JP2020135503A - Travel time prediction method, model learning method, travel time prediction device, model learning device, computer program, and learned model set - Google Patents

Abstract

To provide a travel time prediction method capable of predicting highly accurate travel time.SOLUTION: A travel time prediction method for predicting a travel time of a vehicle, comprising the steps of: predicting a phenomenon of a traffic flow occurring in a predicted target link which is a road link of a travel time prediction target; selecting a model corresponding to the predicted phenomenon from models for predicting the travel times prepared for each phenomenon of traffic flow; and predicting the travel time in the predicted target link based on the selected model.SELECTED DRAWING: Figure 10

Description

The present invention relates to a travel time prediction method, a model learning method, a travel time prediction device, a model learning device, a computer program, and a trained model set.

Conventionally, an information providing service has been provided in which the time required for a vehicle on a road link to pass is calculated and the calculation result is provided to the driver of the vehicle.

A method of predicting such a travel time based on a measurement result of a past travel time is known (see, for example, Patent Document 1).

Japanese Unexamined Patent Publication No. 3-73100

For example, it is possible to predict the travel time by using a linear regression model in which the measurement result of the past travel time is used as an explanatory variable and the prediction result of the future travel time is used as the objective variable.

However, when the travel time is predicted using a model such as a linear regression model, a deviation from the measured value may occur, and improvement in prediction accuracy is expected.

The present disclosure has been made in view of such circumstances, and is a travel time prediction method, a model learning method, a travel time prediction device, a model learning device, a computer program, and a travel time prediction device capable of predicting a highly accurate travel time. The purpose is to provide a trained model set.

In order to achieve the above object, the travel time prediction method according to one embodiment of the present disclosure is a travel time prediction method for predicting the travel time of a vehicle, and is a prediction target which is a road link for which the travel time is predicted. A step of predicting the phenomenon of the traffic flow that occurs in the link, a step of selecting a model corresponding to the predicted phenomenon from the models for predicting the travel time prepared for each phenomenon of the traffic flow, and a step of selecting the model corresponding to the predicted phenomenon. It includes a step of predicting the travel time at the predicted link based on the selected model.

The model learning method according to another embodiment of the present disclosure is a model learning method for learning a model for predicting the travel time of a vehicle, and is a road link for which a travel time is predicted using the model. Based on the step of acquiring the phenomenon of the traffic flow generated in the predicted target link and the travel time observed for one or more road links including the predicted target link, the parameters of the model corresponding to the acquired phenomenon are set. Includes steps to learn.

The travel time prediction device according to another embodiment of the present disclosure is a travel time prediction device for predicting the travel time of a vehicle, and is a road link for which the travel time is predicted, which is a traffic flow generated at the prediction target link. From the phenomenon prediction unit that predicts the phenomenon and the model for predicting the travel time prepared for each traffic flow phenomenon, the model selection unit that selects the model corresponding to the predicted phenomenon was selected. A travel time prediction unit that predicts the travel time in the prediction target link based on the model is provided.

The model learning device according to another embodiment of the present disclosure is a model learning device that learns a model for predicting the travel time of a vehicle, and is a road link for which a travel time is predicted using the model. Based on the phenomenon acquisition unit that acquires the phenomenon of the traffic flow that occurred in the prediction target link and the travel time observed for one or more road links including the prediction target link, the model corresponding to the acquired phenomenon. It is equipped with a model learning unit that learns parameters.

The computer program according to another embodiment of the present disclosure is a computer program for causing a computer to function as a travel time prediction device for predicting the travel time of a vehicle, and the computer is used as a road for which a travel time is predicted. Corresponds to the predicted phenomenon from the phenomenon prediction unit that predicts the phenomenon of the traffic flow that occurs in the prediction target link that is the link and the model for predicting the travel time prepared for each traffic flow phenomenon. It functions as a model selection unit that selects a model and a travel time prediction unit that predicts the travel time in the prediction target link based on the selected model.

The computer program according to another embodiment of the present disclosure is a computer program for making a computer function as a model learning device for learning a model for predicting a vehicle travel time, and the computer is used as a model. Based on the phenomenon acquisition unit that acquires the phenomenon of the traffic flow that occurred in the prediction target link, which is the target road link for which the travel time is predicted, and the travel time observed for one or more road links including the prediction target link. Therefore, it functions as a model learning unit that learns the parameters of the model corresponding to the acquired phenomenon.

A trained model set according to another embodiment of the present disclosure is a set of trained models for making a computer function to predict the travel time of a vehicle at a prediction target link, which is a road link for which a travel time is predicted. The trained model accepts as input the travel time observed for one or more road links including the prediction target link, and outputs the future travel time in the prediction target link. , The computer is made to function, the trained model is associated with each traffic flow phenomenon generated in the prediction target link, the traffic flow phenomenon generated in the prediction target link is acquired, and the prediction target link is obtained. The trained model is generated by learning the parameters of the model corresponding to the acquired phenomenon based on the travel time observed for one or more road links including.

It should be noted that the above-mentioned computer program or trained model can be distributed via a computer-readable non-volatile recording medium such as a CD-ROM (Compact Disc-Read Only Memory) or a communication network such as the Internet. Needless to say. Further, the present disclosure can be realized as a semiconductor integrated circuit that realizes a part or all of the travel time prediction device or the model learning device, or can be realized as a system including the travel time prediction device or the model learning device.

According to the present disclosure, it is possible to predict a highly accurate travel time.

FIG. 1 is a diagram schematically showing a road link. FIG. 2 is a diagram for explaining a method of predicting travel time using a linear regression model. FIG. 3 is a diagram showing a prediction result of travel time using a linear regression model. FIG. 4 is a block diagram showing a configuration of a travel time prediction device according to an embodiment. FIG. 5A is a diagram for explaining the first phenomenon. FIG. 5B is a diagram for explaining the second phenomenon. FIG. 5C is a diagram for explaining the third phenomenon. FIG. 5D is a diagram for explaining the fourth phenomenon. FIG. 5E is a diagram for explaining the fifth phenomenon. FIG. 6 is a diagram in which the five traffic flow phenomena are classified based on the traffic flow phenomena of congestion and non-congestion. FIG. 7A is a diagram for explaining a method of determining the congestion of the prediction target link. FIG. 7B is a diagram for explaining a method of determining the congestion of the prediction target link. FIG. 8A is a diagram for explaining a method of determining the traveling direction at the end of the traffic jam. FIG. 8B is a diagram for explaining a method of determining the traveling direction at the end of the traffic jam. FIG. 9A is a diagram for explaining the prediction of the travel time of the predicted target link using the linear regression model. FIG. 9B is a diagram for explaining the prediction of the travel time of the predicted target link using the linear regression model. FIG. 10 is a flowchart showing a procedure of travel time prediction processing by the travel time prediction device. FIG. 11 is a flowchart showing details of the traffic flow phenomenon prediction process (step S2 in FIG. 10). FIG. 12 is a diagram schematically showing the travel time prediction process by the travel time prediction device. FIG. 13 is a block diagram showing a configuration of a model learning device according to an embodiment. FIG. 14 is a diagram schematically showing an input / output data set classification process and a model parameter learning process. FIG. 15 is a flowchart showing a procedure of model learning processing by the model learning device. FIG. 16 is a flowchart showing details of the traffic flow phenomenon prediction process (step S2 in FIG. 10).

[Knowledge on which this disclosure is based]
First, the findings that form the basis of this disclosure will be described.

As described above, the travel time can be calculated or predicted for each road link (hereinafter, also referred to as “link”).

FIG. 1 is a diagram schematically showing a road link. For example, it is assumed that road links are connected in the order of link 1 → link 2 → link 3, and a vehicle 1 such as an automobile travels in the order of link 1 → link 2 → link 3.

Here, the road link is a road divided by information provision units such as between intersections. Further, the link travel time (hereinafter, also referred to as “travel time”) indicates the time required for the vehicle 1 to move from one end of the road link to the other. That is, when the vehicle 1 travels in the order of link 1 → link 2 → link 3, the travel time of link 1, the travel time of link 2, and the travel time of link 3 are obtained.

As a method of predicting the travel time before traveling on the road link, a method of predicting the travel time using the linear regression model described below can be considered.

FIG. 2 is a diagram for explaining a method of predicting travel time using a linear regression model. The horizontal axis shows the time and the vertical axis shows the road link. As shown in FIG. 1, the road links are arranged in the order of link 1, link 2, and link 3 from the upstream side in the traveling direction of the vehicle. In addition, each point of the black circle indicates the travel time at its own coordinate position.

The link 2 is a link for which the travel time is predicted (hereinafter, referred to as a “prediction target link”), and the travel time at the prediction target link at a time two periods ahead of the current time is predicted. For example, the time between adjacent times is 5 minutes.

Assuming that the current time is t, the input data set 8 input to the linear regression model has the travel time of links 1 to 3 observed at time t and the travel of links 1 to 3 observed at time t-1. It's time.
The input data set 8 is input to the linear regression model to predict the travel time of the predicted link at time t + 2.

