Traffic information prediction device, traffic information prediction method, and traffic information prediction program

Abstract

To provide a traffic information prediction device, a traffic information prediction method, and a traffic information prediction program for easily acquiring traffic information prediction results and improving the accuracy of the prediction results.SOLUTION: A traffic information prediction device includes: a first acquisition unit inputting a first congestion length, a first travel speed, and a first traffic amount at a first point in time into a learned model, and acquiring a second travel speed which is a predicted value of a travel speed at a second point in time later than the first point in time output from the learned model and a second traffic amount which is a predicted value of the traffic amount at the second point in time; a second acquisition unit inputting the second travel speed and the second traffic amount acquired by the first acquisition unit into a mathematical model and performing a prediction process to acquire a second congestion length that is a predicted value of the congestion length at the second point in time output from the mathematical model; and an output unit that outputs the second congestion length acquired by the second acquisition unit.SELECTED DRAWING: Figure 3

Description

[Technical field]

[0001] The present disclosure relates to a traffic information prediction device, a traffic information prediction method, and a traffic information prediction program. [Background technology]

[0002] Conventionally, a technology has been developed for predicting traffic information such as congestion length by performing a simulation. For example, Patent Document 1 (JP Patent Publication No. 2013-210941) discloses the following technology. That is, a traffic flow prediction device quantifies the relationship between traffic density and traffic volume in an arbitrarily specified area from probe information consisting of time, latitude and longitude information, which is a traffic flow observation value, or probe traffic information consisting of time, section number, number of vehicles passing through the section, and average required time for passing through the section, and estimates traffic volume using the relationship, and includes a first approximation curve calculation means for calculating a first approximation curve showing the relationship between traffic density and traffic volume in the area based on the probe traffic information of vehicles that have traveled through the area, a second approximation curve calculation means for calculating a second approximation curve showing the relationship between traffic density and traffic volume in the area based on the result of a traffic flow simulation using a traffic flow simulation model, and an extension curve for matching the first approximation curve with the second approximation curve. the traffic volume and link capacity so as to minimize the distance between the value indicating the relationship between traffic density and traffic volume in the area obtained by multiplying the expansion rate by each of the values ​​indicating the relationship between traffic density and traffic volume in the area when the first approximation curve is obtained, and the value indicating the relationship between traffic density and traffic volume in the area when the second approximation curve is obtained; and a result determination means for repeating the process of determining the traffic volume and the link capacity by the parameter adjustment means and the execution of the traffic flow simulation using the traffic volume and the link capacity determined by the parameter adjustment means as inputs until the result of executing the traffic flow simulation satisfies a predetermined termination condition. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] JP 2013-210941 A Summary of the Invention [Problem to be solved by the invention]

[0004] In the technology described in Patent Document 1, the wider the road network to be simulated is, or the more meshes the road network is divided into in the simulation, the longer it takes to obtain prediction results. In addition, it takes a lot of time and effort to search for optimal values ​​for parameters used in the simulation.

[0005] In addition, when predicting congestion length on roads where congestion occurs infrequently using a trained model, the teacher data used in statistical learning to generate the trained model contains a lot of traffic information during non-congestion periods. In this case, the prediction accuracy of congestion length by the trained model generated by the statistical learning may decrease. A technology is desired that can easily obtain traffic information prediction results and improve the accuracy of the prediction results.

[0006] The present disclosure has been made to solve the above-mentioned problems, and its purpose is to provide a traffic information prediction device, a traffic information prediction method, and a traffic information prediction program that are capable of easily obtaining traffic information prediction results and improving the accuracy of the prediction results. [Means for solving the problem]

[0007] The traffic information prediction device disclosed herein includes a first acquisition unit that inputs a first congestion length, a first travel speed, and a first traffic volume at a first time into a learned model, and acquires a second travel speed which is a predicted value of the travel speed at a second time after the first time and a second traffic volume which is a predicted value of the traffic volume at the second time output from the learned model, a second acquisition unit that performs a prediction process that inputs the second travel speed and the second traffic volume acquired by the first acquisition unit into a mathematical model, and acquires a second congestion length which is a predicted value of the congestion length at the second time output from the mathematical model, and an output unit that outputs the second congestion length acquired by the second acquisition unit.

[0008] One aspect of the present disclosure can be realized not only as a traffic information prediction device having such a characteristic processing unit, but also as a semiconductor integrated circuit that realizes part or all of the traffic information prediction device, or as a system that includes the traffic information prediction device. Effect of the Invention

[0009] According to the present disclosure, it is possible to easily obtain predicted results of traffic information and improve the accuracy of the predicted results. [Brief description of the drawings]

[0010] [Figure 1] FIG. 1 is a diagram illustrating a configuration of a traffic information prediction system according to an embodiment of the present disclosure. [Diagram 2] FIG. 2 is a diagram for explaining a congestion length measured by a congestion length sensor in the traffic information prediction system according to the embodiment of the present disclosure. [Diagram 3] FIG. 3 is a diagram illustrating a configuration of a traffic information prediction device according to an embodiment of the present disclosure. [Figure 4] FIG. 4 is a diagram for explaining a method for calculating a predicted value of a congestion length by the traffic information prediction device according to the embodiment of the present disclosure. [Diagram 5] FIG. 5 is a diagram for explaining the re-learning process performed by the traffic information prediction device according to the embodiment of the present disclosure. [Figure 6] FIG. 6 is a diagram for explaining the re-learning process performed by the traffic information prediction device according to the embodiment of the present disclosure. [Figure 7] FIG. 7 is a flowchart defining an operation procedure when the traffic information prediction device according to the embodiment of the present disclosure performs a prediction process. [Figure 8] FIG. 8 is a flowchart that defines an operation procedure when the traffic information prediction device according to the embodiment of the present disclosure performs the re-learning process. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0011] First, the contents of the embodiments of the present disclosure will be listed and described. (1) A traffic information prediction device according to an embodiment of the present disclosure includes a first acquisition unit that inputs a first congestion length, a first travel speed, and a first traffic volume at a first time into a learned model, and acquires a second travel speed that is a predicted value of the travel speed at a second time after the first time and a second traffic volume that is a predicted value of the traffic volume at the second time output from the learned model, a second acquisition unit that performs a prediction process that inputs the second travel speed and the second traffic volume acquired by the first acquisition unit into a mathematical model, and acquires a second congestion length that is a predicted value of the congestion length at the second time output from the mathematical model, and an output unit that outputs the second congestion length acquired by the second acquisition unit.

[0012] In this way, by using a trained model to predict travel speed and traffic volume, and by using the trained model to predict traffic congestion length, the trained model can predict traffic congestion length in a shorter time than by using a simulation to predict travel speed, traffic volume, and traffic congestion length. In addition, since the trained model can predict traffic congestion length using a mathematical model that mathematically describes the relationship between traffic congestion length, travel speed, and traffic volume based on knowledge of traffic engineering, the trained model can suppress output of erroneous traffic congestion length compared to a trained model that predicts traffic congestion length using teacher data that includes a lot of traffic information during non-traffic jams. Therefore, the traffic information prediction result can be easily obtained, and the accuracy of the prediction result can be improved.

