CN115376308B - A method for predicting vehicle travel time - Google Patents

Abstract

The present invention discloses a method for predicting the driving time of a car, by establishing a LSTM-GAN model to predict the traffic speed on different roads in the future time period, according to the predicted traffic speed, the road is divided into sections with different speeds, and the time periods corresponding to the speeds of different sections, and the total driving time required for the car to reach the destination is calculated according to the speeds of different sections and the corresponding time periods; the model includes a generator and a discriminator; the generator captures the spatiotemporal characteristics of traffic flow data and outputs the initially predicted traffic speed data to the discriminator; the discriminator simultaneously inputs the corresponding actual traffic flow data to learn the feature vectors of the potential traffic flow data of the two, and finally uses the feature vectors to construct a classification model, and judges the truth of the input initially predicted traffic flow data, and outputs the initially predicted traffic flow data judged to be true as the predicted traffic flow data. The present invention can improve the prediction accuracy.

Description

A method for predicting vehicle travel time

Technical Field

The invention relates to the technical field of path planning, and in particular to a method for predicting vehicle travel time.

Background technique

Traffic congestion in cities leads to a lot of time spent on the road, which reduces people's work efficiency. Energy consumption in the transportation industry is on the rise, and stop-and-go and long-term low-gear driving can easily cause increased fuel consumption. Poor roads have become the main reason for the increase in fuel consumption, and the planning of transportation routes has become increasingly important.

The travel time of a car to its destination is an important factor in route selection, and the travel time can be calculated by predicting traffic flow. Existing traffic flow prediction methods ignore the spatiotemporal interaction between traffic flows on adjacent roads and the difference in traffic congestion on different road sections. The prediction error of travel time will affect the decision of route selection. Considering the spatiotemporal interaction of traffic flows between different roads, a long short-term memory-generative adversarial network (LSTM-GAN) deep learning algorithm is proposed to predict traffic flow, which improves the prediction accuracy.

Summary of the invention

1. Technical problems to be solved:

In view of the above technical problems, the present invention provides a method for predicting the driving time of a car, which can accurately calculate the driving time of the car, help calculate the energy consumption cost, and is particularly helpful for the route planning of electric vehicles.

2. Technical solution:

A method for predicting the driving time of a car, characterized in that: by establishing an LSTM-GAN model to predict the traffic speed on the road in the future time period, according to the predicted traffic speed, the road is divided into sections with different speeds and time periods corresponding to the speeds of different sections, and the total driving time required for the car to reach the destination is calculated according to the speeds of different sections and the corresponding time periods;

The LSTM-GAN model includes a generator H and a discriminator D; the generator H captures the temporal and spatial characteristics of the input traffic flow data and outputs the initially predicted traffic speed data to the discriminator D; the discriminator D inputs the initially predicted traffic flow data and the actual traffic flow data corresponding to its prediction to learn the feature vectors of the potential traffic flow data of the two, and finally uses the feature vectors to construct a classification model, and judges the truth of the input initially predicted traffic flow data, and outputs the initially predicted traffic flow data judged to be true as the predicted traffic flow data;

The traffic flow data adopts a traffic speed matrix sequence, and the traffic speed matrices of different time periods on the same road are arranged according to a preset period; the generator H of the LSTM-GAN model is a three-layer structure; the traffic speed matrix sequence is input into the first CNN layer, and the first CNN layer inputs the spatial characteristics of the traffic speed matrix sequences on all roads learned by it into the second LSTM layer; the second LSTM layer inputs the temporal characteristics of the continuous traffic speed matrix captured by it into the third CNN layer, and the third CNN layer generates the initial prediction data of the traffic speed matrix of the next time period; the discriminator D is a three-layer structure; the initial prediction data of the traffic speed matrix of the next time period generated by the generator H and the real traffic speed matrix are both input into the fourth CNN layer; the fourth CNN layer inputs the potential spatial features learned by it into the fifth bidirectional LSTM layer; the fifth bidirectional LSTM layer inputs the potential temporal features captured by it into the sixth layer; the sixth layer optimizes the accuracy of the generator and the discriminator through the loss function, obtains the global optimal solution, and outputs the prediction result of the traffic speed.

Furthermore, the following steps are specifically included:

Step 1: Obtain historical traffic flow data and preprocess the traffic flow data into a traffic speed matrix sequence; the traffic speed matrix sequence is a traffic speed matrix sequence of a road arranged in a preset period {v _t } = (v(t ₀ ), v(t ₁ ), …, v(t _n )), where the traffic speed matrix at time t is:

In formula (A1), v _n,1 represents the speed corresponding to the (n, 1) node on the road at time t;

Step 2: Input the traffic speed matrix sequence into the generator H of the LSTM-GAN model. After multiple trainings, the generator H predicts and generates the initial predicted speed matrix sequence of the corresponding road at time t+1.

