CN111292534A - A Traffic State Estimation Method Based on Clustering and Deep Sequence Learning - Google Patents

Abstract

The invention provides a traffic state estimation method based on kmeans clustering and deep sequence learning, belongs to the field of intelligent traffic systems, and mainly solves the problem of estimating the traffic state of the whole expressway under the condition that traffic flow data of partial road sections in an urban expressway cannot be acquired in real time. The method is characterized by comprising the following steps: (1) dividing a rapid road network; (2) modeling and data acquisition of the expressway; (3) preprocessing and normalizing data; (4) calculating Euclidean distances among traffic flow data through a kmeans clustering algorithm, and determining the traffic state grade of each data point; (5) and (3) designing a deep sequence learning Seq2Seq model, and performing traffic state recognition on the whole road network through model iterative learning. The invention fully considers the relation of traffic flow among road sections, exerts the advantages of a machine learning algorithm in the traffic field, obtains the traffic state of the whole road network in time and can provide reliable traffic information for a driving subject.

Description

A Traffic State Estimation Method Based on Clustering and Deep Sequence Learning

technical field

The invention relates to the field of intelligent traffic systems, in particular to a traffic state estimation method based on kmeans clustering and deep sequence learning.

Background technique

With the continuous development of my country's social economy and the continuous growth of urban population, more and more families own one or more private cars. The rapid increase in the number of vehicles has led to increasing traffic pressure in major cities in my country, which has seriously affected the The operation efficiency of the urban traffic network increases the travel time of residents. In addition, the low-speed driving of vehicles in congestion will aggravate the waste of energy. Frequent stalls and starts will also increase exhaust emissions and pollute the living environment of residents. Therefore, how to accurately estimate the traffic flow of the urban road network and relieve the traffic pressure under the condition of meeting the needs of people's travel has become an important direction of the development of traffic management and the focus of academic research.

Traffic state estimation refers to the process of inferring the traffic state of the entire road network using the traffic flow data observed in the road network. At present, it is mainly divided into two types: model-driven and data-driven: model-driven algorithms generally use traffic flow models to describe the transmission relationship between road sections, and use mathematical formulas to deduce traffic state changes; data-driven algorithms usually use machine learning to Analyze historical traffic flow data and mine the relationship between the data, and then estimate or predict the traffic status of the road section.

However, due to technical and financial reasons, the current road detectors in the urban road network cannot achieve seamless coverage and can only detect the traffic flow data of some road sections, resulting in most existing technologies focusing on a single road section. , it is impossible to effectively estimate the road sections for which no data has been detected, and it cannot meet the needs of the traveling crowd for the traffic information of the road network. Therefore, no effective solution has been proposed for the above problem.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a traffic state estimation method based on kmeans clustering and deep sequence learning in view of the above-mentioned problems, aiming to solve the problem that the traffic flow data of some road sections in urban expressways cannot be obtained in real time. The problem of estimating the traffic state of a road segment.

The technical scheme of the present invention is implemented according to the following steps:

S1. Expressway division: According to the CTM theory of the cellular transmission model, an urban expressway is divided into several balanced sections to ensure that the traffic flow density inside each section is evenly distributed, and the cross-sectional flow and traffic speed are roughly the same;

S2. Data collection: use simulation software to model the selected expressway, set up a virtual detector, and obtain the historical parameter data of traffic flow of each road section, wherein the characteristics of the data include traffic flow, speed and time occupancy rate of the road section;

S3. Data preprocessing: Eliminate the collected duplicate data and abnormal data, and normalize different traffic flow characteristic data to convert them into values in the [0,1] interval;

S4. Traffic status division: According to the basic characteristics of the traffic flow of the road section, the traffic status is divided into two status levels: free flow and congested flow, and the kmeans clustering algorithm is used to cluster and analyze the historical traffic flow data of each road section. The Euclidean distance in the three-dimensional space is used to judge the category of each data point, so as to achieve the purpose of calibrating the data set of each road segment;

S5. Traffic state estimation: construct a training data set according to a certain proportion of the calibrated data, and design a deep sequence learning model Seq2Seq model. The input of the model is the traffic flow data sequence of some sections of the expressway, and the output is the traffic flow of all sections of the expressway. The state sequence is used to estimate the traffic state of the entire road segment through iterative learning, and the estimation result is obtained.