[0013] (2) In the above (1), the second acquisition unit may further acquire the correction information output from the learned model, and the second acquisition unit may input the second travel speed and the second traffic volume acquired by the first acquisition unit, as well as the acquired correction information, to the mathematical model, and may perform the prediction process to acquire the second congestion length output from the mathematical model.

[0014] With this configuration, for example, in a prediction process using a mathematical model, it is possible to reduce the error relative to the true congestion length at the time being predicted, thereby further improving the accuracy of the output results of the mathematical model.

[0015] (3) In the above (1) or (2), the output unit may further output the second travel speed and the second traffic volume acquired by the first acquisition unit in correspondence with the second congestion length acquired by the second acquisition unit.

[0016] A mathematical model is generated deductively according to theoretical grounds such as physical laws. On the other hand, a trained model is generated inductively, for example, by learning teacher data. Therefore, the cause of the prediction result by the trained model is difficult for a user to grasp, compared to the cause of the prediction result by the mathematical model. As described above, by using a mathematical model to predict the traffic congestion length, and outputting the predicted values ​​of the travel speed and traffic volume input to the mathematical model in correspondence with the predicted value of the traffic congestion length, the predicted values ​​of the travel speed and traffic volume that are the basis of the predicted value of the traffic congestion length are easier for the user to grasp, compared to a configuration in which a trained model is used to predict the traffic congestion length, and therefore the user can be given a sense of security and satisfaction with the prediction result of the mathematical model.

[0017] (4) In any of (1) to (3) above, the second acquisition unit may use, as the mathematical model in the prediction process, a second calculation formula for congestion length, which is a modification of a first calculation formula in an hourglass model that outputs a travel speed.

[0018] In the field of traffic engineering, the hourglass model is used, for example, in the prediction process of travel speed. As described above, in the prediction process of congestion length, by using a mathematical model that is a modification of the existing hourglass model, it is possible to mathematically describe the relationship between congestion length, travel speed, and traffic volume using information on travel speed and information on traffic volume that can be easily obtained in a traffic control system.

[0019] (5) In any of (1) to (4) above, the traffic information prediction device may further include a learning processing unit that performs a re-learning process to train the trained model using teacher data in which travel speed, traffic volume, and congestion length at a predetermined time are associated with each other, and a loss function related to a first difference between the second congestion length and a true congestion length at the second time, a second difference between the second travel speed and the true travel speed at the second time, and a third difference between the second traffic volume and the true traffic volume at the second time, wherein the loss function may include a first weighting coefficient, a second weighting coefficient, and a third weighting coefficient that are multiplied by the first difference, the second difference, and the third difference, respectively, and the learning processing unit may change the first weighting coefficient to a value different from the second weighting coefficient and the third weighting coefficient in the re-learning process depending on the magnitude of the congestion length in the teacher data.

[0020] With this configuration, for example, when the training data includes the length of a traffic jam during a serious traffic jam, which occurs infrequently, a loss function with a large weighting coefficient that is multiplied by the difference between the predicted traffic jam length and the true traffic jam length can be used in the re-learning process of the trained model to obtain output results of the trained model that reflect the traffic jam length during such a serious traffic jam, thereby further improving the accuracy of the prediction results.

[0021] (6) In any of (1) to (5) above, the first acquisition unit may further input to the trained model at least one of first information regarding a specified section of a road, the first information affecting traffic in the specified section or a section different from the specified section, and second information regarding periodicity of the traffic in the specified section.

[0022] In this way, by further inputting information that may be a cause of congestion in the road section being predicted into the trained model, the accuracy of the output results of the trained model can be improved.

[0023] (7) A traffic information prediction method according to an embodiment of the present disclosure is a traffic information prediction method in a traffic information prediction device, and includes the steps of inputting a first congestion length, a first travel speed, and a first traffic volume at a first time into a learned model, and acquiring a second travel speed which is a predicted value of the travel speed at a second time after the first time and a second traffic volume which is a predicted value of the traffic volume at the second time which are output from the learned model, inputting the acquired second travel speed and the second traffic volume into a mathematical model, and performing a prediction process to acquire a second congestion length which is a predicted value of the congestion length at the second time which is output from the mathematical model, and outputting the acquired second congestion length.

[0024] In this way, by using a trained model to predict travel speed and traffic volume, and by using the trained model to predict traffic congestion length, the trained model can predict traffic congestion length in a shorter time than by using a simulation to predict travel speed, traffic volume, and traffic congestion length. In addition, since the trained model can predict traffic congestion length using a mathematical model that mathematically describes the relationship between traffic congestion length, travel speed, and traffic volume based on knowledge of traffic engineering, the trained model can suppress output of erroneous traffic congestion length compared to a trained model that predicts traffic congestion length using teacher data that includes a lot of traffic information during non-traffic jams. Therefore, the traffic information prediction result can be easily obtained, and the accuracy of the prediction result can be improved.

[0025] (8) A traffic information prediction program according to an embodiment of the present disclosure is a traffic information prediction program used in a traffic information prediction device, and causes a computer to function as a first acquisition unit that inputs a first congestion length, a first travel speed, and a first traffic volume at a first time into a learned model, and acquires a second travel speed that is a predicted value of the travel speed at a second time after the first time and a second traffic volume that is a predicted value of the traffic volume at the second time that are output from the learned model, a second acquisition unit that performs a prediction process that inputs the second travel speed and the second traffic volume acquired by the first acquisition unit into a mathematical model, and acquires a second congestion length that is a predicted value of the congestion length at the second time that is output from the mathematical model, and an output unit that outputs the second congestion length acquired by the second acquisition unit.

[0026] In this way, by using a trained model to predict travel speed and traffic volume, and by using the trained model to predict traffic congestion length, the trained model can predict traffic congestion length in a shorter time than by using a simulation to predict travel speed, traffic volume, and traffic congestion length. In addition, since the trained model can predict traffic congestion length using a mathematical model that mathematically describes the relationship between traffic congestion length, travel speed, and traffic volume based on knowledge of traffic engineering, the trained model can suppress output of erroneous traffic congestion length compared to a trained model that predicts traffic congestion length using teacher data that includes a lot of traffic information during non-traffic jams. Therefore, the traffic information prediction result can be easily obtained, and the accuracy of the prediction result can be improved.

[0027] Hereinafter, embodiments of the present disclosure will be described with reference to the drawings. In the drawings, the same or corresponding parts are denoted by the same reference numerals, and the description thereof will not be repeated. In addition, at least some of the embodiments described below may be arbitrarily combined.

[0028] [Traffic Information Prediction System] 1 is a diagram showing a configuration of a traffic information prediction system according to an embodiment of the present disclosure. Referring to FIG. 1, a traffic information prediction system 501 includes a traffic information prediction device 101, one or more congestion length sensors 51, one or more traffic volume sensors 52, and one or more optical beacons 53.