Step 3: The initial predicted speed matrix sequence at time t+1 generated by the generator H and the corresponding real speed matrix sequence at time t+1 are simultaneously input into the discriminator D, and the discriminator D discriminates the initial predicted matrix sequence. At the beginning of the prediction, let the discriminator first learn the distribution of the real data and achieve effective recognition. If the probability output by the discriminator D is 1, the initial predicted matrix sequence is judged to be the real matrix sequence. If the probability output by the discriminator D is 0, the initial predicted matrix sequence is judged to be the generated matrix sequence; the generator learns the probability distribution of the historical data of traffic flow on the basis of a large amount of data, and the generated data is close to the real data, and is recognized by the discriminator to predict the traffic speed.

Step 4: Based on the traffic speed predicted in step 3, divide the road into sections with different traffic speeds and their corresponding time periods, so as to calculate the driving time for the car to reach the destination.

Furthermore, in step 2, the generator H learns the probability distribution of real traffic flow data through a large amount of historical data, and then uses the learned probability distribution to predict future traffic flow; in the early stage of learning, the generator H cannot pass the recognition of the discriminator D when generating data, and is identified as generated data. After the generator has undergone multiple iterative training, the generated data is close to the real data and is recognized by the discriminator D; the iterative optimization process improves the performance of the generator H and the discriminator D; when the discriminator D cannot correctly identify the data generated by the generator and the real data, it means that the generator H has learned the distribution of the real data, which improves the prediction accuracy.

Furthermore, in step 3, the discriminator D uses cross entropy as a loss function to judge the similarity between the real traffic speed matrix sequence and the distribution of the traffic speed matrix sequence initially predicted in step 2; the cross entropy as a loss function is specifically:

(1) Where: Represents a real matrix sequence, where i and j represent the node numbers corresponding to the road; /> is the real data distribution; p _v (v) is the prior distribution; Δt is the time interval; /> for The probability of being taken from the real data; H(v) is the initial predicted data from the output of the generator H; D(H(v)) is the probability from H(v) to the discriminator D;/> Represents the expectation of the initial predicted data distribution;

Through the generator H, equation (1) is minimized to obtain the optimal solution; in the continuous space, equation (1) is rewritten as equation (2):

The expected output of the discriminator D is between 0 and 1. When the distribution of the real data is derived, the goal of the discriminator D is to make the output probability As close to 1 as possible; when its input data comes from the generated data H(v), the discriminator D tries to correctly judge the data source and make D(H(v)) as close to 0 as possible, while the goal of the generator H is to make D(H(v)) as close to 1 as possible through iterative training; this means that the data generated by the generator H is getting closer and closer to the real data; that is, through the zero-sum game between the generator H and the discriminator D, the loss function of the generator H is Obj ^H (θ _H ) = -Obj ^D (θ _D ,θ _H );

Thus, the objective function of the entire LSTM-GAN model is established as shown in formula (3):

Further, the road sections with different traffic speeds in step 4 are compared with the historical average speed of the road, specifically: the road sections with queued vehicles and the road sections with speeds lower than the historical average speed are defined as downstream sections, the road sections with speeds higher than the historical average speed are defined as upstream sections, and the intersections that need to pass through traffic lights are defined as waiting sections; wherein the car reaches the destination via at least one road, the required travel time for each road on the route is calculated, and the time required for each road is summed up to be the predicted total time required for the car to reach the destination; the calculation of the required travel time for each road on the route specifically includes the following steps:

S41: Divide the driving time of the car on the corresponding road into free driving time, queuing waiting time and intersection passing time; the free driving time is the driving time on the upstream section of the road; the queuing waiting time is the time required for being on the downstream section of the road and waiting for the traffic light; the intersection passing time is the average time required for the vehicle to pass through the intersection and enter the next road, and one road corresponds to one intersection passing time; its specific expression is shown in the following formula (4):

(4)Where: (t) represents the function that changes with time; is the time function required to travel on the road; /> and/> The lengths of the time periods are free travel time, queue waiting time, and intersection passing time, respectively;

S42: Free travel time As shown in formula (5):

(5)Where: d _i,j is the total distance that needs to be traveled on this road; The length of the queued vehicles in the middle and downstream section of the road; /> is the predicted average speed on the road;