ここで
ｘ_ｉ（ｔ）：道路リンクｉの時刻ｔにおけるリンク旅行時間
ｘ_ｉ（ｔ−ｊ）：道路リンクｉの時刻ｔのｊ期前におけるリンク旅行時間
ｗ_ｉ，ｊ：ｘ_ｉ（ｔ−ｊ）に対するモデルパラメータ
ｘ´_ｉ（ｔ）：道路リンクｉの時刻ｔにおけるリンク旅行時間の予測値 The linear regression model described with reference to FIG. 2 can be expressed by the following equation 1.

Here, x _i (t): Link travel time at time t of road link i x _i (t-j): Link travel time before j period of time t of road link i w _{i, j} : x _i (t-) model parameters for the j) _x'i (t): the predicted value of the link travel time at time t of the road link i

ここで
予測先時間（何期先を予測するか）：ｑ期
モデル時間次数：ｐ
リンク数：ｍ Further, if the linear regression model is generalized, it can be expressed by the following equation 2.

Here, the predicted destination time (how many periods ahead is predicted): q period model time order: p
Number of links: m

ここで、
ｘ（ｔ）：時刻ｔにおけるリンク旅行時間の観測値
ｘ´（ｔ）：時刻ｔにおけるリンク旅行時間の予測値
ｔｉ：サンプルｉのリンク旅行時間の観測時刻および予測時刻
ｎ：観測および予測されたリンク旅行時間のサンプル数 The model parameters shown in Equation 2 are defined so as to satisfy the following Equation 3 as much as possible.

here,
x (t): Observed value of link travel time at time t x'(t): Predicted value of link travel time at time t ti: Observation time and predicted time of link travel time of sample i n: Observed and predicted Sample number of link travel times

There is a method of least squares as a method of obtaining a model parameter that satisfies Equation 3 as much as possible. The least squares method is a method of obtaining a model parameter that minimizes the value Error of the function shown in the following equation 5 (hereinafter referred to as “error function”) when the error is defined by the following equation 4.
e (t) = x'(t) -x (t) ... (Equation 4)

The error function is a quadratic function related to the model parameters wi _{and j} . Therefore, by solving the simultaneous equations in which the value obtained by partially differentiating the error function with respect to the model parameters wi _{and j} is 0, the model parameters wi _{and j} that minimize the value error of the error function can be obtained.

Next, the experimental results by the linear regression model including the model parameters wi _{and j} thus obtained will be described.

FIG. 3 is a diagram showing a prediction result of travel time using a linear regression model. The horizontal axis shows the time, and the vertical axis shows the travel time.

The thick solid line shows the observed value 41 of the travel time. The observation value 41 indicates the value of the travel time observed based on the probe information of the vehicle 1 that actually traveled on the prediction target link. The probe information is information indicating the traveling position and traveling time of the vehicle 1. Further, when the observation time is t, the travel time at time t is the average value of the travel time of the vehicle 1 that has passed the prediction target link between the time t-1 and the time t. That is, the observed value 41 is an average value of the travel time observed based on the probe information of the vehicle 1 that has passed the prediction target link between the time t-1 and the time t.
The dashed line shows the predicted value 42 of the travel time predicted using the linear regression model.

As can be seen by comparing the observed value 41 and the predicted value 42, there are times when the travel time prediction is correct, but there are times when the travel time is not predicted. In particular, the prediction tends to be wrong in the time zone around 7 o'clock when the travel time gradually increases (changes from non-congestion to non-congestion) and the time zone around 11 o'clock when the travel time gradually decreases (changes from congestion to non-congestion). There is.

As described above, there are various patterns in the phenomenon of traffic flow, and the inventor found it difficult to express these multiple phenomena with one linear regression model.

On the other hand, it is also conceivable to predict the travel time using a nonlinear model such as a neural network. In order to improve the prediction accuracy, a non-linear model must be prepared for each road link. However, the number of road links is enormous. For example, as of the end of March 2018, the "Digital Road Map Database" provided by the Japan Digital Road Map Association contains only basic road data (main roads with a width of 5.5 m or more or prefectural roads nationwide). It contains 1.54 million links. Further, the parameter learning time of the nonlinear model is at least several tens of times longer than the parameter learning time of the linear regression model. Therefore, it takes an enormous amount of time to train the parameters of the nonlinear model. Therefore, it is not realistic to predict the travel time using a nonlinear model.

Therefore, the inventor investigated a linear regression model that can predict the travel time with high accuracy. However, the present disclosure is also applicable to nonlinear models and does not exclude nonlinear models. If the parameter learning of the nonlinear model becomes possible in an operationally realistic time due to the progress of computer technology in the future, it is possible to predict the travel time by the nonlinear model.

[Summary of Embodiments of the present disclosure]
First, the outlines of the embodiments of the present disclosure will be listed and described.
(1) The travel time prediction method according to an embodiment of the present disclosure is a travel time prediction method for predicting the travel time of a vehicle, and traffic generated at a prediction target link which is a road link for which the travel time is predicted. A step of predicting a flow phenomenon, a step of selecting a model corresponding to the predicted phenomenon from among the models for predicting the travel time prepared for each traffic flow phenomenon, and the selected model. Includes a step of predicting the travel time at the predicted link based on.

According to this configuration, the travel time at the predicted link can be predicted by using a model that matches the traffic flow phenomenon predicted to occur at the predicted link. In this way, the travel time can be predicted using the model most suitable for the prediction of the travel time. Therefore, it is possible to predict the travel time with high accuracy.

(2) Preferably, in the step of predicting the phenomenon, the phenomenon is predicted based on the travel time observed for a plurality of road links including the prediction target link.

According to this configuration, not only the travel time observed at the predicted link, but also the phenomenon that occurs at the predicted link using the travel time observed at at least one of the road links upstream and downstream of the predicted link. Can be predicted. Since the phenomenon can be predicted in consideration of the situation around the prediction target link in this way, the phenomenon can be predicted accurately.

(3) Further, in the step of predicting the phenomenon, the prediction is made at a second time point in the future from the first time point based on the traffic flow phenomenon occurring at the prediction target link at the first time point. You may predict the phenomenon of the traffic flow that occurs at the target link.

According to this configuration, it is possible to predict the phenomenon that will occur in the prediction target link in the future based on the phenomenon that is currently occurring in the prediction target link.

(4) Further, the phenomenon may use the congestion generated at the road link as an index.

According to this configuration, it is possible to predict a phenomenon related to traffic congestion. Congestion has a significant impact on travel time predictions. Therefore, by using a model corresponding to the above phenomenon, it is possible to predict the travel time with high accuracy.

(5) Further, in the step of predicting the phenomenon, the phenomenon may be predicted by determining which of the first phenomenon to the fifth phenomenon the prediction target link corresponds to.
First phenomenon: the prediction target link, a predetermined number of road links on the upstream side of the prediction target link in the traveling direction of the vehicle, and a predetermined number of road links on the downstream side of the prediction target link in the traveling direction of the vehicle. No congestion has occurred in any of the second phenomenon: No congestion has occurred at the prediction target link, and the end of the congestion occurring at a predetermined number of road links on the downstream side of the prediction target link is Third phenomenon approaching the prediction target link: Congestion is occurring at the prediction target link, and the end of the congestion occurring at a predetermined number of road links on the upstream side of the prediction target link is the prediction target. Fourth phenomenon of moving away from the link: Congestion is occurring at the prediction target link, and the end of the congestion occurring at the predetermined number of road links on the upstream side of the prediction target link approaches the prediction target link. Fifth phenomenon: There is no congestion at the prediction target link, and the end of the congestion occurring at the predetermined number of road links on the downstream side of the prediction target link is away from the prediction target link.

According to this configuration, it is possible to predict the travel time with high accuracy by preparing a model corresponding to each of the five phenomena.

(6) Further, in the step of predicting the phenomenon, the traveling direction of the end position of the traffic jam generated in the plurality of road links including the prediction target link is determined, and the phenomenon is predicted based on the determination result of the traveling direction. You may.

By determining the traveling direction at the end of the traffic jam, for example, it is possible to determine whether the end of the traffic jam existing downstream of the prediction target link is close to the prediction target link and will eventually be involved in the traffic jam. In addition, the end of the traffic jam existing upstream of the prediction target link has moved downstream, and it is possible to determine whether the traffic jam will be resolved eventually. As a result, the phenomenon can be predicted accurately.

(7) Further, in the step of predicting the phenomenon, the traveling direction may be determined based on the time change of the travel time on the road link located upstream or downstream of the prediction target link.

The greater the proportion of congested sections to the road links, the greater the travel time. Therefore, according to this configuration, the traveling direction of the end position of the traffic jam can be accurately determined.

(8) Further, in the step of predicting the phenomenon, the end position of the traffic jam is determined by determining the congestion situation occurring in a plurality of road links including the prediction target link, and based on the time change of the end position. The traveling direction may be determined.

By determining whether or not a traffic jam has occurred at a plurality of road links, the end position of the traffic jam can be accurately determined. Therefore, the traveling direction of the end position of the traffic jam can be accurately determined from the time change of the end position.

(9) The model learning method according to another embodiment of the present disclosure is a model learning method for learning a model for predicting the travel time of a vehicle, and is a road for which the travel time is predicted using the model. The model corresponding to the acquired phenomenon based on the step of acquiring the phenomenon of the traffic flow generated in the predicted target link which is a link and the travel time observed for one or more road links including the predicted target link. Includes steps to learn the parameters of.

