp，V_tp，V_dri，V_short，V_long) (11)

at each small grid passing through, the two-way recurrent neural network is input to obtain the passing time of the grid, namely W^T·h_i+ b, total travel time overhead

Comprises the following steps:

to calculate the total timeWeight matrix and bias parameter of the overhead, W^TRepresents 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 T^fThe real time overhead vector of the reverse order is T^b(ii) a The time overhead vector estimated by the neural network is:

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:

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.:

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 extraction_spAnd V_tpDriving state characteristic V_driAnd historical short and long term traffic condition characteristics V_shortAnd V_longV 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 h_tThen the estimated time to pass through the current region is W^T·h_t+ b, and the total time overhead estimate is:

where n represents the total number of grids passed,

for the trained weight matrix and bias parameters, W, of the total time overhead^TRepresenting the transpose of the W matrix.

Images