Further, in the step S1, the section division rule of the expressway is:

S1.1. According to the number and location of ramps in the expressway network, the location where the number of lanes changes, and the location where the radius of curvature of the road changes, the road network can be divided into several segments, each segment is called a link segment; the ramps include On-ramps and off-ramps, the number of lanes changing as an increase or decrease;

S1.2. Each link road segment is further divided into several smaller road segments of equal length according to its length, each small road segment is called a cell, and each cell is guaranteed to be balanced.

Further, in the step S2, the expressway model is divided according to the road section division rules described in S1, and the traffic demand of each road section and the diversion ratio between the entrance and exit ramps and the main road are dynamically set according to the real data, so that the simulated Traffic flow can evolve according to real traffic flow changes. The historical traffic flow parameter data collected by each road section can be the data of several consecutive working days, which can be simulated by changing the random seed of the software.

Further, in the step S3, since the dimension difference between different characteristic variables of the traffic flow data is relatively large, for example, the traffic flow can reach hundreds or thousands, and the traffic flow speed is only a few dozen, which is directly used in the subsequent model. Training can lead to inaccurate results. Therefore, it is necessary to normalize the collected data so that different characteristic variables can fall within a specific area. The normalization formula is:

Among them: x' is the value after data normalization, x is the value before data normalization, x _min is the minimum value in the data set, and x _max is the maximum value in the data set.

Further, in the step S4, the traffic state level is divided according to the basic graph formed by the traffic flow data inside the cell, and it is divided into two types: free flow and congested flow. It is not affected by the outside world, and the number 0 is used to represent this state; the traffic operation in the congestion flow state is extremely unstable, and the interference between vehicles is severe, and the number 1 is used to represent this state.

Further, in the step S4, the kmeans clustering algorithm is used to cluster the historical traffic flow parameter data, the historical traffic flow parameter data of each group is calibrated (0 or 1 state), and different traffic state categories are calculated. The mean value of the data is compared and the differences are compared. The specific steps of the clustering algorithm are as follows:

S4.1. Determine the number of cluster categories k, and randomly select k data points from the data set as the initial cluster center;

S4.2. Calculate the Euclidean distance between each data point and the cluster center, and divide it into the category where the cluster center is closer. The calculation formula is expressed as:

Among them, x _i is the i-th data point in the dataset, μ _j is the center point of the j-th cluster category, and k is the number of cluster categories.

S4.3. Calculate the arithmetic mean of data points in different categories according to the clustering results, and use this value to replace the previous cluster center point. The update formula is expressed as:

Among them, c _j is the number of data points contained in the jth cluster category.

S4.4. Compare the difference between the current cluster center point and the center point before the update. If they are the same, the iteration stops and the algorithm ends; if they are different, return to step S4.2 to continue the iteration.

Further, in the step S4, the objective function of the algorithm is expressed as:

Among them, E represents the squared error of the algorithm, and C _j represents the jth cluster.

Further, in the step S5, the deep sequence Seq2Seq model is a model with a "many-to-many" structure. Before training the model, the input and output data of the model need to be fitted in advance, and the expressway network needs to be fitted. The traffic flow data and state data are arranged in spatial order, and the road network traffic flow data (continuous) sequence and traffic state (discrete) binary sequence corresponding one by one are obtained. The traffic flow data sequence is used as the input and the state sequence is used as the output. to train a Seq2Seq model. In addition, since the data of some road sections on the actual road network cannot be obtained in real time, it is necessary to clear the data of that part of the road sections in the input sequence when training the model, so that the trained model can use the data of some road sections in the future estimation process. The state sequence of the entire expressway is estimated.

Further, in the step S5, the Seq2Seq model is composed of a two-layer LSTM neural network, so as to solve the gradient disappearance and gradient explosion problems faced by the RNN neural network. The previous layer of LSTM network is used as the encoder, which is responsible for parsing the input traffic flow data sequence. The encoder formula is expressed as:

h _t =f(h _t-1 ,x _t ) (5)

where f represents the activation function of the encoder, x _t represents the traffic flow data sequence at time t, and h _t represents the hidden state of the encoder at time t.

The latter layer of LSTM network is used as the decoder, which is responsible for parsing the output of the encoder, and calculates the traffic state probability distribution sequence of the entire expressway according to certain rules. The output formula of the decoder and the model is expressed as:

s _t =f(s _t-1 ,y _t-1 ,c) (6)

p(y _t |y _t-1 ,y _t-2 ,...,y ₁ ,c)=g(s _t ,y _t-1 ,c) (7)

Among them, f represents the activation function of the decoder, c represents the output of the encoder, s _t represents the hidden state of the decoder at time t, y _t represents the output result of the decoder at time t, and g represents the activation function of the output layer.