[0029] The traffic information prediction device 101 performs a prediction process to predict a congestion length L in a predetermined segment, that is, a predetermined section (hereinafter also referred to as a "target section A") of a road R. The traffic information prediction device 101 is, for example, a server.

[0030] (Traffic jam length detector) The congestion length detector 51 is, for example, a vehicle detector that measures a congestion length L in the target section A.

[0031] FIG. 2 is a diagram for explaining a congestion length measured by a congestion length sensor in the traffic information prediction system according to the embodiment of the present disclosure.

[0032] 1 and 2, the target section A includes one or more intersections CS. In the example shown in Fig. 2, the target section A includes one intersection CS. The congestion length sensor 51 is provided, for example, at the intersection CS.

[0033] The congestion length sensor 51 measures, for example, the congestion length L from the downstream end of the target section A. The downstream end of the target section A is, for example, a stop line D of an entrance road to an intersection CS. The entrance road is a road for the vehicle 1 to enter the intersection CS.

[0034] The downstream end of the target section A is not limited to the stop line D, but may be a point a predetermined distance away from the intersection CS.

[0035] The congestion length detector 51 measures the length from the position of the stop line D to the position of the rearmost vehicle 1 as the congestion length L.

[0036] The congestion length sensor 51 measures the congestion length L and transmits the measurement result, for example, periodically. Specifically, the congestion length sensor 51 measures the congestion length L, for example, every 50 seconds. Then, the congestion length sensor 51 transmits congestion length information including the measurement result and the measurement time to the traffic information prediction device 101 via the network 151.

[0037] (traffic sensor) 1 again, the traffic volume detector 52 is, for example, a vehicle detector that measures a traffic volume Q in the target section A. For example, in the target section A, a plurality of traffic volume detectors 52 are provided.

[0038] More specifically, for example, each traffic volume sensor 52 is an image-type vehicle sensor. Specifically, for example, each traffic volume sensor 52 captures an image of the target section A and measures the traffic volume Q based on the image capture result. Note that the traffic volume sensor 52 is not limited to an image-type vehicle sensor, and may be an optical vehicle sensor or an ultrasonic vehicle sensor, etc.

[0039] Each traffic volume sensor 52 measures the traffic volume Q and transmits the measurement results, for example, periodically. Specifically, each traffic volume sensor 52 measures the traffic volume Q, for example, every 50 seconds. Then, each traffic volume sensor 52 transmits traffic volume information including its own ID (identifier), the measurement result, and the measurement time to the traffic information prediction device 101 via the network 151. Here, each traffic volume sensor 52 measures the traffic volume Q and transmits the traffic volume information at the same timing as the congestion length sensor 51.

[0040] (Light beacon) The optical beacon 53 is provided on the roadside in the target section A. The optical beacon 53 measures, for example, the passing time of the vehicle 1 passing through the target section A. For example, a plurality of optical beacons 53 are provided in the target section A.

[0041] More specifically, for example, the vehicle 1 is provided with an optical beacon vehicle-mounted device 54 capable of communicating with the optical beacon 53.

[0042] For example, when the vehicle 1 passes near the optical beacon 53, the optical beacon vehicle unit 54 communicates with the optical beacon 53 and transmits to the optical beacon 53 the ID of the vehicle 1 and passage time information indicating the passage time.

[0043] When the optical beacon 53 receives passage time information from the optical beacon vehicle-mounted device 54 of each vehicle passing through the target section A, the optical beacon 53 transmits the received plurality of pieces of passage time information to the traffic information prediction device 101 via the network 151 .

[0044] More specifically, each optical beacon 53 transmits, for example, periodically to the traffic information prediction device 101, vehicle passage information including its own ID and the ID and passage time of the vehicle 1 indicated in the passage time information received from the optical beacon on-board device 54 of each vehicle 1. Specifically, each optical beacon 53 transmits the vehicle passage information to the traffic information prediction device 101, for example, every 50 seconds.

[0045] [Traffic Information Prediction Device Configuration] Fig. 3 is a diagram showing a configuration of a traffic information prediction device according to an embodiment of the present disclosure. Referring to Fig. 3, the traffic information prediction device 101 includes a creation unit 11, a first acquisition unit 12, a second acquisition unit 13, an output unit 14, a learning processing unit 15, and a storage unit 16. A part or all of the creation unit 11, the first acquisition unit 12, the second acquisition unit 13, the output unit 14, and the learning processing unit 15 are realized, for example, by a processing circuit including one or more processors. The storage unit 16 is, for example, a non-volatile memory included in the processing circuit.

[0046] (Creation Department) <Time series traffic congestion length data> For example, the creation unit 11 creates time-series congestion length data based on congestion length information from the congestion length sensor 51.

[0047] More specifically, for example, when the creation unit 11 receives congestion length information from the congestion length sensor 51 , the creation unit 11 stores the received congestion length information in the storage unit 16 .

[0048] The creation unit 11 creates time-series congestion length data that associates the congestion length in the target section A with the measurement time based on multiple pieces of congestion length information stored in the memory unit 16 every time a certain amount of time has elapsed since its traffic information prediction device 101 is started.

[0049] In the conversion process, the creation unit 11 performs, for example, a smoothing process on the created time-series congestion length data.

[0050] As described above, the congestion length sensor 51 measures the congestion length in the target section A, for example, every 50 seconds. Therefore, the time-series congestion length data before the smoothing process indicates the congestion length every 50 seconds in the target section A. The creation unit 11 converts the time-series congestion length data created in the smoothing process, for example, into data indicating the congestion length every 5 minutes in the target section A. Then, the creation unit 11 outputs the time-series congestion length data after the smoothing process (hereinafter also referred to as "traffic jam length data L1") to the first acquisition unit 12.

[0051] <Time series traffic volume data> Also, for example, the creation unit 11 creates time-series data on statistical values ​​of traffic volume in the target section A based on traffic volume information from each traffic volume detector 52.

[0052] More specifically, for example, when the creation unit 11 receives traffic volume information from each traffic volume sensor 52 , the creation unit 11 stores the received multiple pieces of traffic volume information in the storage unit 16 .

[0053] The creation unit 11 creates time-series traffic volume data in which the measurement results of traffic volume Q correspond to the measurement times for each ID of the traffic volume sensor 52, based on the multiple traffic volume information stored in the memory unit 16, every time a certain time period has elapsed since the startup of its own traffic information prediction device 101.

[0054] Then, the creation unit 11 performs a smoothing process on the created time-series traffic volume data in the same manner as the smoothing process on the time-series congestion length data. That is, the creation unit 11 converts the created time-series traffic volume data into data indicating the measurement results of each traffic volume sensor 52 in the target section A every 5 minutes.