The length of the queued vehicles in the downstream section As shown in the following formula (6):

(6)Where: N _i,j (t) is the number of vehicles on the road; is the average headway between queued vehicles; μ _i,j is the maximum vehicle flow on the road; λ _i,j is the green signal ratio at the intersection;

Since the road congestion density is determined by the maximum vehicle flow μ _i,j on the road and the speed limit The ratio of is obtained; according to the relationship between road congestion density and average speed; the number of vehicles Ni _,j (t) on the road is shown in formula (7):

According to equations (5), (6) and (7), and/> As shown in formula (8) and (9) respectively:

S43: Among them, waiting time in queue

When a vehicle arrives at an intersection with a traffic light, the waiting time in the queue is the waiting time for the green light. Or wait for the red light time/> As shown in (10) and (11) respectively:

(10) Where: α _i,j represents the signal period of the road intersection;

Waiting time in queue The following equation (12) approximates:

(12)In the formula, _pj is the probability that a vehicle arrives at the intersection during the green light period;

S44: The time of crossing the intersection

The average time for a vehicle to pass through an intersection is shown in formula (13):

Where: β _j (t) is the preset time for ending the turn at the intersection based on statistical data.

3. Beneficial effects:

(1) The present invention establishes a traffic flow prediction model through the long short-term memory-generative adversarial network (LSTM-GAN) deep learning algorithm. This prediction model can obtain more accurate prediction results when facing traffic speed changes in different time periods, effectively improving the accuracy of the prediction, and also proving the important role of time-varying characteristics in traffic speed prediction.

(2) The present invention divides the road into different sections according to the prediction results of the traffic flow prediction model, thereby calculating the time it takes for a car to reach its destination. This solution has been proven to be effective through simulation, which helps to accurately calculate the travel time and shows that the prediction of travel time is closely related to the degree of traffic congestion.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a schematic diagram of the structure of the LSTM-GAN model in the present invention;

FIG2 is a flow chart of traffic speed prediction using the LSTM-GAN model in the present invention;

FIG3 is a schematic diagram of dividing road segments in the present invention;

FIG. 4 is a traffic network diagram in a specific embodiment.

Detailed ways

As shown in Figures 1 to 3, a method for predicting the driving time of a car is characterized by: by establishing an LSTM-GAN model to predict the traffic speed on different roads in the future time period, according to the predicted traffic speed, the road is divided into sections with different speeds and time periods corresponding to the speeds of different sections, and the total driving time required for the car to reach the destination is calculated according to the speeds of different sections and the corresponding time periods;

The LSTM-GAN model includes a generator H and a discriminator D; the generator H captures the temporal and spatial characteristics of the input traffic flow data and outputs the initially predicted traffic speed data to the discriminator D; the discriminator D inputs the initially predicted traffic flow data and the actual traffic flow data corresponding to its prediction to learn the feature vectors of the potential traffic flow data of the two, and finally uses the feature vectors to construct a classification model, and judges the truth of the input initially predicted traffic flow data, and outputs the initially predicted traffic flow data judged to be true as the predicted traffic flow data;

The traffic flow data adopts a traffic speed matrix sequence, and arranges the traffic speed matrices of different time periods on the same road according to a preset period; the generator H of the LSTM-GAN model is a three-layer structure; the traffic speed matrix sequence is input into the first CNN layer, and the first CNN layer inputs the spatial characteristics of the traffic speed matrix sequences on all roads learned by it into the second LSTM layer; the second LSTM layer inputs the temporal characteristics of the continuous traffic speed matrix captured by it into the third CNN layer, and the third CNN layer generates the initial prediction data of the traffic speed matrix of the next time period; the discriminator D is a three-layer structure; the initial prediction data of the traffic speed matrix of the next time period generated by the generator H and the real traffic speed matrix are both input into the fourth CNN layer; the fourth CNN layer inputs the potential spatial features learned by it into the fifth bidirectional LSTM layer; the fifth bidirectional LSTM layer inputs the potential temporal features captured by it into the sixth layer; the sixth layer optimizes the accuracy of the generator and the discriminator through the loss function, obtains the global optimal solution, and outputs the prediction result of the traffic speed.