Further, in order to enable the model to remember more traffic flow data information of road sections in the encoding stage and improve the subsequent estimation accuracy, the present invention introduces an attention mechanism on the basis of the existing model for optimization, by hiding the encoder at all times. A dynamic weight is added to the state, so that the output c of the encoder can be dynamically updated, ensuring that the model can obtain more important information from the input sequence. The update formula is expressed as:

e _ij =a(s _i-1 ,h _j ) (10)

where c _i represents the dynamically updated encoder output, a _ij represents the weight between the jth hidden state of the encoder and the ith hidden state of the decoder, and e _ij represents an alignment model used to measure the jth hidden state of the encoder Correlation between the hidden state and the ith hidden state of the decoder.

Compared with existing methods, the beneficial effects of the present invention are as follows:

Aiming at the situation that the traffic flow data of some sections of the urban expressway cannot be obtained in real time, a traffic state estimation method based on kmeans clustering and deep sequence learning is proposed. First, the state calibration of the road historical traffic flow data is carried out by kmeans clustering algorithm, and then a deep sequence learning Seq2Seq model is designed, which uses the calibrated traffic flow data to train the model, and iteratively learns to obtain the traffic state sequence of the entire expressway. In the existing traffic state estimation problem, the road detector cannot achieve seamless coverage and can only detect the traffic flow data of some road sections. The relationship between traffic flow between road sections is fully considered, and the advantages of machine learning algorithms in the field of traffic are used to obtain the traffic status of the entire road network in time, and provide reliable and accurate information for driving subjects.

Description of drawings

1 is a flowchart of a traffic state estimation method based on kmeans clustering and deep sequence learning described in the present invention;

Fig. 2 is a schematic diagram of expressway division;

Fig. 3 is a schematic diagram of simulation modeling taking Jingtong Expressway as an example;

Fig. 4 is the traffic flow data clustering diagram of kmeans clustering algorithm;

Figure 5 is a structural diagram of the deep sequence learning Seq2Seq model;

Detailed ways

In order to clearly illustrate the present invention, the present invention will be further described below with reference to the embodiments and accompanying drawings. Obviously, the content specifically described below is illustrative rather than restrictive, and should not limit the protection scope of the present invention.

As shown in FIG. 1 , the present invention discloses a traffic state estimation method based on kmeans clustering and deep sequence learning. The estimation method includes the following steps:

S1. Expressway division: In this example, the section between Guomao Bridge and Yuantong Bridge (from west to east) of Beijing Jingtong Expressway is selected as an example for analysis. This expressway is about 7km long and has 7 exit ramps and 6 entrances. The ramp, and there are lane changes and turns in the road section, it will be divided into several balanced road sections according to the CTM theory. , traffic speed, etc. are roughly the same; the specific division rules are as follows:

S1.1. Divide the road network into several sections according to the number and location of ramps (including on-ramp and off-ramp) in the expressway network, the location where the number of lanes changes (the number of lanes increases or decreases), and the location where the radius of curvature of the road changes. , each segment is called a link segment;

S1.2. Each link road segment is further divided into several smaller road segments of equal length according to its length, and each small road segment is called a cell.

Through the above rules, the Jingtong Expressway is divided into 18 cells, so that each cell is balanced.

S2. Data collection: As shown in Figure 3, simulation software is used to model the selected Jingtong Expressway, and the traffic demand and ramp diversion ratio are dynamically set according to the actual road network traffic flow changes, and the required traffic flow The data is obtained by setting up a virtual detector in each cell. The detector counts every 30s. By changing the random seed, the traffic flow data of the morning peak from 6:00 to 10:00 on weekdays (Monday to Friday) is collected. A total of 43,200 sets of data were collected, among which the characteristics of the data included traffic flow, speed and time occupancy of the road section. The data of the first 4 days was selected as the training set, and the data of the last day was used as the test set.

S3. Data preprocessing: after the data is collected, the duplicate data and abnormal data need to be cleaned up first, and secondly, due to the large difference between the dimensions of the collected traffic flow data characteristic variables, such as the traffic flow can reach the number of Hundreds of thousands, and the traffic flow speed is only a few dozen, directly used for subsequent model training will lead to inaccurate results, so it is necessary to normalize the collected data so that different feature variables can fall within [0,1 ] interval, the normalization formula is:

Among them: x' is the value after data normalization, x is the value before data normalization, x _min is the minimum value in the data set, and x _max is the maximum value in the data set.