[0055] When the creation unit 11 completes the smoothing process on the time-series traffic volume data, it calculates, as a statistical value, an average value of the measurement results of the multiple traffic volume sensors 52 for each measurement time of the traffic volume Q, based on the time-series traffic volume data after the smoothing process. That is, the creation unit 11 creates time-series data (hereinafter also referred to as "traffic volume data Q1") in which the average value of the traffic volume Q is associated with the measurement time of the traffic volume Q. Then, the creation unit 11 outputs the created traffic volume data Q1 to the first acquisition unit 12.

[0056] The creation unit 11 is not limited to calculating the average value of the traffic volume Q, and may calculate a representative value of the traffic volume Q in the target section A as the statistical value.

[0057] <Time series travel speed data> Furthermore, for example, the creation unit 11 creates time-series data on the statistical values ​​of the travel speeds of the multiple vehicles 1 passing through the target section A, based on the vehicle passage information from each optical beacon 53.

[0058] More specifically, for example, when generating unit 11 receives vehicle passage information from each optical beacon 53 , generating unit 11 stores the received vehicle passage information in storage unit 16 .

[0059] The storage unit 16 stores therein, for example, a position list indicating the correspondence between the ID of the optical beacon 53 and the position of the optical beacon 53 in the target section A.

[0060] The creation unit 11 acquires multiple vehicle passing information from the memory unit 16 every time a certain amount of time has elapsed since its traffic information prediction device 101 was started up, and by referring to the position list in the memory unit 16, identifies, for each vehicle passing information, the position of the optical beacon 53 corresponding to the ID of the optical beacon 53 indicated by the vehicle passing information.

[0061] Then, based on the multiple vehicle passage information acquired from the memory unit 16 and the positions of each identified optical beacon 53, the creation unit 11 creates time-series driving trajectory data for each vehicle ID that corresponds the position of the optical beacon 53 with the passing time indicated by the vehicle passage information.

[0062] When the creation unit 11 creates the time-series travel locus data, it performs a smoothing process on the time-series travel locus data in the same manner as the smoothing process on the time-series congestion length data and the time-series traffic volume data. That is, the creation unit 11 converts the created time-series travel locus data into data indicating the position of each vehicle 1 in the target section A every 5 minutes.

[0063] Upon completing the smoothing process on the time-series travel locus data, the creation unit 11 calculates statistical values ​​of the travel speeds S of the multiple vehicles 1 passing through the target section A.

[0064] Specifically, for example, the creation unit 11 creates time-series travel speed data for each ID of the vehicle 1, based on the time-series driving trajectory data after smoothing processing, in which the travel speed S of the vehicle 1 at the position of each optical beacon 53 corresponds to the passing time.

[0065] Then, based on the created time-series travel speed data, the creation unit 11 calculates an average value of the travel speeds S of the multiple vehicles 1 at each passing time as a statistical value, and creates time-series data (hereinafter also referred to as "travel speed data S1") in which the calculated average value is associated with the passing time. Then, the creation unit 11 outputs the created travel speed data S1 to the first acquisition unit 12.

[0066] The creation unit 11 may calculate a representative value of the travel speeds S of a plurality of vehicles 1 passing through the target section A as the statistical value, instead of the average value of the travel speeds S.

[0067] (Prediction of traffic jam length) For example, the storage unit 16 stores a trained model E1 for predicting a travel speed S and a traffic volume Q at a time t2 that is after a certain time t1. The time t2 is, for example, a predicted time that is one hour after the time t1. The time t1 is an example of a first time, and the time t2 is an example of a second time.

[0068] More specifically, for example, the trained model E1 predicts travel speed S and traffic volume Q at each of a plurality of predicted times included in a time period T2 that follows a certain time period T1. For example, each of time period T1 and time period T2 is one hour long, and time period T2 starts immediately after the end of time period T1.

[0069] The trained model E1 predicts, for example, the travel speed S and the traffic volume Q at regular time intervals from the start time of the time slot T2. Specifically, for example, the trained model E1 predicts the travel speed S and the traffic volume Q at five minute intervals from the start time of the time slot T2.

[0070] The storage unit 16 further stores a mathematical model E2 that outputs a predicted value of the congestion length L at time t2.

[0071] More specifically, for example, the mathematical model E2 is a model that mathematically describes the relationship between travel speed S, traffic volume Q, and traffic congestion length L in accordance with physical laws. The mathematical model E2 predicts the traffic congestion length L at each of a plurality of predicted times included in time period T2, for example. Specifically, for example, the mathematical model E2 predicts the traffic congestion length L at regular time intervals from the start time of time period T2. For example, the mathematical model E2 predicts the traffic congestion length L every 5 minutes from the start time of time period T2.

[0072] For example, the mathematical model E2 is expressed by the following formula (1). In formula (1), D is the length of the target section A. Sa is the average travel speed of multiple vehicles 1 passing through the target section A. Sb is the travel speed of vehicle 1 in a section of the target section A where no congestion occurs. C is a correction term for correcting errors in the prediction process. k and a are constants. The correction term C is an example of correction information.

[0073]

number

[0074] More specifically, for example, the correction term C is a term for correcting an error with respect to the true congestion length L0 at time t2.

[0075] Here, when generating the mathematical model, a condition related to traffic information in the target section A may be set that does not hold true in the actual target section A. In this case, even if the input value to the mathematical model is an accurate value, the output value from the mathematical model, i.e., the predicted value of the congestion length, may have an error with respect to the actual congestion length. Therefore, in this embodiment, the mathematical model E2 includes a correction term C for correcting the error.

[0076] FIG. 4 is a diagram for explaining a method for calculating a predicted value of a congestion length by the traffic information prediction device according to the embodiment of the present disclosure.

[0077] Referring to Figures 3 and 4, the first acquisition unit 12 inputs congestion length data L1, travel speed data S1 and traffic volume data Q1, section information regarding the target section A, and periodicity information regarding the periodicity of traffic in the target section A to the learned model E1.

[0078] More specifically, for example, the storage unit 16 further stores section information and periodicity information. The section information and periodicity information are registered in advance in the storage unit 16 by, for example, a user.

[0079] The section information is information that affects traffic in the target section A or a section different from the target section A.

[0080] Specifically, for example, the section information indicates at least one of the length D of the target section A, the number of entrance roads to the intersection CS in the target section A, the number of exit roads from the intersection CS, the number of lanes in the target section A, and the number of intersections CS in the target section A. The exit roads are roads for the vehicle 1 to exit from the intersection CS.

[0081] The periodicity information indicates, for example, the statistics of congestion length, the statistics of travel speed, and the statistics of traffic volume at the same time as the predicted time included in the time slot T2.

[0082] Specifically, for example, the periodicity information indicates the average congestion length, the average travel speed, and the average traffic volume at the same time as the predicted time included in the time slot T2.

[0083] When the first acquisition unit 12 receives the congestion length data L1, the traffic volume data Q1, and the travel speed data S1 from the creation unit 11, the first acquisition unit 12 acquires the learned model E1, the section information, and the periodicity information from the storage unit 16. Then, the first acquisition unit 12 inputs each data received from the creation unit 11, and the section information and the periodicity information acquired from the storage unit 16 to the learned model E1. The learned model E1 is generated by, for example, deep learning.

