• PDF version [PDF: 662KB]

East Nippon Expressway Co., Ltd. (hereinafter NEXCO EAST) and Grid Co., Ltd. (hereinafter Grid) utilize AI to develop technology that enables traffic jam forecasters to perform traffic jam prediction several months ahead. Was successful.

Grid is a technology venture company with one of the leading AI technologies in Japan, and the AI development platform "ReNom" that can respond to various issues developed and provided by Grid. ^* And, using NEXCO EAST 's congestion prediction technology, we developed a prediction model equivalent to a traffic forecaster. As a result of comparing this prediction model with the traffic jam results during the traffic congestion period, we were able to confirm a certain degree of accuracy and set the stage for practical application.

1 AI traffic congestion prediction

Traffic congestion forecasts such as traffic congestion periods and congestion calendars that are more than a few months ahead can be done by a congestion forecaster who was previously in charge of NEXCO EAST 's congestion forecasting work, superimposing past congestion records and arranging days of the week, changes in road conditions, and surrounding areas. It was decided and predicted after considering the situation of the event.

The congestion prediction model using AI learns past factor data that is likely to greatly affect the occurrence of congestion and predicts whether or not congestion will occur at a future date and time.

This development is targeted at the Kan-Kan-Etsu Expressway, and a large amount of data from the following [1] to [2] for about 14 years from 2004 to 2018 was used for learning. When learning, teacher data is created by combining NEXCO EAST 's traffic congestion prediction know-how with grid model engineering technology.
[1] Speed and traffic volume data every 5 minutes obtained from a device called a traffic counter
[2] Calendar pattern for each year (day of the week arrangement, holiday arrangement, etc.)

• An AI development platform that allows you to freely build advanced algorithms such as deep learning according to the task, and for simple tasks, even if you are not an expert you can develop models with a GUI interface without writing a program .

[Reference] Conventional traffic forecast by traffic forecaster

The traffic jam forecasts performed by traffic jam forecasters are roughly classified into the following tasks [1] to [4].

• [1] Overlapping past traffic jam results (for 3 years)
• [2] Scrutiny of past traffic jam results
• [3] Addition of the latest traffic trends
• [4] Correction work (combining adjacent traffic jams, considering the effect on connecting routes)

2 Accuracy of traffic congestion prediction by AI

When the traffic congestion forecast (GW, Obon) on this year's Kan-Etsu Expressway was compared with actual traffic congestion forecasts by AI and traffic forecasters, the missed rate of the forecast ^* , Missed rate ^* Both were about 20%, and it was confirmed that the prediction was possible with almost the same accuracy as the prediction by the traffic jam forecaster.

• Missed rate: “Number of missed times (predicted that no traffic jam will occur, but the actual number of traffic jams that occurred”) / “Total number of traffic jams”
  Missing rate: “Number of missed events (the number of predicted traffic jams that predicted that a traffic jam will occur, but did not actually occur)” / “the total number of traffic jam predictions”

In addition, when comparing the traffic jam forecast for the year-end and New Year holidays of the Kan-Etsu Expressway this year, the traffic forecast by AI and the traffic forecast by the traffic forecaster are compared, and about 80% show the same tendency.

<< Example of forecast [1] >> Kan-Kan-Etsu Expressway In-bound Line January 2 (Wednesday) Comparison of AI forecast and forecaster forecast

<< Example of forecast [2] >> Kan-Kan-Etsu Expressway In-bound Line January 3 (Thursday) Comparison of AI forecast and forecaster forecast

3 Future development of traffic congestion prediction using AI

• The technology developed this time can be applied to other than the targeted Kan-Etsu Expressway. In the future, we will study other routes such as the Tohoku Expressway and aim to Large target routes.
• At this time, it is difficult for traffic congestion forecasters to make predictions based on changes in road conditions and toll systems, so in traffic congestion forecasting operations, we will use the traffic forecasters to assist in making predictions such as oversights.
• If the accuracy of traffic congestion prediction by AI will be further improved in the future and cooperation with the conventional prediction system will be made, the traffic congestion forecaster will perform a lot of work time such as [1] overlay work and [4] correction work. It is expected to be shortened and traffic congestion forecasting work will be cut in half.
• In the future, we will examine the possibility of new learning data such as weather information and accident occurrence status, and aim to further improve accuracy.