S4. Traffic state division: According to the relationship between the traffic flow data characteristics of the road section, the kmeans clustering algorithm is used to divide the data into two state levels of free flow and congested flow. Among them, the traffic operation in the free flow state is relatively stable, and the driving vehicles are hardly affected by the outside world, which can be represented by the number 0; the traffic operation in the congested flow state is extremely unstable, and the interference between vehicles is severe, which can be represented by the number 1 this state. The specific steps of the clustering algorithm are as follows:

S4.1. Determine the number of clustering categories k=2, and randomly select 2 data points from the traffic flow data set as the initial clustering center;

S4.2. Calculate the Euclidean distance between each data point and the cluster center, and divide it into the category where the cluster center is closer. The calculation formula is expressed as:

Among them, x _i is the i-th data point in the dataset, μ _j is the center point of the j-th cluster category, and k is the number of cluster categories.

S4.3. Calculate the arithmetic mean of data points in different categories according to the clustering results, and use this value to replace the previous cluster center point. The update formula is expressed as:

Among them, c _j is the number of data points contained in the jth cluster category.

S4.4. Compare the difference between the current cluster center point and the center point before the update. If they are the same, the iteration stops and the algorithm ends; if they are different, return to step S4.2 to continue the iteration.

Further, in the step S4, the objective function of the algorithm is expressed as:

Among them, E represents the squared error of the algorithm, and C _j represents the jth cluster.

Specifically, Figure 4 shows the kmeans clustering result graph of the traffic state of one of the road sections. The triangular data points in the figure represent congested flow, and the circular data points represent free flow. It can be seen from the figure that in the free flow state, the traffic flow occupancy rate of the road section is relatively low, and it is approximately linearly related to the flow rate. The moving vehicles are hardly disturbed by external factors and can maintain high-speed driving; in the congested flow state, the road section is The occupancy rate of the traffic flow begins to rise rapidly, and the flow will gradually decrease after reaching the peak. At this time, the traffic operation is extremely unstable, the degree of data dispersion is high, and the vehicles interfere with each other, so they can only drive at low speeds on the road. Therefore, k-means clustering can well divide the traffic flow data of road sections, the boundaries between different categories are obvious, and the clustering effect is good, which is in line with the changes of the basic map of the road section and the actual traffic flow.

S5. Traffic state estimation: Design the deep sequence learning model Seq2Seq model, which is a model with a "many-to-many" structure. Before training the model, the input and output data of the model need to be fitted in advance, and the expressway network The traffic flow data and state data are arranged in spatial order, and the road network traffic flow data (continuous) sequence and traffic state (discrete) binary sequence corresponding one by one are obtained. The traffic flow data sequence is used as the input and the state sequence is used as the output. to train a Seq2Seq model. In addition, since the data of some road sections on the actual road network cannot be obtained in real time, it is necessary to clear the data of that part of the road sections in the input sequence when training the model. Based on this, the input of this experimental design model is a traffic flow data sequence consisting of cells 1, 2, 3, 4, 5, 7, 9, 11, 13, 15, and 17, and the output is all 18 cells. It consists of binary (0 or 1) sequences of traffic states, with a total of 2400 sequence pairs.

Specifically, the model structure designed in this experiment is shown in Figure 5. It consists of two layers of LSTM neural networks in order to solve the gradient disappearance and gradient explosion problems faced by the RNN neural network. The previous layer of LSTM network is used as the encoder, which is responsible for parsing the input traffic flow data sequence. The encoder formula is expressed as:

h _t =f(h _t-1 ,x _t ) (5)

where f represents the activation function of the encoder, x _t represents the traffic flow data sequence at time t, and h _t represents the hidden state of the encoder at time t.

The latter layer of LSTM network is used as the decoder, which is responsible for parsing the output of the encoder, and calculates the traffic state probability distribution sequence of the entire expressway according to certain rules. The output formula of the decoder and the model is expressed as:

s _t =f(s _t-1 ,y _t-1 ,c) (6)

p(y _t |y _t-1 ,y _t-2 ,...,y ₁ ,c)=g(s _t ,y _t-1 ,c) (7)

Among them, f represents the activation function of the decoder, c represents the output of the encoder, s _t represents the hidden state of the decoder at time t, y _t represents the output result of the decoder at time t, and g represents the activation function of the output layer.