[0084] Then, the first acquisition unit 12 acquires a predicted value of travel speed S (hereinafter also referred to as "travel speed data S2") and a predicted value of traffic volume (hereinafter also referred to as "traffic volume data Q2"), as well as a correction term C, at each of a plurality of predicted time points included in time zone T2, which are output from the trained model E1 to which the congestion length data L1, traffic volume data Q1, travel speed data S1, section information, and periodicity information have been input. The travel speed data S2 indicates, for example, a predicted value of travel speed S every 5 minutes in time zone T2. The traffic volume data Q2 indicates, for example, a predicted value of traffic volume Q every 5 minutes in time zone T2.

[0085] Specifically, for example, the first acquisition unit 12 inputs the congestion length data L1, the traffic volume data Q1, the travel speed data S1, the section information, and the periodicity information to the trained model E1 every time it receives the congestion length data L1, the traffic volume data Q1, and the travel speed data S1 from the creation unit 11. Then, the first acquisition unit 12 acquires the travel speed data S2, the traffic volume data Q2, and the correction term C output from the trained model E1. That is, the first acquisition unit 12 acquires new output results of the trained model E1 every time it receives new congestion length data L1, traffic volume data Q1, and travel speed data S1 from the creation unit 11.

[0086] When the first acquisition unit 12 acquires the travel speed data S2, the traffic volume data Q2, and the correction term C, the first acquisition unit 12 outputs the acquired travel speed data S2, the traffic volume data Q2, and the correction term C to the second acquisition unit 13.

[0087] However, when predicting the length of congestion on road R, where congestion occurs infrequently, using a trained model, the training data used to generate the trained model will contain a lot of traffic information during non-congested periods, which may reduce the accuracy of predicting the length of congestion.

[0088] Through careful study, the inventors of the present application have discovered that predicting congestion length using a mathematical model based on the hourglass model can suppress a decrease in prediction accuracy compared to predicting congestion length using methods other than the mathematical model.

[0089] Upon receiving the travel speed data S2, the traffic volume data Q2, and the correction term C from the first acquisition unit 12, the second acquisition unit 13 performs a prediction process to acquire the congestion length L at time t2.

[0090] For example, in the prediction process, the second acquisition unit 13 uses the mathematical model E2 expressed by the above formula (1). The mathematical model E2 is a calculation formula for the congestion length that is a modification of the calculation formula in the hourglass model that outputs the travel speed.

[0091] More specifically, for example, when the second acquisition unit 13 receives the travel speed data S2, the traffic volume data Q2, and the correction term C from the first acquisition unit 12, the second acquisition unit 13 acquires the mathematical model E2 from the storage unit 16. Then, the second acquisition unit 13 inputs the travel speed data S2, the traffic volume data Q2, and the correction term C received from the first acquisition unit 12 into the mathematical model E2.

[0092] Then, the second acquisition unit 13 acquires the predicted values ​​of the congestion length L at each of the multiple predicted time points included in the time period T2 (hereinafter also referred to as "congestion length data L2") output from the mathematical model E2. Upon acquiring the congestion length data L2, the second acquisition unit 13 outputs the acquired congestion length data L2, as well as the travel speed data S2 and the traffic volume data Q2 received from the first acquisition unit 12, to the output unit 14.

[0093] Specifically, for example, every time the second acquisition unit 13 receives new travel speed data S2, traffic volume data Q2, and correction term C from the first acquisition unit 12, the second acquisition unit 13 inputs the travel speed data S2, traffic volume data Q2, and correction term C to the mathematical model E2 and acquires congestion length data L2 output from the mathematical model E2. The second acquisition unit 13 outputs the acquired congestion length data L2, and the travel speed data S2 and traffic volume data Q2 received from the first acquisition unit 12 to the output unit 14.

[0094] 3 again, the output unit 14 outputs the congestion length data L2 acquired by the second acquisition unit 13, and the travel speed data S2 and traffic volume data Q2 acquired by the first acquisition unit 12. More specifically, for example, the output unit 14 outputs the travel speed data S2 and the traffic volume data Q2 acquired by the first acquisition unit 12 in association with the congestion length data L2 acquired by the second acquisition unit 13.

[0095] Specifically, for example, when the output unit 14 receives the congestion length data L2, the travel speed data S2, and the traffic volume data Q2 from the second acquisition unit 13, it associates the travel speed data S2 and the traffic volume data Q2 with the congestion length data L2 and displays them on a display device not shown.

[0096] (Relearning process) For example, the learning processing unit 15 performs a re-learning process to re-learn the learned model E1 using the teacher data K and the loss function F. The teacher data K is, for example, data in which a true travel speed S0, a true traffic volume Q0, and a true congestion length L0 are associated with each predicted time included in the time slot T2.

[0097] More specifically, for example, the storage unit 16 stores teacher data K and a loss function F. The teacher data K and the loss function F are registered in advance in the storage unit 16 by a user.

[0098] The loss function F is a function related to the difference between the predicted value of the congestion length L at the prediction time and the true congestion length L0, the difference between the predicted value of the travel speed S at the prediction time and the true travel speed S0, and the difference between the predicted value of the traffic volume Q at the prediction time and the true traffic volume Q0. Specifically, the loss function F is expressed by the following formula (2). F=λ1×F1+λ2×F2+λ3×F3+λ4×F4 (2)

[0099] In formula (2), F1 is a function related to the difference between the predicted value of the congestion length L and the true congestion length L0. F2 is a function related to the difference between the predicted value of the congestion length L and the true congestion length L0 divided by the length D of the target section A. F3 is a function related to the difference between the predicted value of the travel speed S and the true travel speed S0. F4 is a function related to the difference between the predicted value of the traffic volume Q and the true traffic volume Q0. λ1, λ2, λ3, and λ4 are weighting coefficients multiplied by the functions F1, F2, F3, and F4, respectively. The weighting coefficients λ1 and λ2 are examples of a first weighting coefficient. The weighting coefficient λ3 is an example of a second weighting coefficient. The weighting coefficient λ4 is an example of a third weighting coefficient.

[0100] For example, the learning processing unit 15 changes the weighting coefficients λ1 and λ2 to values ​​different from the weighting coefficients λ3 and λ4 in accordance with the magnitude of the congestion length L in the teacher data K in the re-learning process.

[0101] More specifically, for example, when the true congestion length L0 in the teacher data K is equal to or greater than a predetermined threshold Th1, the learning processing unit 15 sets the weighting coefficients λ1 and λ2 to values ​​greater than the weighting coefficients λ3 and λ4.

[0102] 5 and 6 are diagrams for explaining the re-learning process by the traffic information prediction device according to the embodiment of the present disclosure. Fig. 5 is a diagram for explaining an example of the re-learning process in the early stage of re-learning when the number of re-learning times of the trained model E1, for example, the number of epochs, is small, and Fig. 6 is a diagram for explaining an example of the re-learning process in the later stage of re-learning when the number of epochs is large.