According to this configuration, a model is learned for each traffic flow phenomenon that occurs at the predicted link. That is, the trained model is a model that matches the traffic flow phenomenon that is predicted to occur at the predicted link. Therefore, by predicting the phenomenon of the traffic flow and predicting the travel time by using the model corresponding to the predicted phenomenon, the travel time can be predicted by using the optimum model for predicting the travel time. Therefore, it is possible to predict the travel time with high accuracy.

(10) The travel time prediction device according to another embodiment of the present disclosure is a travel time prediction device that predicts the travel time of a vehicle, and occurs at a prediction target link that is a road link for which the travel time is predicted. A phenomenon prediction unit that predicts a traffic flow phenomenon, a model selection unit that selects a model corresponding to the predicted phenomenon from among models for predicting travel time prepared for each traffic flow phenomenon, and a model selection unit. It includes a travel time prediction unit that predicts the travel time in the prediction target link based on the selected model.

This configuration includes a processing unit corresponding to a characteristic step included in the above-mentioned travel time prediction method. Therefore, according to this configuration, it is possible to obtain the same actions and effects as the above-mentioned travel time prediction method.

(11) The model learning device according to another embodiment of the present disclosure is a model learning device that learns a model for predicting the travel time of a vehicle, and is a road for which the travel time is predicted using the model. Corresponds to the acquired phenomenon based on the phenomenon acquisition unit that acquires the phenomenon of the traffic flow generated in the prediction target link, which is a link, and the travel time observed for one or more road links including the prediction target link. A model learning unit for learning the parameters of the model is provided.

This configuration includes a processing unit corresponding to the characteristic steps included in the model learning method described above. Therefore, according to this configuration, the same actions and effects as those of the above-mentioned model learning method can be obtained.

(12) The computer program according to another embodiment of the present disclosure is a computer program for causing a computer to function as a travel time prediction device for predicting the travel time of a vehicle, and the computer predicts the travel time. The phenomenon predicted from the phenomenon prediction unit that predicts the phenomenon of the traffic flow that occurs in the prediction target link, which is the target road link, and the model for predicting the travel time prepared for each traffic flow phenomenon. It functions as a model selection unit that selects a model corresponding to the above, and a travel time prediction unit that predicts the travel time in the prediction target link based on the selected model.

According to this configuration, the computer can function as the above-mentioned travel time prediction device. Therefore, the same actions and effects as those of the above-mentioned travel time prediction device can be obtained.

(13) The computer program according to another embodiment of the present disclosure is a computer program for causing a computer to function as a model learning device for learning a model for predicting a vehicle travel time, and the computer is used as a model learning device. The travel observed for the phenomenon acquisition unit that acquires the phenomenon of the traffic flow generated in the prediction target link, which is the target road link for which the travel time is predicted using the model, and one or more road links including the prediction target link. It functions as a model learning unit that learns the parameters of the model corresponding to the acquired phenomenon based on the time.

According to this configuration, the computer can function as the model learning device described above. Therefore, the same actions and effects as those of the above-mentioned model learning device can be obtained.

(14) The trained model set according to another embodiment of the present disclosure is a trained model set for making a computer function to predict the travel time of a vehicle at a prediction target link, which is a road link for which a travel time is predicted. A model set including a set of models, the trained model accepts as input the travel time observed for one or more road links including the prediction target link, and outputs a future travel time in the prediction target link. The trained model is associated with each traffic flow phenomenon generated at the prediction target link, acquires the traffic flow phenomenon generated at the prediction target link, and obtains the prediction target. The trained model is generated by learning the parameters of the model corresponding to the acquired phenomenon based on the travel time observed for one or more road links including the link.

According to this configuration, a model is learned for each traffic flow phenomenon that occurs at the predicted link. That is, the trained model is a model that matches the traffic flow phenomenon that is predicted to occur at the predicted link. Therefore, by predicting the phenomenon and predicting the travel time using a model corresponding to the predicted phenomenon, it is possible to predict the travel time using the model most suitable for predicting the travel time. Therefore, it is possible to predict the travel time with high accuracy.

[Details of Embodiments of the present disclosure]
Hereinafter, embodiments of the present disclosure will be described with reference to the drawings. It should be noted that all of the embodiments described below show a preferred specific example of the present disclosure. The numerical values, shapes, materials, components, arrangement positions and connection forms of the components, steps, the order of steps, etc. shown in the following embodiments are examples, and are not intended to limit the present disclosure. Further, among the components in the following embodiments, the components not described in the independent claims indicating the highest level concept of the present disclosure will be described as arbitrary components constituting the more preferable form.

<Structure of travel time prediction device>
FIG. 4 is a block diagram showing a configuration of a travel time prediction device according to an embodiment.
The travel time prediction device 2 is a device that predicts the travel time at a future time from the observed value of the travel time of the vehicle 1 traveling on a plurality of links including the prediction target link. Here, it is assumed that the travel time at the prediction target link of the time two periods ahead of the current time is predicted. The time between adjacent times (time for one period) is, for example, 5 minutes.

The travel time prediction device 2 includes a communication I / F (interface) unit 21, a storage unit 22, a probe information acquisition unit 23, a travel time observation unit 24, a phenomenon prediction unit 25, a model selection unit 26, and a travel. It includes a time prediction unit 27 and a travel time provision unit 28.

The communication I / F unit 21 is a communication interface for wirelessly communicating with the vehicle 1. For example, the communication I / F unit 21 performs wireless communication with the vehicle 1 based on a standard such as 4G (4th generation mobile communication system) or 5G (5th generation mobile communication system).

The storage unit 22 stores various data necessary for predicting the travel time of the prediction target link. For example, the storage unit 22 stores the probe information 22a, the travel time data 22b, and the learned model 22c.

The probe information 22a is information indicating the traveling position and traveling time of the vehicle 1 for each vehicle 1. The travel time data 22b is information indicating an observed value of the travel time of the road link, and is the travel time of the vehicle 1 that has passed the road link between the previous time t-1 and the time t at each time t. Shows the average value of. The trained model 22c is information indicating the parameters of the linear regression model used for predicting the travel time. A linear regression model is provided for each type of traffic flow phenomenon that occurs at the predicted link.

The probe information acquisition unit 23 acquires probe information by receiving probe information from the vehicle 1 via the communication I / F unit 21. The probe information acquisition unit 23 stores the acquired probe information in the storage unit 22 as the probe information 22a.

The travel time observation unit 24 reads the probe information 22a from the storage unit 22. Based on the read probe information 22a, the travel time observation unit 24 determines the travel time of the vehicle 1 that has passed the road link between the previous time t-1 and the time t for each time t up to the current time. Is observed for each vehicle. That is, the travel time observation unit 24 extracts the vehicle 1 whose passing time at the most downstream position of the road link is between the time t-1 and the time t. The travel time observation unit 24 observes the travel time by calculating the time required for the extracted vehicle 1 to travel from the most upstream position to the most downstream position of the road link based on the probe information 22a. The travel time observation unit 24 calculates the average value of the travel time observed at each time t, and stores it in the storage unit 22 as the travel time data 22b.

The phenomenon prediction unit 25 reads the travel time data 22b from the storage unit 22, and predicts the phenomenon of the traffic flow generated in the prediction target link based on the travel time data 22b. Specifically, the phenomenon prediction unit 25 determines which of the first to fifth phenomena the traffic flow corresponds to, which will be described below. As a result, the phenomenon prediction unit 25 predicts the phenomenon of the traffic flow that will occur in the future.

≪About the first phenomenon≫
FIG. 5A is a diagram for explaining the first phenomenon.
The first phenomenon shows a phenomenon in which congestion does not occur including the vicinity of the prediction target link 4. That is, the first phenomenon is that the prediction target link 4, the predetermined number of road links (hereinafter, referred to as “upstream links”) 5 on the upstream side of the prediction target link 4, and the predetermined number of road links on the downstream side of the prediction target link 4. It shows a phenomenon that no congestion occurs in any of the road links (hereinafter, referred to as “downstream links”) 6. In other words, the first phenomenon indicates a phenomenon in which the current state in which the prediction target link is not congested is expected to continue in the future. Therefore, when the phenomenon prediction unit 25 determines that the current traffic flow phenomenon is the first phenomenon, the observed value (current value) of the current travel time is used as the predicted value of the travel time two periods ahead. The travel time is predicted using the first model. The number of upstream links and the number of downstream links may be the same or different.

≪About the second phenomenon≫
FIG. 5B is a diagram for explaining the second phenomenon.
The second phenomenon shows a phenomenon in which no congestion has occurred at the prediction target link 4 and the end position of the congestion section 7 downstream of the prediction target link 4 (hereinafter, also referred to as “congestion end”) is approaching. That is, the second phenomenon indicates a phenomenon in which no congestion has occurred in the prediction target link 4, and the end of the congestion located in a predetermined number of downstream links 6 is approaching the prediction target link 4 with the passage of time. Here, the traffic jam section 7 indicates a section of a road link in which the vehicle speed converted from the travel time by the link length is less than a predetermined speed threshold value. In other words, the second phenomenon indicates a phenomenon in which congestion is not currently occurring in the predicted link, but is expected to occur in the future. When the phenomenon prediction unit 25 determines that the current traffic flow phenomenon is the second phenomenon, the travel time is predicted using the second model, which is a linear regression model.