Further, in order to enable the model to remember more traffic flow data information of road sections in the encoding stage and improve the subsequent estimation accuracy, the present invention introduces an attention mechanism on the basis of the existing model for optimization, by hiding the encoder at all times. A dynamic weight is added to the state, so that the output c of the encoder can be dynamically updated, ensuring that the model can obtain more important information from the input sequence. The update formula is expressed as:

e _ij =a(s _i-1 ,h _j ) (10)

where c _i represents the dynamically updated encoder output, a _ij represents the weight between the jth hidden state of the encoder and the ith hidden state of the decoder, and e _ij represents an alignment model used to measure the jth hidden state of the encoder Correlation between the hidden state and the ith hidden state of the decoder.

The Seq2Seq model designed above is trained using the traffic flow data sequence that has been fitted, and the model is verified and tested by the five-fold crossover method. When the tested model reaches the predetermined performance index, the actual road network can be used to collect data. The obtained real-time data is used to estimate the traffic state of the road network, and then the real-time traffic state sequence of the entire expressway is obtained.

To sum up, the present invention proposes a traffic state estimation method based on kmeans clustering and deep sequence learning. First, the state calibration of the road historical traffic flow data is carried out by kmeans clustering algorithm, and then a deep sequence learning Seq2Seq model is designed, which uses the calibrated traffic flow data to train the model, and iteratively learns to obtain the traffic state sequence of the entire expressway. In the existing traffic state estimation problem, the road detector cannot achieve seamless coverage and can only detect the traffic flow data of some road sections. The present invention fully considers the relationship of traffic flow between road sections, exerts the advantages of machine learning algorithm in the field of traffic, obtains the traffic status of the entire road network in time, and can provide reliable and accurate information for driving subjects.

Finally, it should be noted that the above-mentioned embodiments of the present invention are only examples for clearly illustrating the present invention, rather than limiting the implementation of the present invention. For those of ordinary skill in the art, on the basis of the above descriptions Changes or changes in other different forms can also be made, and all expressway information cannot be exhaustively listed here, and all obvious changes or changes derived from the technical solutions of the present invention are still within the protection scope of the present invention.

Claims (9)

1. A traffic state estimation method based on clustering and deep sequence learning is characterized by comprising the following steps:

s1, fast path division: dividing an urban expressway into a plurality of balanced road sections according to a Cellular Transmission Model (CTM) theory, and ensuring that the traffic flow density inside each divided road section is uniformly distributed, and the section flow, the traffic flow speed and the like are approximately the same;

s2, data acquisition: modeling the selected express way by adopting simulation software, setting a virtual detector, and acquiring historical parameter data of traffic flow of each road section, wherein the characteristics of the data comprise the traffic flow of the section of the road section, the speed and the time occupancy of the road section;

s3, preprocessing data: removing the collected repeated data and abnormal data, and carrying out normalization processing on different traffic flow characteristic data to convert the data into values in an interval of [0,1 ];

s4, dividing traffic states: dividing the traffic state into two state grades of free flow and crowded flow according to the basic map characteristic of the road traffic flow, respectively carrying out cluster analysis on the historical traffic flow data of each road section by adopting a kmeans clustering algorithm, and judging the category of each data point according to the Euclidean distance of the data in a three-dimensional space, thereby achieving the purpose of calibrating the data set of each road section;

s5, traffic state estimation: and constructing a training data set by the calibrated data according to a certain proportion, designing a deep sequence learning model Seq2Seq model, inputting a traffic flow data sequence of a part of road sections in the expressway by the model, outputting a traffic state sequence of all road sections of the expressway, and realizing the traffic state estimation of the whole road section in an iterative learning mode to obtain an estimation result.

2. The traffic state estimation method based on clustering and deep sequence learning of claim 1, wherein in step S1, the segment division rule of the expressway is:

s1.1, dividing a road network into a plurality of road sections according to the number and the positions of ramps in the expressway network, the change positions of the number of lanes and the change positions of the curvature radius of the road, wherein each road section is called a link road section; the ramp comprises an entrance ramp and an exit ramp, and the number of lanes is changed into increase or decrease;

s1.2, each link section is further divided into a plurality of smaller sections with the same length according to the length of each link section, each small section is called a cell, and each cell is guaranteed to be balanced.

3. The method according to claim 1, wherein in step S2, the expressway model is divided according to the road segment division rule of S1, and the traffic demand of each road segment and the split ratio between the on-off ramp and the main road are dynamically set according to the real data, and the historical traffic flow parameter data collected by each road segment can be data of a plurality of working days continuously, and is simulated by changing the random seed of the software.