[0103] 3, 5 and 6, for example, in the early stage of re-learning, the learning processing unit 15 increases the frequency of performing the re-learning process using a data group (hereinafter also referred to as "data group N1") including the true travel speed S0 and true traffic volume Q0 in the teacher data K and the correction term C output from the trained model E1. In addition, in the later stage of re-learning, the learning processing unit 15 increases the frequency of performing the re-learning process using a data group (hereinafter also referred to as "data group N2") including the travel speed data S2, traffic volume data Q2 and the correction term C output from the trained model E1.

[0104] More specifically, for example, when performing re-learning processing, the learning processing unit 15 acquires teacher data K from the memory unit 16, i.e., congestion length data L1, travel speed data S1, traffic volume data Q1, true congestion length L0, true travel speed S0 and true traffic volume Q0.

[0105] When the learning processing unit 15 acquires the congestion length data L1, the travel speed data S1, the traffic volume data Q1, the true congestion length L0, the true travel speed S0, and the true traffic volume Q0 from the memory unit 16, it inputs the congestion length data L1, the travel speed data S1, and the traffic volume data Q1 to the trained model E1 in the memory unit 16. Then, the learning processing unit 15 acquires the travel speed data S2, the traffic volume data Q2, and the correction term C output from the trained model E1.

[0106] When the learning processing unit 15 acquires the travel speed data S2, the traffic volume data Q2, and the correction term C, it creates a data group N1 including the true travel speed S0, the true traffic volume Q0, and the correction term C acquired from the memory unit 16. The learning processing unit 15 also creates a data group N2 including the acquired travel speed data S2, the traffic volume data Q2, and the correction term C.

[0107] In the following description, the event using the data set N1 in the relearning process is also referred to as "event X0," and the event using the data set N2 in the relearning process is also referred to as "event X1."

[0108] For example, when the learning processing unit 15 creates data groups N1 and N2 in the re-learning process, it generates random numbers for the data groups N1 and N2, and selects a data group to be used in the re-learning process (hereinafter also referred to as a "data group for re-learning") depending on the value of the generated random number.

[0109] Specifically, for example, the learning processing unit 15 increases the value of the probability P1 that the event X1 occurs by a predetermined value each time the number of epochs increases. In this embodiment, for example, the learning processing unit 15 sets the value of the probability P1 to 0.1 when the number of epochs is zero, and sets the value of the probability P1 to 1 when the number of epochs reaches a predetermined number. Then, the learning processing unit 15 increases the value of the probability P1 from 0.1 by a predetermined value each time the number of epochs increases until the value of the probability P1 becomes 1.

[0110] Then, in the re-learning process, the learning processing unit 15 performs a process of generating a random number r having a value of "1" with a probability P1, and generating a random number r having a value of "0" with a probability P2 obtained by subtracting the probability P1 from 1.

[0111] Then, when a random number r with a value of "1" is generated, the learning processing unit 15 selects the data group N2 as the re-learning data group. When a random number r with a value of "0" is generated, the learning processing unit 15 selects the data group N1 as the re-learning data group. That is, until the value of the probability P1 becomes 1, when the value of the random number r is small, the learning processing unit 15 increases the frequency of using the data group N1 as the re-learning data group, and when the value of the random number r is large, the learning processing unit 15 increases the frequency of using the data group N2 as the re-learning data group. When the value of the probability P1 becomes 1, the generated random number r becomes "1", so the learning processing unit 15 uses the data group N2 as the re-learning data group.

[0112] For example, when a random number r with a value of "0" is generated, the learning processing unit 15 inputs the true travel speed S0, the true traffic volume Q0, and the correction term C contained in the data group N1 to the mathematical model E2, as shown in Figure 5, and obtains the congestion length data L2 output from the mathematical model E2.

[0113] Also, for example, when a random number r having a value of "1" is generated, the learning processing unit 15 inputs the travel speed data S2, the traffic volume data Q2, and the correction term C included in the data group N2 to the mathematical model E2, as shown in Figure 6, and obtains the congestion length data L2 output from the mathematical model E2.

[0114] When the learning processing unit 15 acquires the congestion length data L2, it assigns the congestion length data L2, travel speed data S2, traffic volume data Q2, true congestion length L0, true travel speed S0, and true traffic volume Q0 to a loss function F to acquire an output value of the loss function F. The learning processing unit 15 updates various parameters of the trained model E1 until the output value of the loss function F converges. Furthermore, when the learning processing unit 15 updates various parameters of the trained model E1, it adds one to the number of epochs and increases the probability P1 used to generate the random number r by a predetermined value.

[0115] When the output value of the loss function F has converged, the learning processing unit 15 stores a new trained model E1 with various parameters updated in the storage unit 16. The first acquisition unit 12 uses the new trained model E1 in the next prediction process for predicting the travel speed S and the traffic volume Q.

[0116] [Operation flow] FIG. 7 is a flowchart defining an operation procedure when the traffic information prediction device according to the embodiment of the present disclosure performs a prediction process.

[0117] Referring to FIG. 7, first, the traffic information prediction device 101 receives congestion length information from the congestion length sensor 51, traffic volume information from the traffic volume sensor 52, and vehicle passing information from the optical beacon 53, and stores them in the memory unit 16 (step S101).

[0118] The traffic information prediction device 101 stores new congestion length information received from the congestion length sensor 51, new traffic volume information received from the traffic volume sensor 52, and new vehicle passing information received from the optical beacon 53 in the memory unit 16 (step S101) until a certain period of time has elapsed since the traffic information prediction device 101 was started (NO in step S102).

[0119] Then, when a certain time has elapsed since the traffic information prediction device 101 was started (YES in step S102), the traffic information prediction device 101 creates time-series congestion length data L1 to be used in the prediction process based on multiple pieces of congestion length information stored in the memory unit 16 (step S103).

[0120] Next, the traffic information prediction device 101 creates time-series traffic volume data Q1 to be used in the prediction process, based on a plurality of pieces of traffic volume information stored in the storage unit 16 (step S104).

[0121] Next, the traffic information prediction device 101 creates time-series travel speed data S1 to be used in the prediction process based on the vehicle passing information stored in the storage unit 16 (step S105). Note that steps S103, S104, and S105 may be executed in a reverse order or in parallel.

[0122] Next, the traffic information prediction device 101 acquires the learned model E1, section information and periodicity information from the memory unit 16, and inputs the created congestion length data L1, traffic volume data Q1 and travel speed data S1, as well as the section information and periodicity information acquired from the memory unit 16 into the learned model E1 (step S106).

[0123] Next, the traffic information prediction device 101 acquires the travel speed data S2, the traffic volume data Q2, and the correction term C output from the trained model E1 (step S107).