Specific embodiment: As shown in Figure 4, in this embodiment, the starting point of the car is node 43 and the end point is node 17. The car needs to pass through the dotted line parts of 7 roads as an example to predict the speed of different sections in the next day and the corresponding average speed every 15 minutes. The road related parameters are shown in Table 1:

Table 1

In order to verify the effectiveness of the proposed LSTM-GAN model in predicting the traffic speed on different roads in the future time period, the mean absolute error (MAE), mean relative error (MRE) and root mean square error (RMSE) are used as evaluation examples in this embodiment. The LSTM-GAN model of the present application is compared with the prediction methods including LSTM, ARIMA and SVR. And because LSTM is a RNN, RNN-GAN is added as a comparison method to verify whether the combination of LSTM and GAN has significant advantages. The average speed in the next 15 minutes is predicted by these methods, and the comparison of the prediction accuracy results is shown in Table 2. Obviously, the proposed LSTM-GAN obtains the lowest MAE, MRE and RMSE values, and has the highest prediction accuracy.

Table 2

Although the present invention has been disclosed as above in terms of preferred embodiments, they are not intended to limit the present invention. Anyone skilled in the art can make various changes or modifications without departing from the spirit and scope of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection defined by the claims of this application.

Claims (2)

1. A method for predicting vehicle travel time, characterized in that: by establishing an LSTM-GAN model to predict the traffic speed on the road in the future time period, according to the predicted traffic speed, the road is divided into sections with different speeds and time periods corresponding to the speeds of different sections, and the total travel time required for the vehicle to reach the destination is calculated according to the speeds of different sections and the corresponding time periods; The LSTM-GAN model includes a generator H and a discriminator D; the generator H captures the spatiotemporal characteristics of the input traffic flow data and outputs the initially predicted traffic speed data to the discriminator D; the discriminator D inputs the initially predicted traffic flow data and the actual traffic flow data corresponding to its prediction to learn the feature vectors of the potential traffic flow data of the two, and finally uses the feature vectors to build a classification model, and judges the truth of the input initially predicted traffic flow data, and outputs the initially predicted traffic flow data judged to be true as the predicted traffic flow data; The traffic flow data adopts a traffic speed matrix sequence, and arranges the traffic speed matrices of different time periods on the same road according to a preset period; the generator H of the LSTM-GAN model is a three-layer structure; the traffic speed matrix sequence is input into the first CNN layer, and the first CNN layer inputs the spatial characteristics of the traffic speed matrix sequences on all roads learned by it into the second LSTM layer; the second LSTM layer inputs the temporal characteristics of the continuous traffic speed matrix captured by it into the third CNN layer, and the third CNN layer generates the initial prediction data of the traffic speed matrix of the next time period; the discriminator D is a three-layer structure; the initial prediction data of the traffic speed matrix of the next time period generated by the generator H and the real traffic speed matrix are both input into the fourth CNN layer; the fourth CNN layer inputs the potential spatial features learned by it into the fifth bidirectional LSTM layer; the fifth bidirectional LSTM layer inputs the potential temporal features captured by it into the sixth layer; the sixth layer optimizes the accuracy of the generator and the discriminator through the loss function, obtains the global optimal solution, and outputs the prediction result of the traffic speed; The specific steps include: Step 1: Obtain historical traffic flow data and preprocess the traffic flow data into a traffic speed matrix sequence; the traffic speed matrix sequence is a traffic speed matrix sequence of a road arranged in a preset period {v _t } = (v(t ₀ ), v(t ₁ ), …, v(t _n )), where the traffic speed matrix at time t is: In formula (A1), v _n,1 represents the speed corresponding to the (n, 1) node on the road at time t; Step 2: Input the traffic speed matrix sequence into the generator H of the LSTM-GAN model. After multiple trainings, the generator H predicts and generates the initial predicted speed matrix sequence of the corresponding road at time t+1. Step 3: The initial predicted speed matrix sequence at time t+1 generated by the generator H and the corresponding real speed matrix sequence at time t+1 are simultaneously input into the discriminator D, and the discriminator D discriminates the initial predicted matrix sequence; at the beginning of the prediction, let the discriminator first learn the distribution of the real data and achieve effective recognition. If the probability output by the discriminator D is 1, the initial predicted matrix sequence is judged to be the real matrix sequence. If the probability output by the discriminator D is 0, the initial predicted matrix sequence is judged to be the generated matrix sequence; based on a large amount of data, the generator learns the probability distribution of historical traffic flow data, and the generated data is close to the real data, and is recognized by the discriminator to predict the traffic speed; Step 4: According to the traffic speed predicted in step 3, the road is divided into sections with different traffic speeds and their corresponding time periods, so as to calculate the driving time for the car to reach the destination; In step 3, the discriminator D uses cross entropy as a loss function to judge the similarity between the real traffic