≪About the third phenomenon≫
FIG. 5C is a diagram for explaining the third phenomenon.
The third phenomenon shows a phenomenon in which congestion occurs at the prediction target link 4 and the end of the congestion located on the upstream side of the prediction target link 4 moves away toward the upstream side of the prediction target link 4. That is, the third phenomenon indicates a phenomenon in which congestion occurs in the prediction target link 4, and the end of the congestion located in a predetermined number of upstream links 5 moves away from the prediction target link 4 with the passage of time. In other words, the third phenomenon indicates a phenomenon in which the current state in which congestion occurs in the predicted link is expected to continue in the future. When the phenomenon prediction unit 25 determines that the current traffic flow phenomenon is the third phenomenon, the travel time is predicted using the third model, which is a linear regression model.

≪About the 4th phenomenon≫
FIG. 5D is a diagram for explaining the fourth phenomenon.
The fourth phenomenon shows a phenomenon in which a traffic jam occurs at the prediction target link 4 and the end of the traffic jam located on the upstream side of the prediction target link 4 is approaching toward the prediction target link 4. That is, the fourth phenomenon indicates a phenomenon in which congestion occurs in the prediction target link 4, and the end of the congestion located in a predetermined number of upstream links 5 approaches the prediction target link 4 with the passage of time. In other words, the fourth phenomenon indicates a phenomenon in which a traffic jam is currently occurring at the predicted target link, but the traffic jam is expected to be eliminated in the future. When the phenomenon prediction unit 25 determines that the current traffic flow phenomenon is the fourth phenomenon, the travel time is predicted using the fourth model, which is a linear regression model.

≪About the 5th phenomenon≫
FIG. 5E is a diagram for explaining the fifth phenomenon.
The fifth phenomenon shows a phenomenon in which the traffic jam does not occur at the prediction target link 4 and the traffic jam occurs downstream of the prediction target link 4, but the end of the traffic jam moves away from the prediction target link 4. That is, the fifth phenomenon shows a phenomenon in which no congestion occurs in the prediction target link 4, and the end of the congestion located in a predetermined number of downstream links 6 moves away from the prediction target link 4 with the passage of time. In other words, the fifth phenomenon indicates a phenomenon in which the current state in which the prediction target link is not congested is expected to continue in the future. Therefore, when the phenomenon prediction unit 25 determines that the current traffic flow phenomenon is the fifth phenomenon, the observed value (current value) of the current travel time is used as the predicted value of the travel time two periods ahead. The travel time is predicted using the fifth model.

FIG. 6 is a diagram in which the five traffic flow phenomena are classified based on the traffic flow phenomena of congestion and non-congestion.

Each phenomenon can be represented by a combination of either the current congestion or non-congestion phenomenon and the expected future congestion or non-congestion phenomenon. That is, in the first phenomenon and the fifth phenomenon, both the present and future phenomena are non-congestion. The second phenomenon is that the current phenomenon is non-congestion and the future phenomenon is congestion. The third phenomenon is that the current and future phenomena are congestion. In the fourth phenomenon, the current phenomenon is congestion, and the future phenomenon is non-congestion.

≪How to judge traffic jam≫
Next, when the phenomenon prediction unit 25 predicts each phenomenon, a method of determining whether or not the prediction target link 4 is congested will be described.

7A and 7B are diagrams for explaining a method of determining the congestion of the prediction target link 4. FIG. 7A shows a situation in which the prediction target link 4 is included in the congestion section 7 and the prediction target link 4 is congested, and FIG. 7B shows a situation in which the end of the congestion is located on the downstream side of the prediction target link 4. , Indicates a situation in which the predicted link 4 is not congested.

The phenomenon prediction unit 25 calculates the average vehicle speed of the vehicle 1 on the prediction target link 4 by dividing the distance between the start and end of the prediction target link 4 by the observed value of the travel time observed on the prediction target link 4. To do. If the average vehicle speed is less than the predetermined speed threshold value, the phenomenon prediction unit 25 determines that the prediction target link 4 is congested, and if the average vehicle speed is equal to or higher than the predetermined speed threshold value, the prediction target link 4 is congested. Judge that it is not. The phenomenon prediction unit 25 may determine whether the vehicle is congested or non-congested in consideration of the duration of the congestion. That is, the phenomenon prediction unit 25 may determine that the vehicle is congested when the average vehicle speed is less than the predetermined speed threshold value continuously for a predetermined time or longer, and may determine that the vehicle is not congested in other cases. As a result, it is possible to eliminate the fluctuation of the state and accurately determine whether the vehicle is congested or non-congested.

≪How to judge the direction of travel at the end of traffic jam≫
Next, a method of determining the traveling direction at the end of the traffic jam will be described when the phenomenon prediction unit 25 predicts each phenomenon.

8A and 8B are diagrams for explaining a method of determining the traveling direction at the end of the traffic jam. FIG. 8A shows the congestion section 7 when the end of the congestion is located on the upstream side of the prediction target link 4, and FIG. 8B shows the congestion section 7 when the end of the congestion is located on the downstream side of the prediction target link 4.

As shown in FIG. 8A, when congestion occurs in the prediction target link 4, the phenomenon prediction unit 25 includes one or more predetermined number of upstream links 5 as the upstream link set 5A in the upstream link set 5A. The total travel time at the upstream link 5 is calculated. When the calculated sum is larger than the previous time, the phenomenon prediction unit 25 determines that the end of the traffic jam is moving upstream from the prediction target link 4. This is because when the end of the traffic jam is far away, the ratio of the traffic jam section 7 in the upstream link set 5A increases, so that the total sum is considered to increase. Further, the phenomenon prediction unit 25 determines that the end of the traffic jam is approaching the prediction target link 4 from the upstream side when the total sum is less than the previous time. This is because when the end of the traffic jam is approaching, the ratio of the traffic jam section 7 in the upstream link set 5A decreases, so that the total sum is considered to decrease.

The phenomenon prediction unit 25 may determine the traveling direction at the end of the traffic jam in consideration of the increase in the total sum or the duration of the phenomenon. That is, the phenomenon prediction unit 25 determines that the end of the traffic jam is moving away when the total sum is continuously increasing for a predetermined time or more, and the end of the traffic jam is when the sum is continuously decreasing for a predetermined time or more. May be judged to be approaching. As a result, it is possible to eliminate the fluctuation of the state and accurately determine the traveling direction at the end of the traffic jam.

As shown in FIG. 8B, when there is no congestion in the prediction target link 4, the phenomenon prediction unit 25 includes one or more predetermined number of downstream links 6 as the downstream link set 6A in the downstream link set 6A. The total travel time at the downstream link 6 is calculated. When the calculated total sum is larger than the previous time, the phenomenon prediction unit 25 determines that the end of the traffic jam is approaching the prediction target link 4 from the downstream side. This is because when the end of the traffic jam is approaching, the ratio of the traffic jam section 7 in the downstream link set 6A increases, so that the total sum is considered to increase. Further, the phenomenon prediction unit 25 determines that the end of the traffic jam is moving downstream from the prediction target link 4 when the total sum is less than the previous time. This is because when the end of the traffic jam is far away, the ratio of the traffic jam section 7 in the downstream link set 6A decreases, so that the total sum is considered to decrease.

The phenomenon prediction unit 25 may determine the traveling direction at the end of the traffic jam in consideration of the increase in the total sum or the duration of the phenomenon. That is, the phenomenon prediction unit 25 determines that the end of the traffic jam is approaching when the summation is continuously increasing for a predetermined time or more, and the end of the traffic jam is when the summation is continuously decreasing for a predetermined time or more. May be judged to be moving away. As a result, it is possible to eliminate the fluctuation of the state and accurately determine the traveling direction at the end of the traffic jam.

Further, the phenomenon prediction unit 25 may obtain the end of the traffic jam from the traffic jam section and determine the traveling direction of the end of the traffic jam. For example, in FIG. 8A, the congestion state is determined for each of the prediction target link 4 and each upstream link 5 and each downstream link 6 other than the prediction target link 4. As a result, the phenomenon prediction unit 25 can identify the congestion section 7 which is a set of road links in which congestion is occurring. Further, the phenomenon prediction unit 25 can determine that the end of the traffic jam is near the most downstream road link of the traffic jam section 7. As a result, the phenomenon prediction unit 25 can determine the traveling direction at the end of the traffic jam from the time change at the end of the traffic jam.

With reference to FIG. 4 again, the model selection unit 26 selects a model used for predicting the travel time from the learned models 22c stored in the storage unit 22 based on the phenomenon predicted by the phenomenon prediction unit 25. Select and read. That is, when the phenomenon prediction unit 25 determines that the traffic flow phenomenon at the current time t is the first phenomenon, the model selection unit 26 selects and reads out the first model. When the phenomenon prediction unit 25 determines that the traffic flow phenomenon at the current time t is the second phenomenon, the model selection unit 26 selects and reads out the second model. When the phenomenon prediction unit 25 determines that the traffic flow phenomenon at the current time t is the third phenomenon, the model selection unit 26 selects and reads out the third model. When the phenomenon prediction unit 25 determines that the traffic flow phenomenon at the current time t is the fourth phenomenon, the model selection unit 26 selects and reads out the fourth model. When the phenomenon prediction unit 25 determines that the traffic flow phenomenon at the current time t is the fifth phenomenon, the model selection unit 26 selects and reads out the fifth model.