[0124] Next, the traffic information prediction device 101 acquires the mathematical model E2 based on the hourglass model from the storage unit 16, and inputs the travel speed data S2, the traffic volume data Q2, and the correction term C to the mathematical model E2 (step S108).

[0125] Next, the traffic information prediction device 101 acquires congestion length data L2, which is the prediction result of the congestion length L output from the mathematical model E2 (step S109).

[0126] Next, the traffic information prediction device 101 associates the travel speed data S2 and the traffic volume data Q2 with the congestion length data L2 and displays them on the display device (step S110), and newly receives congestion length information, traffic volume information and optical beacon information and stores them in the memory unit 16 (step S101).

[0127] FIG. 8 is a flowchart that defines an operation procedure when the traffic information prediction device 101 according to the embodiment of the present disclosure performs the re-learning process.

[0128] 8, first, the traffic information prediction device 101 creates a data group N1 and a data group N2. For example, as described above, the traffic information prediction device 101 acquires the congestion length data L1, the travel speed data S1, the traffic volume data Q1, the true congestion length L0, the true travel speed S0, and the true traffic volume Q0 from the storage unit 16. Then, the traffic information prediction device 101 inputs the acquired congestion length data L1, the travel speed data S1, and the traffic volume data Q1 to the trained model E1 in the storage unit 16. When the traffic information prediction device 101 acquires the travel speed data S2, the traffic volume data Q2, and the correction term C output from the trained model E1, it creates a data group N1 including the true travel speed S0, the true traffic volume Q0, and the correction term C, and a data group N2 including the travel speed data S2, the traffic volume data Q2, and the correction term C (step S201).

[0129] Next, the traffic information prediction device 101 generates a random number r for the created data group N1 and data group N2 (step S202).

[0130] Next, the traffic information prediction device 101 checks whether the value of the generated random number r is 1 or not (step S203).

[0131] If the random number r is not 1 (NO in step S203), the traffic information prediction device 101 selects the data group N1 as a re-learning data group (step S204).

[0132] On the other hand, if the random number r is 1 (YES in step S203), the traffic information prediction device 101 selects the data group N2 as the re-learning data group (step S205).

[0133] Next, the traffic information prediction device 101 inputs the selected re-learning data group to the mathematical model E2 in the storage unit 16. For example, as described above, when the traffic information prediction device 101 selects the data group N1 as the re-learning data group, it inputs the true travel speed S0, the true traffic volume Q0, and the correction term C to the mathematical model E2. Also, for example, when the traffic information prediction device 101 selects the data group N2 as the re-learning data group, it inputs the travel speed data S2, the traffic volume data Q2, and the correction term C to the mathematical model E2 (step S206).

[0134] Next, the traffic information prediction device 101 acquires the congestion length data L2 output from the mathematical model E2 (step S207).

[0135] Next, the traffic information prediction device 101 performs a re-learning process on the learned model E1 by substituting the congestion length data L2, the travel speed data S2, the traffic volume data Q2, the true congestion length L0, the true travel speed S0, and the true traffic volume Q0 into the loss function F, and updates various parameters in the learned model E1 (step S208).

[0136] Next, the traffic information prediction device 101 adds one to the epoch number indicating the number of times the learned model E1 has been re-learned (step S209).

[0137] Next, the traffic information prediction device 101 increases, by a predetermined value, the value of the probability P1 used in the process of generating the random number r, that is, the value of the probability P1 of performing the re-learning process using the data group N2 (step S210).

[0138] Next, the traffic information prediction device 101 checks whether the output value of the loss function F has converged (step S211).

[0139] When the output value of the loss function F has converged (YES in step S211), the traffic information prediction device 101 stores a new trained model E1 in which various parameters have been updated in the storage unit 16 (step S212).

[0140] On the other hand, if the output value of the loss function F does not converge (NO in step S211), the traffic information prediction device 101 newly acquires a data group N1 and a data group N2 (step S201).

[0141] In the traffic information prediction device 101 according to the embodiment of the present disclosure, the trained model E1 is configured to output the travel speed data S2, the traffic volume data Q2, and the correction term C, but is not limited to this. The trained model E1 may be configured to output the travel speed data S2 and the traffic volume data Q2 without outputting the correction term C.

[0142] In addition, in the traffic information prediction device 101 according to the embodiment of the present disclosure, the output unit 14 is configured to output the travel speed data S2 and the traffic volume data Q2 acquired by the first acquisition unit 12 in association with the congestion length data L2 that is the output result of the mathematical model E2, but this is not limited to the above. The output unit 14 may be configured to output the congestion length data L2 without outputting the travel speed data S2 and the traffic volume data Q2.

[0143] In addition, in the traffic information prediction device 101 according to the embodiment of the present disclosure, the second acquisition unit 13 is configured to use the above formula (1) obtained by modifying the calculation formula in the hourglass model that outputs the travel speed as the mathematical model E2 in the prediction process, but this is not limited to this. The second acquisition unit 13 may be configured to use another formula other than the formula obtained by modifying the hourglass model as the mathematical model in the prediction process.

[0144] In the traffic information prediction device 101 according to the embodiment of the present disclosure, the learning processing unit 15 is configured to change the weighting coefficients λ1 and λ2 to values ​​different from the weighting coefficients λ3 and λ4 in the learning process according to the size of the congestion length in the teacher data K, but this is not limited to this. The learning processing unit 15 may be configured to use fixed values ​​of the weighting coefficients λ1 and λ2.

[0145] In addition, in the traffic information prediction system 501 according to the embodiment of the present disclosure, the traffic information prediction device 101 is configured to include the learning processing unit 15 that re-learns the trained model E1, but this is not limited to this. The traffic information prediction device 101 may be configured not to include the learning processing unit 15 and not to update the trained model E1.

[0146] In addition, in the traffic information prediction device 101 according to the embodiment of the present disclosure, the first acquisition unit 12 is configured to input section information and periodicity information to the trained model E1 in addition to the congestion length data L1, travel speed data S1, and traffic volume data Q1, but this is not limited to the above. The first acquisition unit 12 may be configured to input either section information or periodicity information to the trained model E1 in addition to the congestion length data L1, travel speed data S1, and traffic volume data Q1. The first acquisition unit 12 may be configured to input the congestion length data L1, travel speed data S1, and traffic volume data Q1 to the trained model E1, but not to input the section information and periodicity information to the trained model E1.

[0147] In the traffic information prediction system 501 according to the embodiment of the present disclosure, the traffic information prediction device 101 is configured to include the creation unit 11 that creates the congestion length data L1, the traffic volume data Q1, and the travel speed data S1, but this is not limited to this. A device other than the traffic information prediction device 101 may be configured to include the creation unit 11.

[0148] Furthermore, some or all of the functions of the traffic information prediction device 101 according to the embodiment of the present disclosure may be provided by cloud computing. That is, the traffic information prediction device 101 according to the embodiment of the present disclosure may be a cloud server configured by a plurality of servers.

[0149] The above-described embodiments should be considered as illustrative and not restrictive in all respects. The scope of the present invention is defined by the claims, not by the above description, and is intended to include all modifications within the meaning and scope of the claims.