4. The traffic state estimation method based on clustering and deep sequence learning of claim 1, wherein in step S3, the collected data is normalized to make different feature variables fall into a specific area, and the normalization formula is:

wherein x' is the value after data normalization, x is the value before data normalization, x_minIs the minimum value, x, in the data set_maxIs the maximum value in the data set.

5. The method according to claim 1, wherein in step S4, the traffic status is classified according to a basic map formed by the data of the traffic flow inside the cells, and the classification is divided into two types of free flow and congestion flow, wherein the traffic operation in the free flow state is stable, the running vehicles are hardly affected by the outside, and the state is represented by the number 0; traffic running in a congested flow state is extremely unstable and interference between vehicles is severe, and this state is denoted by numeral 1.

6. The traffic state estimation method based on clustering and deep sequence learning according to claim 1, wherein in step S4, a kmeans clustering algorithm is used to cluster the historical traffic flow parameter data, the historical traffic flow parameter data of each group is calibrated (0 or 1 state), and the mean value of different traffic state category data is calculated, and the difference is compared, wherein the clustering algorithm comprises the following specific steps:

s4.1, determining the number k of clustering categories, and randomly selecting k data points from the data set as an initial clustering center;

s4.2, respectively calculating Euclidean distances between each data point and the clustering centers, and dividing the Euclidean distances into categories where the clustering centers which are close to each other are located, wherein the calculation formula is expressed as follows:

s4.3, calculating the arithmetic mean value of the data points contained in different categories according to the clustering result, replacing the previous clustering center point with the value, and updating the formula to be expressed as:

wherein x is_iFor the ith data point, μ in the dataset_jIs the central point of the jth clustering category, and k is the number of the clustering categories;

s4.4, comparing the difference between the current clustering center point and the center point before updating, if the current clustering center point and the center point before updating are the same, stopping iteration, and ending the algorithm; if not, returning to the step S4.2 and continuing iteration;

further, in step S4, the objective function of the algorithm is represented as:

where E represents the squared error of the algorithm, C_jIndicating the jth cluster.

7. The traffic state estimation method based on clustering and deep sequence learning of claim 1, wherein in step S5, the deep sequence Seq2Seq model is a model with a "many-to-many" structure, input and output data of the model need to be fitted in advance before training the model, traffic flow data and state data of the expressway network are arranged in a spatial order to obtain a traffic flow data continuous sequence and a traffic state discrete binary sequence corresponding to each other on road segments, the traffic flow data sequence is used as input, and the state sequence is used as output to train the Seq2Seq model; in addition, since data of a part of road segments on an actual road network cannot be acquired in real time, the data of the part of road segments in the input sequence needs to be removed when the model is trained, so that the trained model can estimate the state sequence of the whole express way by using the data of the part of road segments in the future estimation process.

8. The traffic state estimation method based on clustering and deep sequence learning of claim 1, wherein in step S5, the Seq2Seq model is composed of two layers of LSTM neural networks, so as to solve the problem of gradient disappearance and gradient explosion faced by RNN neural networks; the former layer LSTM network is used as an encoder and is responsible for analyzing the input traffic flow data sequence, and the formula of the encoder is as follows:

h_t＝f(h_t-1,x_t) (5)

where f denotes the activation function of the encoder, x_tRepresenting a sequence of traffic flow data at time t, h_tIndicating the hidden state of the encoder at time t;

the next layer of LSTM network is used as a decoder and is responsible for analyzing the output of the encoder, the traffic state probability distribution sequence of the whole express way is calculated according to a certain rule, and the output formula of the decoder and the model is expressed as follows:

s_t＝f(s_t-1,y_t-1,c) (6)

p(y_t|y_t-1,y_t-2,...,y₁,c)＝g(s_t,y_t-1,c) (7)

where f denotes the activation function of the decoder, c denotes the output of the encoder, s_tIndicating the hidden state of the decoder at time t, y_tRepresenting the output result at the moment t of the decoder and g the activation function of the output layer.

9. The traffic state estimation method according to claim 1, wherein in step S5, an attention mechanism is introduced to optimize based on the existing model, and a dynamic weight is added to the hidden state of the encoder at all times, so that the output c of the encoder can be dynamically updated accordingly, and it is ensured that the model can obtain more important information from the input sequence, and the update formula is represented as:

e_ij＝a(s_i-1,h_j) (10)

wherein, c_iRepresenting the dynamically updated encoder output, a_ijRepresenting the weight between the jth hidden state of the encoder and the ith hidden state of the decoder, e_ijAn alignment model is shown to measure the correlation between the jth hidden state of the encoder and the ith hidden state of the decoder.

Images