The travel time prediction unit 27 reads the travel time data 22b from the storage unit 22, and inputs the read travel time data 22b into the model read from the storage unit 22 by the model selection unit 26 to predict the prediction target for two periods ahead. Predict the travel time of the link.

9A and 9B are diagrams for explaining the prediction of the travel time of the predicted target link using the linear regression model. FIG. 9A schematically shows the prediction of the travel time of the predicted target link at time t + 2 two periods ahead at time t, and FIG. 9B shows the prediction of the travel time of the predicted target link at time t + 3 two periods ahead at time t + 1. Is schematically shown. The horizontal axis of each figure indicates the time, and the vertical axis indicates the road link. As shown in FIG. 1, it is assumed that the road links are arranged in the order of link 1, link 2, link 3, link 4, link 5, link 6, link 7, link 8, and link 9 from the upstream side. In addition, each point shall indicate the travel time at its own coordinate position. Further, the link 5 is set as the prediction target link.

As shown in FIG. 9A, when the current time is t, the phenomenon prediction unit 25 predicts the phenomenon of the traffic flow based on the phenomenon prediction data set 10. The phenomenon prediction data set 10 includes the travel times of links 1 to 4 and links 6 to 9 observed at time t-3 to time t, and the travel time of link 5 observed at time t.

That is, the phenomenon prediction unit 25 predicts whether or not the traffic flow phenomenon corresponds to the first phenomenon based on the travel time of each of the links 1 to 9 observed at time t. For example, when the phenomenon prediction unit 25 determines whether or not each link is congested from the travel time of links 1 to 9 at time t, and determines that no congestion has occurred in all of links 1 to 9. Is predicted to correspond to the first phenomenon.

Further, the phenomenon prediction unit 25 determines the traffic flow based on the travel time of the link 5 observed at time t and the travel time of links 6 to 9 observed at each time point from time t-3 to time t. It is determined whether or not the phenomenon corresponds to the second phenomenon. For example, the phenomenon prediction unit 25 calculates the total travel time of links 6 to 9 at each time point from time t-3 to time t. The phenomenon prediction unit 25 determines from the travel time of link 5 at time t that no congestion has occurred at link 5, and the total travel time of links 6 to 9 calculated is at time t-3 to time t. When it is determined that the number of traffic jams is gradually increasing and the end of the traffic jam is approaching the prediction target link, it is determined that the second phenomenon is applicable.

Further, the phenomenon prediction unit 25 determines the traffic flow based on the travel time of links 1 to 4 observed at each time point from time t-3 to time t and the travel time of link 5 observed at time t. It is determined whether or not the phenomenon corresponds to the third phenomenon or the fourth phenomenon. For example, the phenomenon prediction unit 25 calculates the total travel time of links 1 to 4 at each time point from time t-3 to time t. The phenomenon prediction unit 25 determines from the travel time of link 5 at time t that congestion has occurred at link 5, and the total of the calculated travel times of links 1 to 4 is at time t-3 to time t. When it is determined that the number of traffic jams is gradually increasing and the end of the traffic jam is far from the prediction target link, it is determined that the third phenomenon is applicable. Further, the phenomenon prediction unit 25 determines that the traffic jam is occurring at the link 5 from the travel time of the link 5 at the time t, and the total of the calculated travel times of the links 1 to 4 is the time t-3 to the time. When it is determined that the traffic jam is gradually decreasing at t and the end of the traffic jam is approaching the prediction target link, it is determined that the fourth phenomenon is applicable.

Further, the phenomenon prediction unit 25 determines the traffic flow based on the travel time of the link 5 observed at time t and the travel time of links 6 to 9 observed at each time point from time t-3 to time t. It is determined whether or not the phenomenon corresponds to the fifth phenomenon. For example, the phenomenon prediction unit 25 calculates the total travel time of links 6 to 9 at each time point from time t-3 to time t. The phenomenon prediction unit 25 determines from the travel time of link 5 at time t that no congestion has occurred at link 5, and the total travel time of links 6 to 9 calculated is at time t-3 to time t. When it is judged that the traffic jam is gradually decreasing and the end of the traffic jam is far from the prediction target link, it is judged that the fifth phenomenon is applicable.

Further, the input data set 8 input to the linear regression model includes the travel time of links 4 to 6 observed at time t, the travel time of links 4 to 6 observed at time t-1, and time t−. The travel time of links 4 to 6 observed in 2 and the travel time of links 4 to 6 observed at time t-3. That is, the travel times of links 4 to 6 from the current time to three periods before the current time are input to the linear regression model.

The travel time prediction unit 27 inputs the input data set 8 into the linear regression model and predicts the travel time of the prediction target link (link 5) at time t + 2.

The linear regression model can be expressed by the following equation 6. The meaning of each constant and each variable is the same as that shown in Equation 1.

As shown in FIG. 9B, when the current time reaches t + 1, the phenomenon prediction unit 25 predicts the phenomenon of the traffic flow based on the phenomenon prediction data set 10. The phenomenon prediction data set 10 includes the travel times of links 1 to 4 and 6 to 9 observed at each time point from time t-2 to time t + 1, and the travel time of link 5 observed at time t + 1. The prediction method is the same as that described with reference to FIG. 9A except that the time is different. Therefore, the detailed description will not be repeated.

Further, the input data set 8 input to the linear regression model has the travel time of links 4 to 6 observed at time t + 1, the travel time of links 4 to 6 observed at time t, and the travel time of links 4 to 6 observed at time t-1. The observed travel times of links 4 to 6 and the travel times of links 4 to 6 observed at time t-2. That is, the travel times of links 4 to 6 from the current time (time t + 1) to three periods before the current time (time t-2) are input to the linear regression model.

The travel time prediction unit 27 inputs the input data set 8 into the linear regression model and predicts the travel time of the prediction target link (link 5) at time t + 3.

The linear regression model described with reference to FIGS. 9A and 9B is a second model to a fourth model corresponding to the second phenomenon to the fourth phenomenon, respectively.

On the other hand, in the first model and the fifth model corresponding to the first phenomenon and the fifth phenomenon, the observed value of the current travel time is used as the predicted value of the travel time two periods ahead. Therefore, when the model selected by the model selection unit 26 is the first model or the fifth model, the travel time prediction unit 27 predicts that the observed value of the current travel time is the travel time two periods ahead. To do. For example, with reference to FIG. 9A, the travel time prediction unit 27 predicts that the travel time of link 5 at time t is the travel time of link 5 at time t + 2. Further, referring to FIG. 9B, the travel time prediction unit 27 predicts that the travel time of the link 5 at the time t + 1 is the travel time of the link 5 at the time t + 3.

In the above description, the phenomenon prediction unit 25 calculates the total travel time of links 1 to 4 or the total travel time of links 6 to 9 at four time points, but these totals are used to predict the phenomenon. You don't have to calculate each time. For example, the phenomenon prediction unit 25 may store the calculation result of the sum in the storage unit 22, and may read the sum from the storage unit 22 when predicting a phenomenon at another time.

With reference to FIG. 4 again, the travel time providing unit 28 transmits the travel time information predicted by the travel time prediction unit 27 to the vehicle 1 or the like via the communication I / F unit 21 to determine the travel time. provide.

<Travel time prediction processing>
FIG. 10 is a flowchart showing a procedure of travel time prediction processing by the travel time prediction device 2. It is assumed that the travel time data 22b used for the prediction process is already stored in the storage unit 22. That is, it is assumed that the probe information acquisition unit 23 sequentially acquires the probe information 22a, and the travel time observation unit 24 generates the travel time data 22b based on the probe information 22a and stores it in the storage unit 22.

With reference to FIG. 10, the phenomenon prediction unit 25 reads out the phenomenon prediction data set from the travel time data 22b stored in the storage unit 22 (S1). For example, the phenomenon prediction unit 25 reads out the phenomenon prediction data set 10 shown in FIG. 9A.

The phenomenon prediction unit 25 predicts a traffic flow phenomenon based on the read phenomenon prediction data set (S2). That is, the phenomenon prediction unit 25 determines which of the first to fifth phenomena the traffic flow corresponds to, and outputs the determination result to the model selection unit 26 as the prediction result of the traffic flow phenomenon.

FIG. 11 is a flowchart showing details of the traffic flow phenomenon prediction process (step S2 in FIG. 10).

With reference to FIG. 11, the phenomenon prediction unit 25 determines whether or not a traffic jam has occurred at the prediction target link (S21).

When congestion occurs in the prediction target link (YES in S21), the phenomenon prediction unit 25 determines whether or not the end of the congestion is far from the prediction target link (S22).