speed matrix sequence and the distribution of the traffic speed matrix sequence initially predicted in step 2; the cross entropy as a loss function is specifically: (1) Where: θ _D represents the output of the discriminator D; θ _H represents the output of the generator H; Represents a real matrix sequence, where i and j represent the node numbers corresponding to the road; /> is the real data distribution; p _v (v) is the prior distribution; Δt is the time interval; /> The probability of being taken from the real data; H(v) is the initial predicted data from the output of the generator H; D(H(v)) is the probability from H(v) to the discriminator D;/> Represents the expectation of the initial predicted data distribution; Through the generator H, equation (1) is minimized to obtain the optimal solution; in the continuous space, equation (1) is rewritten as equation (2): The expected output of the discriminator D is between 0 and 1. When the distribution of the real data is derived, the goal of the discriminator D is to make the output probability As close to 1 as possible; when its input data comes from the generated data H(v), the discriminator D tries to correctly judge the data source and make D(H(v)) as close to 0 as possible, while the goal of the generator H is to make D(H(v)) as close to 1 as possible through iterative training; this means that the data generated by the generator H is getting closer and closer to the real data; that is, through the zero-sum game between the generator H and the discriminator D, the loss function of the generator H is Obj ^H (θ _H )＝-Obj ^D (θ _D ,θ _H ); Thus, the objective function of the entire LSTM-GAN model is established as shown in formula (3): The road sections with different traffic speeds in step 4 are to compare the actual traffic speed of the road with the historical average speed of the road, specifically: the road sections with queued vehicles and the road sections with a speed lower than the historical average speed are defined as downstream sections, the road sections with a speed higher than the historical average speed are defined as upstream sections, and the intersections that need to pass through traffic lights are defined as waiting sections; wherein the car reaches the destination via at least one road, the required travel time for each road on the route is calculated, and the time required for each road is summed up to be the predicted total time required for the car to reach the destination; the calculation of the required travel time for each road on the route specifically includes the following steps: S41: Divide the driving time of the car on the corresponding road into free driving time, queuing waiting time and intersection passing time; the free driving time is the driving time on the upstream section of the road; the queuing waiting time is the time required for being on the downstream section of the road and waiting for the traffic light; the intersection passing time is the average time required for the vehicle to pass through the intersection and enter the next road, and one road corresponds to one intersection passing time; its expression is specifically shown in the following formula (4): (4)Where: (t) represents the function that changes with time; is the time function required to travel on the road; and/> The lengths of the time periods are free travel time, queue waiting time, and intersection passing time, respectively; S42: Free travel time As shown in formula (5): (5)Where: d _i,j is the total distance that needs to be traveled on this road; The length of the queued vehicles in the downstream section of the road; /> is the predicted average speed on the road; The length of the queued vehicles in the downstream section As shown in the following formula (6): (6)Where: N _i,j (t) is the number of vehicles on the road; is the average headway between queued vehicles; μ _i,j is the maximum vehicle flow on the road; λ _i,j is the green signal ratio at the intersection; Since the road congestion density is determined by the maximum vehicle flow μ _i,j on the road and the speed limit The ratio of is obtained; according to the relationship between road congestion density and average speed; the number of vehicles Ni _,j (t) on the road is shown in formula (7): According to equations (5), (6) and (7), and/> As shown in formula (8) and (9) respectively: S43: Among them, waiting time in queue When a vehicle arrives at an intersection with a traffic light, the waiting time in the queue is the waiting time for the green light. Or wait for the red light time/> As shown in (10) and (11) respectively: (10) Where: α _i,j represents the signal period of the road intersection; Waiting time in queue The following equation (12) approximates: (12) In the formula, is the expected value of the waiting time in the queue; _pj is the probability that a vehicle arrives at the intersection during the green light period; S44: The time of crossing the intersection The average time for a vehicle to pass through an intersection is shown in formula (13): Where: β _j (t) is the preset time for ending the turn at the intersection based on statistical data. 2. A method for predicting vehicle driving time according to claim 1, characterized in that: in step 2, the generator H learns the probability distribution of real traffic flow data through a large amount of historical data, and then uses the learned probability distribution to predict future traffic flow; in the early stage of learning, the generator H cannot pass the recognition of the discriminator D when generating data, and is identified as generated data. After the generator has undergone multiple iterative training, the generated data is close to the real data and is recognized by the discriminator D; the iterative optimization process improves the performance of the generator H and the discriminator D; when the discriminator D cannot correctly identify the data generated by the generator and the real data, it means that the generator H has learned the distribution of the real data, thereby improving the prediction accuracy.