[0150] Each process (each function) in the above-mentioned embodiment is realized by a processing circuit including one or more processors. The processing circuit may be configured as an integrated circuit or the like in which one or more memories, various analog circuits, and various digital circuits are combined in addition to the one or more processors. The one or more memories store programs (instructions) that cause the one or more processors to execute each of the above processes. The one or more processors may execute each of the above processes according to the programs read from the one or more memories, or may execute each of the above processes according to a logic circuit designed in advance to execute each of the above processes. The processor may be various processors suitable for computer control, such as a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), a DSP (Digital Signal Processor), an FPGA (Field Programmable Gate Array), and an ASIC (Application Specific Integrated Circuit). The physically separated processors may execute each of the above processes in cooperation with each other. For example, the processors mounted on each of a plurality of physically separated computers may cooperate with each other via a network such as a LAN (Local Area Network), a WAN (Wide Area Network), the Internet, etc. The program may be installed in the memory from an external server device or the like via the network, or may be distributed in a state stored in a recording medium such as a CD-ROM (Compact Disc Read Only Memory), a DVD-ROM (Digital Versatile Disc Read Only Memory), or a semiconductor memory, and installed in the memory from the recording medium.

[0151] The above description includes the following additional features. [Appendix 1] a first acquisition unit that inputs a first congestion length, a first travel speed, and a first traffic volume at a first time into a trained model, and acquires a second travel speed, which is a predicted value of the travel speed at a second time after the first time, and a second traffic volume, which is a predicted value of the traffic volume at the second time, output from the trained model; a second acquisition unit that performs a prediction process of inputting the second travel speed and the second traffic volume acquired by the first acquisition unit into a mathematical model and acquiring a second congestion length, which is a predicted value of a congestion length at the second time output from the mathematical model; an output unit that outputs the second congestion length acquired by the second acquisition unit, The first acquisition unit further inputs section information related to a predetermined section of a road, the section information affecting traffic in the predetermined section or a section different from the predetermined section, into the trained model; The predetermined section includes an intersection, A traffic information prediction device, wherein the section information indicates at least one of the length of the specified section, the number of incoming roads to the intersection, the number of outgoing roads from the intersection, the number of lanes in the specified section, and the number of intersections.

[0152] [Appendix 2] A processing circuit is provided, The processing circuitry includes: inputting a first congestion length, a first travel speed, and a first traffic volume at a first time into a trained model, and acquiring a second travel speed, which is a predicted value of the travel speed at a second time after the first time, and a second traffic volume, which is a predicted value of the traffic volume at the second time, which are output from the trained model; inputting the acquired second travel speed and the acquired second traffic volume into a mathematical model, and acquiring a second congestion length which is a predicted value of the congestion length at the second time outputted from the mathematical model; A traffic information prediction device that outputs the acquired second congestion length. [Explanation of symbols]

[0153] 1 vehicle 11 Creation Department 12 First acquisition part 13 Second acquisition part 14 Output section 15 Learning processing section 16 Memory section 51 Traffic jam length detector 52 Traffic Detector 53 Light Beacon 54 Optical beacon vehicle unit 101 Traffic information prediction device 501 Traffic Information Prediction System C correction term E1 Pre-trained model E2 Mathematical Model

Claims

1. a first acquisition unit that inputs a first congestion length, a first travel speed, and a first traffic volume at a first time into a trained model, and acquires a second travel speed, which is a predicted value of the travel speed at a second time after the first time, and a second traffic volume, which is a predicted value of the traffic volume at the second time, output from the trained model; a second acquisition unit that performs a prediction process of inputting the second travel speed and the second traffic volume acquired by the first acquisition unit into a mathematical model and acquiring a second congestion length, which is a predicted value of a congestion length at the second time, as an output value output from the mathematical model; and an output unit that outputs the second congestion length acquired by the second acquisition unit.

2. The trained model further outputs correction information for correcting an error in the prediction process, The second acquisition unit further acquires the correction information output from the trained model, 2. The traffic information prediction device according to claim 1, wherein the second acquisition unit inputs the second travel speed and the second traffic volume acquired by the first acquisition unit, and the acquired correction information into the mathematical model, and performs the prediction process to acquire the second congestion length output from the mathematical model.

3. 3. The traffic information prediction device according to claim 1, wherein the output unit further outputs the second travel speed and the second traffic volume acquired by the first acquisition unit in association with the second congestion length acquired by the second acquisition unit.

4. 3. The traffic information prediction device according to claim 1, wherein the second acquisition unit uses, as the mathematical model in the prediction process, a second calculation formula for congestion length, which is a modification of a first calculation formula in an hourglass model that outputs a travel speed.

5. The traffic information prediction device further includes: a learning processing unit that performs a learning process to train the trained model using teacher data in which travel speed, traffic volume, and congestion length at a predetermined time are associated with each other, and a loss function related to a first difference between the second congestion length and a true congestion length at the second time, a second difference between the second travel speed and the true travel speed at the second time, and a third difference between the second traffic volume and the true traffic volume at the second time, the loss function includes a first weighting factor, a second weighting factor, and a third weighting factor multiplied by the first difference, the second difference, and the third difference, respectively; 3. The traffic information prediction device according to claim 1, wherein the learning processing unit changes the first weighting coefficient to a value different from the second weighting coefficient and the third weighting coefficient depending on the size of the congestion length in the teacher data during the learning process.

6. 3. The traffic information prediction device according to claim 1, wherein the first acquisition unit further inputs to the learned model at least one of first information regarding a specified section of a road, the first information affecting traffic in the specified section or a section different from the specified section, and second information regarding periodicity of the traffic in the specified section.

7. A traffic information prediction method in a traffic information prediction device, comprising: inputting a first congestion length, a first travel speed, and a first traffic volume at a first time into a trained model, and acquiring a second travel speed, which is a predicted value of the travel speed at a second time after the first time, and a second traffic volume, which is a predicted value of the traffic volume at the second time, outputted from the trained model; a step of inputting the acquired second travel speed and the acquired second traffic volume into a mathematical model, and performing a prediction process of acquiring a second congestion length, which is a predicted value of a congestion length at the second time output from the mathematical model; and outputting the acquired second congestion length.

8. A traffic information prediction program for use in a traffic information prediction device, Computer, a first acquisition unit that inputs a first congestion length, a first travel speed, and a first traffic volume at a first time into a trained model, and acquires a second travel speed, which is a predicted value of the travel speed at a second time after the first time, and a second traffic volume, which is a predicted value of the traffic volume at the second time, output from the trained model; a second acquisition unit that performs a prediction process of inputting the second travel speed and the second traffic volume acquired by the first acquisition unit into a mathematical model and acquiring a second congestion length, which is a predicted value of a congestion length at the second time output from the mathematical model; an output unit that outputs the second congestion length acquired by the second acquisition unit; A traffic information prediction program to function as a

Images