When the end of the traffic jam is far from the prediction target link (YES in S22), the phenomenon prediction unit 25 determines that the traffic flow phenomenon corresponds to the third phenomenon, and determines that the traffic congestion state is maintained at the prediction target link. Predict (S23).

If the end of the traffic jam is not far from the link to be predicted (NO in S22), the phenomenon prediction unit 25 considers that the end of the traffic jam is approaching the link to be predicted and determines that the traffic flow phenomenon corresponds to the fourth phenomenon. However, it is predicted that the congestion state of the prediction target link will be resolved eventually (S24).

When there is no congestion in the prediction target link (NO in S21), the phenomenon prediction unit 25 determines whether or not there is congestion in the vicinity of the upstream and downstream of the prediction target link (S25). If there is no congestion in the surrounding area (YES in S25), the phenomenon prediction unit 25 determines that the traffic flow phenomenon corresponds to the first phenomenon, and predicts that the non-congestion state will be maintained at the prediction target link (YES in S25). S26).

When there is a traffic jam around the prediction target link (NO in S25), the phenomenon prediction unit 25 determines whether or not the end of the traffic jam is far from the prediction target link (S27).

When the end of the traffic jam is far from the prediction target link (YES in S27), the phenomenon prediction unit 25 determines that the traffic flow phenomenon corresponds to the fifth phenomenon, and the non-congestion state is maintained at the prediction target link. (S28).

If the end of the traffic jam is not far from the link to be predicted (NO in S27), the phenomenon prediction unit 25 considers that the end of the traffic jam is approaching the link to be predicted and determines that the traffic flow phenomenon corresponds to the second phenomenon. However, it is predicted that the prediction target link will eventually be congested (S29).

With reference to FIG. 10 again, the model selection unit 26 receives the prediction result of the traffic flow phenomenon from the phenomenon prediction unit 25, and based on the prediction result, is in the trained model 22c stored in the storage unit 22. To select a trained model to be used for predicting travel time (S3).

The model selection unit 26 reads the selected learned model from the storage unit 22, and outputs the read learned model to the travel time prediction unit 27 (S4).

The travel time prediction unit 27 reads an input data set from the travel time data 22b stored in the storage unit 22 (S5). For example, the input data set 8 shown in FIG. 9A is read out.

The travel time prediction unit 27 receives the trained model from the model selection unit 26, inputs the read input data set into the trained model, and predicts the travel time of the predicted target link two periods ahead (S6). .. The travel time prediction unit 27 outputs the travel time prediction result to the travel time providing unit 28.

The travel time providing unit 28 receives the travel time prediction result from the travel time prediction unit 27 and transmits the travel time prediction result to the vehicle 1 or the like via the communication I / F unit 21 to provide the travel time (S7).

<Typical diagram of travel time prediction processing>
FIG. 12 is a diagram schematically showing the travel time prediction process by the travel time prediction device 2. For space reasons, FIG. 12 shows the travel times of links 1 to 3 and the travel times of links 7 to 9 together. The travel time prediction process shown in FIG. 10 can be represented as shown in FIG. 12 from another viewpoint.

That is, the phenomenon of the traffic flow is predicted based on the phenomenon prediction data set 10 (ST1).

The input data set 8 is distributed to either model based on the prediction result of the traffic flow phenomenon (ST2). That is, when the prediction result is the first phenomenon, the input data set 8 takes out the travel time at the time t of the prediction target link, and sets the taken out travel time as the predicted value of the travel time at time t + 2. It is sorted. When the prediction result is the second phenomenon, the input data set 8 is distributed to the second model, which is a linear regression model. When the prediction result is the third phenomenon, the input data set 8 is distributed to the third model, which is a linear regression model. When the prediction result is the fourth phenomenon, the input data set 8 is distributed to the fourth model, which is a linear regression model. Further, when the prediction result is the fifth phenomenon, the input data set 8 takes out the travel time at the time t of the prediction target link, and sets the taken out travel time as the predicted value of the travel time at the time t + 2. It is sorted.

Each of the first model to the fifth model predicts the travel time at time t + 2 from the input input data set 8 (ST3a to ST3e).

One of the travel times of the first model to the fifth model is selected based on the traffic flow phenomenon prediction result, and is used as the prediction result of the travel time at the time t + 2 which is the output data 9 (ST4).

<Configuration of model learning device>
FIG. 13 is a block diagram showing a configuration of a model learning device according to an embodiment.
The model learning device 3 generates a trained model by learning the parameters of the linear regression model used by the travel time prediction device 2 to predict the travel time. The model learning device 3 includes a communication I / F unit 31, a storage unit 32, a probe information acquisition unit 33, a travel time observation unit 34, a phenomenon prediction unit 35, a data classification unit 36, and a model learning unit 37. To be equipped.

The communication I / F unit 31 is a communication interface for wirelessly communicating with the vehicle 1. For example, the communication I / F unit 31 wirelessly communicates with the vehicle 1 based on a standard such as 4G or 5G.

The storage unit 32 stores various data necessary for learning the parameters of the model. For example, the storage unit 32 stores the probe information 32a, the travel time data 32b, the classified input / output data set 32c, and the learned model 32d.

The probe information 32a is information indicating the traveling position and traveling time of the vehicle 1 for each vehicle 1. The travel time data 32b is information indicating an observed value of the travel time of the road link, and is the travel time of the vehicle 1 that has passed the road link between the previous time t-1 and the time t at each time t. Shows the average value of. The classified input / output data set 32c is a data set extracted from the travel time data 32b. The classified input / output data set 32c is a learning data set that is classified according to the phenomenon of traffic flow and is used for learning the parameters of the model. The trained model 32d is information indicating the parameters of the linear regression model used for predicting the travel time generated by the model learning device 3. A linear regression model is provided for each type of traffic flow phenomenon that occurs at the predicted link.

The probe information acquisition unit 33 acquires probe information by receiving probe information from the vehicle 1 via the communication I / F unit 31. The probe information acquisition unit 33 stores the acquired probe information in the storage unit 32 as the probe information 32a.

The travel time observation unit 34 reads the probe information 32a from the storage unit 32. Based on the read probe information 32a, the travel time observation unit 34 determines the travel time of the vehicle 1 that has passed the road link between the previous time t-1 and the time t for each time t up to the current time. Is observed for each vehicle. That is, the travel time observation unit 34 extracts the vehicle 1 whose passing time at the most downstream position of the road link is between the time t-1 and the time t. The travel time observation unit 34 observes the travel time by calculating the time required for the extracted vehicle 1 to travel from the most upstream position to the most downstream position of the road link based on the probe information 32a. The travel time observation unit 34 calculates the average value of the travel time observed at each time t, and stores it in the storage unit 32 as the travel time data 32b.

The phenomenon prediction unit 35 functions as a phenomenon acquisition unit, reads travel time data 32b from the storage unit 32, and predicts a traffic flow phenomenon generated at the prediction target link based on the travel time data 32b. The phenomenon prediction unit 35 determines at each time point that the traffic flow phenomenon at that time point is any of the above-mentioned first to fifth phenomena. The method of predicting the phenomenon of the traffic flow is the same as the prediction method by the phenomenon prediction unit 25 of the travel time prediction device 2. Therefore, the detailed description will not be repeated.

The data classification unit 36 classifies the data included in the travel time data 32b stored in the storage unit 32 based on the prediction result of the phenomenon by the phenomenon prediction unit 35, and generates an input / output data set for each phenomenon. The data classification unit 36 stores the generated input / output data set in the storage unit 32 as the classified input / output data set 32c.

The model learning unit 37 learns the parameters of the linear regression model for each traffic flow phenomenon based on the classified input / output data set 32c classified for each traffic flow phenomenon. The model learning unit 37 stores the learned parameters in the storage unit 32 as the learned model 32d.

<I / O data set classification method and model parameter learning method>
Next, the method of classifying the input / output data set and the method of learning the model parameters by the model learning device 3 will be described in detail.

FIG. 14 is a diagram schematically showing an input / output data set classification process and a model parameter learning process.

The data classification unit 36 extracts an input / output data set (hereinafter, referred to as “input / output data set t”) at time t from the travel time data 32b. The input / output data set t is composed of a set of an input data set 8 and an output data 9. As described with reference to FIG. 9A, for example, the input data set 8 includes the travel times of links 4-6 observed at time t and the travel times of links 4-6 observed at time t-1. , The travel time of links 4 to 6 observed at time t-2 and the travel time of links 4 to 6 observed at time t-3. That is, the input data set 8 is the current time and the travel time of links 4 to 6 from the current time to three periods before. The output data 9 is the travel time of the link 5 observed at time t + 2. That is, the output data 9 is the travel time of the link 5 two periods ahead of the current time.

The data classification unit 36 classifies the input / output data set t based on the phenomenon of the traffic flow predicted to occur at time t (ST11). Here, it is assumed that the traffic flow phenomenon is predicted by the phenomenon prediction unit 35 based on the phenomenon prediction data set 10. The first model and the fifth model used in the first phenomenon and the fifth phenomenon do not require learning. Therefore, the data classification unit 36 classifies the input / output data set t only when the traffic flow phenomenon corresponds to any of the second phenomenon to the fourth phenomenon. For example, the second phenomenon is classified into input / output data sets t-2, t-5 and t-10. Further, the input / output data sets t-1, t-3, t-4, t-6 and t-9 are classified into the third phenomenon. Further, the input / output data sets t-7 and t-8 are classified into the fourth phenomenon.

The model learning unit 37 learns a linear regression model for each phenomenon by calculating the parameters of the linear regression model using the least squares method based on the input / output data set classified for each traffic flow phenomenon (). ST12a to ST12c). That is, the model learning unit 37 has a model parameter for each phenomenon that minimizes the square error between the predicted value of the travel time predicted from the input data set 8 and the observed value of the travel time indicated by the output data 9. Is calculated. The details of the least squares method are as described above. Therefore, the detailed description will not be repeated.

<Model learning process>
FIG. 15 is a flowchart showing a procedure of model learning processing by the model learning device 3.

With reference to FIG. 15, the model learning device 3 repeats the following steps S11 to S13 for the input / output data set for each time of the prediction target link (loop A).

That is, the phenomenon prediction unit 35 reads out the phenomenon prediction data set 10 from the travel time data 32b stored in the storage unit 32, and the data classification unit 36 reads the travel time data 32b stored in the storage unit 32. The input / output data set is read from the inside (S11).

The phenomenon prediction unit 35 predicts a traffic flow phenomenon based on the read phenomenon prediction data set 10 (S12). The phenomenon prediction unit 35 outputs the prediction result of the traffic flow phenomenon to the data classification unit 36.

The data classification unit 36 receives the prediction result of the traffic flow phenomenon from the phenomenon prediction unit 35, classifies the read input / output data set for each phenomenon, and stores it in the storage unit 32 as the classified input / output data set 32c.

The model learning device 3 repeats the following steps S14 and S15 for each traffic flow phenomenon (loop B).

That is, the model learning unit 37 reads the input / output data set from the classified input / output data set 32c stored in the storage unit 32 (S14).

The model learning unit 37 calculates the parameters of the linear regression model by the least squares method based on the read input / output data set. The model learning unit 37 stores the calculated parameters in the storage unit 32 as the learned model 32d (S15).

<About the prediction result by the travel time prediction device>
With reference to FIG. 3, the prediction result of predicting the travel time by the travel time prediction device 2 will be described using the trained model learned by the model learning device 3.

The fine solid line shows the travel time predicted value 43 by the travel time predicting device 2. Further, the predicted value 44 indicates a predicted value of the travel time using the second model in the time zone in which the traffic flow phenomenon is determined to be the second phenomenon. The predicted value 45 indicates a predicted value of the travel time using the fourth model in the time zone in which the traffic flow phenomenon is determined to be the fourth phenomenon.

Comparing the observed value 41, the predicted value 42, and the predicted value 44, 14 out of 15 predicted values 44 are closer to the observed value 41 than the predicted value 42, and one is closer to the observed value than the predicted value 42. The result was far from 41. Comparing the observed value 41, the predicted value 42, and the predicted value 45, 21 of the 25 predicted values 45 are closer to the observed value 41 than the predicted value 42, and 4 are closer to the predicted value 42 than the predicted value 42. The result is far from the observed value 41. In other words, it was shown that the travel time can be predicted with higher accuracy when the linear regression model is changed for each phenomenon than when it is not changed.

The inventor defined a linear regression model under the following model conditions 1 to 3.
(Model condition 1) An input data set for prediction is extracted from the travel time observed on the road links existing in the range of 10 km upstream and 10 km downstream of the prediction target link.
(Model condition 2) The interval of time t is set to an interval of 5 minutes, and the travel time one hour ahead is predicted from the travel time observed at the road link at time t and time t-1.
(Model condition 3) Set a forecast target link near the Yokohama-Aoba interchange on the Tomei Expressway.
In addition, the inventor trained the linear regression model under the following learning conditions.
(Learning condition 1) The input / output data set for learning is extracted from the travel time observed at the road links existing in the range of 10 km upstream and 10 km downstream of the predicted target link.
(Learning condition 2) The learning period of the linear regression model is 9 months from July 2016 to March 2017.

In addition, the inventor conducted an experiment of predicting travel time under the following prediction conditions 1 and 2.
(Forecast condition 1) 13 traffic jams with a duration of more than 3.5 hours included in the period from April 2017 to June 2017 are extracted.
(Prediction condition 2) In the period when the traffic flow phenomenon is judged to be the second phenomenon and the fourth phenomenon, the proximity to the observed value is determined by changing the linear regression model for each phenomenon and the linear regression model for each phenomenon. Compare with the case without change.

As a result of the experiment, it was possible to obtain a prediction closer to the observed value by changing the linear regression model for each phenomenon than when not changing it. The same result as in FIG. 3 was obtained in 10 out of 13 traffic jams. It was. For the remaining three, there was not much difference between the two. As a result, the model learning device 3 was able to show the superiority of the model learning and the travel time prediction by the travel time prediction device 2.

<Effect of embodiment>
As described above, it is possible to predict the travel time at the predicted target link by using a model that matches the traffic flow phenomenon predicted to occur at the predicted target link according to the embodiment. In this way, the travel time can be predicted using the model most suitable for the prediction of the travel time. Therefore, it is possible to predict the travel time with high accuracy.

The phenomenon of traffic flow is predicted using not only the travel time observed at the predicted target link but also the travel time observed at at least one of the road links upstream and downstream of the predicted target link. Since the phenomenon can be predicted in consideration of the situation around the prediction target link in this way, the phenomenon can be predicted accurately.

Further, the traffic flow phenomenon that will occur in the prediction target link in the future is predicted based on the phenomenon that is currently occurring in the prediction target link, as shown in FIG.

In addition, the above phenomenon is related to the traffic congestion that occurs at the road link. Therefore, the phenomenon can be predicted in consideration of the change in the traffic condition. Congestion has a significant impact on travel time predictions. Therefore, it is possible to predict the travel time with high accuracy.

In addition, the traveling direction of the end position of the traffic jam that occurs in a plurality of links including the prediction target link is determined, and the traffic flow phenomenon is predicted based on the determination result of the traveling direction. By determining the traveling direction of the end of the traffic jam, for example, it can be determined that the end of the traffic jam existing downstream of the prediction target link is approaching the prediction target link and will eventually be involved in the traffic jam. In addition, the end of the traffic jam existing upstream of the forecast target link has moved downstream, and it can be determined that the traffic jam will be eliminated eventually. As a result, the phenomenon can be predicted accurately.

In addition, the traveling direction at the end of the traffic jam is determined based on the time change of the travel time at the road link located upstream or downstream of the predicted target link. The greater the proportion of congested sections to the road links, the greater the travel time. Therefore, the traveling direction of the end position of the traffic jam can be accurately determined.

It is also possible to determine the end position of a traffic jam by determining whether or not there is a traffic jam on a plurality of road links. Therefore, the traveling direction of the end position of the traffic jam can be accurately determined from the time change of the end position.

In addition, a model is learned for each traffic flow phenomenon that occurs at the predicted link. That is, the trained model is a model that matches the traffic flow phenomenon that is predicted to occur at the predicted link. Therefore, by predicting the phenomenon of the traffic flow and predicting the travel time by using the model corresponding to the predicted phenomenon, the travel time can be predicted by using the optimum model for predicting the travel time. Therefore, it is possible to predict the travel time with high accuracy.

[Modification example]
In the traffic flow phenomenon prediction process shown in FIG. 5 (step S2 in FIG. 10), if the end of the traffic jam is not far from the prediction target link (NO in S22, NO in S27), the end of the traffic jam becomes the prediction target link. The traffic flow was predicted assuming that it was approaching. However, if the end of the traffic jam has not moved, it cannot be said that the end of the traffic jam is approaching the predicted link. Therefore, it may be determined whether or not the end of the traffic jam is close to the prediction target link.

FIG. 16 is a flowchart showing details of the traffic flow phenomenon prediction process (step S2 in FIG. 10). The difference from FIG. 11 is that steps S31 and S32 have been added.

That is, when congestion occurs at the prediction target link and the end of the congestion is not far from the prediction target link (YES in S21, NO in S22), the phenomenon prediction unit 25 has the prediction target link at the end of the congestion. It is determined whether or not the object is approaching (S31).

When the end of the traffic jam is approaching the prediction target link (YES in S31), the phenomenon prediction unit 25 determines that the traffic flow phenomenon corresponds to the fourth phenomenon (S24). If the end of the traffic jam is not close to the prediction target link (NO in S31), the phenomenon prediction unit 25 determines that there is no change in the end of the traffic jam and does not predict the phenomenon.

Further, when there is no congestion at the prediction target link, but there is congestion in the surroundings, and the end of the congestion is not far from the prediction target link (NO in S21, NO in S25, NO in S27), the phenomenon The prediction unit 25 determines whether or not the end of the traffic jam is approaching the prediction target link (S32).

When the end of the traffic jam is approaching the prediction target link (YES in S32), the phenomenon prediction unit 25 determines that the traffic flow phenomenon corresponds to the second phenomenon (S29). If the end of the traffic jam is not close to the prediction target link (NO in S32), the phenomenon prediction unit 25 determines that there is no change in the end of the traffic jam and does not predict the phenomenon.

It should be noted that the travel time may not be predicted at the time when the phenomenon prediction unit 25 does not predict the traffic flow phenomenon, or even if the traffic flow phenomenon at the previous time is maintained. Good. In the latter case, the travel time is predicted using a model corresponding to the phenomenon of the traffic flow at the previous time.

Further, the phenomenon prediction unit 25 predicts that the traffic flow phenomenon corresponds to the sixth phenomenon, which is a new phenomenon, at the time when it is determined that there is no change at the end of the traffic jam (NO in S31 or NO in S32). May be good. In this case, the travel time prediction unit 27 may predict the travel time according to a sixth model different from the first model to the fifth model.

[Additional Notes]
Although the travel time prediction device 2 and the model learning device 3 according to the embodiment of the present disclosure have been described above, the present disclosure is not limited to this embodiment.

For example, in the above embodiment, the phenomenon of traffic flow was predicted based on the travel time observed at the road link. However, if a sensor such as an image-type vehicle detector or an ultrasonic vehicle detector installed on the roadside can detect the location of congestion or the end position of congestion, the traffic flow is based on the detection result of the sensor. The phenomenon of may be acquired.

Further, the method of classifying traffic flow phenomena is not limited to the above-mentioned first to fifth classifications. For example, a model may be provided for each time zone or for each other criterion such as each season to predict travel time.

Also, travel time observations are not limited to those based on probe information. For example, the travel time may be observed by reading the license plate of the vehicle 1 between two points and collating the vehicle 1.

In addition, each of the above devices is specifically a computer system composed of a microprocessor, a ROM (Read Only Memory), a RAM (Random Access Memory), an HDD (Hard Disk Drive), a display unit, a keyboard, a mouse, and the like. It may be configured as. For example, each device achieves its function by expanding the computer program stored in the HDD into RAM and operating the microprocessor according to the computer program.

In addition, some or all of the components constituting each of the above devices may be composed of one or a plurality of semiconductor devices such as system LSIs.

Further, the above-mentioned computer program or trained model may be recorded and distributed on a computer-readable non-volatile recording medium such as an HDD, a CD-ROM, or a semiconductor memory. Further, the computer program may be transmitted and distributed via a telecommunication line, a wireless or wired communication line, a network typified by the Internet, data broadcasting, or the like.
Further, each of the above devices may be realized by a plurality of computers.

In addition, some or all the functions of each of the above devices may be provided by cloud computing. That is, some or all the functions of each device may be realized by the cloud server. For example, in the travel time prediction device 2, the storage unit 22 is realized by a cloud server, and the travel time prediction device 2 acquires a learned model according to a traffic flow phenomenon from the cloud server and predicts the travel time. Good.

Further, at least a part of the above-described embodiment and the above-described modification may be arbitrarily combined.

It should be considered that the embodiments disclosed this time are exemplary in all respects and not restrictive. The scope of the present disclosure is shown by the scope of claims, not the above meaning, and is intended to include all modifications within the meaning and scope equivalent to the scope of claims.

1 Vehicle 2 Travel time prediction device 3 Model learning device 4 Prediction target link 5 Upstream link 5A Upstream link set 6 Downstream link 7 Congestion section 8 Input data set 9 Output data 10 Phenomenon prediction data set 21 Communication I / F unit 22 Storage unit 22a Probe information 22b Travel time data 22c Learned model 23 Probe information acquisition unit 24 Travel time observation unit 25 Phenomenon prediction unit 26 Model selection unit 27 Travel time prediction unit 28 Travel time provision unit 31 Communication I / F unit 32 Storage unit 32a Probe Information 32b Travel time data 32c Classified input / output data set 32d Trained model 33 Probe information acquisition unit 34 Travel time observation unit 35 Phenomenon prediction unit 36 Data classification unit 37 Model learning unit 41 Observation value 42 Prediction value 43 Prediction value 44 Prediction value 45 Predicted value

Claims (14)

It is a travel time prediction method that predicts the travel time of a vehicle.
The step of predicting the phenomenon of traffic flow that occurs at the predicted target link, which is the target road link for predicting the travel time,
A step of selecting a model corresponding to the predicted phenomenon from the models for predicting the travel time prepared for each traffic flow phenomenon, and a step of selecting the model corresponding to the predicted phenomenon.
A travel time prediction method including a step of predicting a travel time in the predicted target link based on the selected model. The travel time prediction method according to claim 1, wherein in the step of predicting the phenomenon, the phenomenon is predicted based on the travel time observed for a plurality of road links including the prediction target link. In the step of predicting the phenomenon, it occurs at the prediction target link at a second time point after the first time point based on the phenomenon of the traffic flow occurring at the prediction target link at the first time point. The travel time prediction method according to claim 1 or 2, which predicts a traffic flow phenomenon. The travel time prediction method according to claim 3, wherein the phenomenon is a traffic jam generated at a road link as an index. In the step of predicting the phenomenon, any one of claims 1 to 4, wherein the phenomenon is predicted by determining which of the first to fifth phenomena the prediction target link corresponds to. Described travel time prediction method.
First phenomenon: the prediction target link, a predetermined number of road links on the upstream side of the prediction target link in the traveling direction of the vehicle, and a predetermined number of road links on the downstream side of the prediction target link in the traveling direction of the vehicle. No congestion has occurred in any of the second phenomenon: No congestion has occurred at the prediction target link, and the end of the congestion occurring at a predetermined number of road links on the downstream side of the prediction target link is Third phenomenon approaching the prediction target link: Congestion is occurring at the prediction target link, and the end of the congestion occurring at a predetermined number of road links on the upstream side of the prediction target link is the prediction target. Fourth phenomenon of moving away from the link: Congestion is occurring at the prediction target link, and the end of the congestion occurring at the predetermined number of road links on the upstream side of the prediction target link approaches the prediction target link. Fifth phenomenon: There is no congestion at the prediction target link, and the end of the congestion occurring at the predetermined number of road links on the downstream side of the prediction target link is away from the prediction target link. In the step of predicting the phenomenon, the traveling direction of the end position of the traffic jam generated in the plurality of road links including the prediction target link is determined, and the phenomenon is predicted based on the determination result of the traveling direction. Alternatively, the travel time prediction method according to claim 5. The travel time prediction method according to claim 6, wherein in the step of predicting the phenomenon, the traveling direction is determined based on the time change of the travel time on the road link located upstream or downstream of the prediction target link. In the step of predicting the phenomenon, the end position of the traffic jam is determined by determining the congestion situation occurring in a plurality of road links including the prediction target link, and the traveling direction is determined based on the time change of the end position. The travel time prediction method according to claim 6, which is determined. It is a model learning method that learns a model for predicting the travel time of a vehicle.
The step of acquiring the phenomenon of the traffic flow generated in the predicted target link, which is the target road link for predicting the travel time using the model, and
A model learning method including a step of learning the parameters of the model corresponding to the acquired phenomenon based on the travel time observed for one or more road links including the predicted target link. It is a travel time prediction device that predicts the travel time of a vehicle.
A phenomenon prediction unit that predicts the phenomenon of traffic flow that occurs at the prediction target link, which is the target road link for predicting travel time,
A model selection unit that selects a model corresponding to the predicted phenomenon from among the models for predicting the travel time prepared for each traffic flow phenomenon.
A travel time prediction device including a travel time prediction unit that predicts a travel time in the prediction target link based on the selected model. A model learning device that learns a model for predicting the travel time of a vehicle.
A phenomenon acquisition unit that acquires the phenomenon of the traffic flow generated at the prediction target link, which is the target road link for which the travel time is predicted using the model, and the phenomenon acquisition unit.
A model learning device including a model learning unit that learns parameters of the model corresponding to the acquired phenomenon based on the travel time observed for one or more road links including the prediction target link. A computer program that allows a computer to function as a travel time predictor that predicts the travel time of a vehicle.
The computer
A phenomenon prediction unit that predicts the phenomenon of traffic flow that occurs at the prediction target link, which is the target road link for predicting travel time,
A model selection unit that selects a model corresponding to the predicted phenomenon from among the models for predicting the travel time prepared for each traffic flow phenomenon.
A computer program that functions as a travel time prediction unit that predicts the travel time in the prediction target link based on the selected model. A computer program for making a computer function as a model learning device for learning a model for predicting a vehicle's travel time.
The computer
A phenomenon acquisition unit that acquires the phenomenon of the traffic flow generated at the prediction target link, which is the target road link for which the travel time is predicted using the model, and the phenomenon acquisition unit.
A computer program that functions as a model learning unit that learns the parameters of the model corresponding to the acquired phenomenon based on the travel time observed for one or more road links including the predicted target link. A model set that includes a set of trained models for making a computer function to predict the travel time of a vehicle at the predicted link, which is the road link for which the travel time is predicted.
The trained model accepts the observed travel time for one or more road links including the predicted link as input and causes the computer to function to output the future travel time at the predicted link.
The trained model is associated with each traffic flow phenomenon generated at the predicted target link.
The phenomenon of the traffic flow generated in the predicted target link is acquired, and the parameters of the model corresponding to the acquired phenomenon are learned based on the travel time observed for one or more road links including the predicted target link. A trained model set in which the trained model is generated.

Images