Technologies for measuring urban and regional activities and expressing them in quantitative data have become widely available, and AI technology for predicting the future based on past spatiotemporal data is attracting attention in society. AI-based prediction is expected to find applications across a wide range of areas and to be used by everyone from individuals to businesses to local governments in order to help them make informed decisions. However, potential errors in such predictions may cause confusion in society. The development of a spatiotemporal AI that can provide reliable predictions will thus be an essential technology for supporting reliable decision-making in the society of the future.

A research team consisting of Koh Takeuchi, Senior Lecturer, Graduate School of Kyoto University, and others has developed QTNN (Queueing-Theory-based Neural Network), a spatiotemporal AI^*1 that accurately predicts when and where traffic congestion will occur. To verify QTNN, experiments were conducted to predict queue length one hour ahead at 1,098 road locations in Tokyo. As a result, QTNN not only successfully predicted whether congestion would occur, but also provided highly accurate predictions of queue length with an average error of 40 m or less. As shown in Fig. 1, this result represents a remarkable 12.6% reduction in prediction errors compared to the state-of-the-art deep learning^*2 methods currently available (DCRNN, AGCRN, GWNT, and Mega CRN).

*1 Spatiotemporal AI
This term refers to AI developed to learn and analyze the complex spatiotemporal relationships of a phenomenon (such as the correlation between temporal changes in traffic conditions and spatial changes due to road connections) based on information on when and where data was measured. Spatiotemporal AI can be used to analyze a wide variety of data measured in cities and regions, and is thus expected to find broad applications beyond traffic congestion in fields related to travel and industry.

*2 Deep learning
A type of multi-layer neural network that learns patterns from vast amounts of data. Although it is a technology driving the recent AI boom, it has also been criticized for its black-box nature, as its internal computations are not transparent. In addition, even with deep learning, the accuracy of traffic data analysis is still considered to be insufficient, unlike text, image, and voice analysis. Deep learning is also used as a core technology in generative AI and large language models. Hopes are high for the emergence of a new type of generative AI that leverages temporal and spatial information.

In addition to predicting congestion by learning the relationship between traffic flow/condition changes and road networks, QTNN also adjusts the mathematical model used to determine the traffic conditions (traffic flow model^*3) according to deep learning prediction results prior to outputting its final predictions.

This mechanism makes it possible for users to learn the reasons behind the predictions, such as “There will be a rapid increase in vehicle influx around 6:00, resulting in a significant increase in queue length,” and “The queue length is reaching its peak, causing a decrease in traffic flow and average speed; due to the traffic flow on surrounding roads, this queue will continue until around 10:00.” Internal calculations performed in deep learning are usually done within a “black box,” meaning that the reasons behind the predictions cannot be ascertained. QTNN alleviates this drawback and enables people to interpret the reasons and determine what to do next. This characteristic leads to reliable predictions.

*3 Traffic flow model
This is a mathematical model used to describe traffic conditions. Typically, it describes the relationship between three variables: traffic flow, average speed, and density.

What lead to the development of QTNN is that traffic congestion has become a serious social and economic problem for Japan’s road networks. In addition to making daily lives more stressful, traffic congestion slows the speed of transport trucks and other vehicles as well as significantly affects economic activities. According to the Ministry of Land, Infrastructure, Transport and Tourism, traffic congestion results in a loss of approximately 10 trillion yen annually nationwide. Traffic congestion also contributes to higher emissions of CO₂ and other greenhouse gases, which are contained in vehicle exhaust, thus adversely affecting climate issues.

Countries around the world share in such traffic congestion problems, and congestion prediction technologies have long been studied to prevent or reduce queue formation. In recent years, AI-based prediction technologies have attracted attention. The ability to accurately predict upcoming traffic congestion will lead to smoother traffic flows and less congestion because it will enable proactive route guidance and signal control.

Various efforts are underway to enhance traffic control systems by utilizing AI technology, although there is one thing that even AI seems to struggle with: congestion prediction. The reason is that accurately forecasting the presence and length of traffic congestion has been a difficult pursuit because queue formation varies widely by time of day, location, and length, and traffic conditions change rapidly after queues form. Around the world, some studies have been conducted on AI-based congestion prediction with a primary focus on average speed and traffic flow, but none of these address queue length prediction.

QTNN predicts congestion by the two-stage approach shown in Fig. 2. QTNN first predicts the average speed and traffic flow at each road segment through deep learning (with STGNN: Spatio-Temporal Graph Neural Network) of past data on the congestion status (queue length, average speed, and traffic flow) of numerous roads. Based on these predictions, QTNN then adjusts (in the QT-layer) the mathematical model (traffic flow model) derived from traffic engineering knowledge. Future queue length is predicted using this two-stage approach. The research team has realized congestion predictions that are aligned with traffic engineering knowledge while leveraging state-of-the-art deep learning techniques.

The data used in this demonstration experiment includes one year of data on average speed, traffic flow, and queue length that was measured every 5 minutes at 1,098 locations on general roads within Tokyo’s 23 wards (where the average length of a road segment was 882 m, with a median of 750 m, and queue length was determined by means of measuring the length of a line of vehicles repeatedly stopping and starting, moving at speeds below a certain threshold).

An example of the road network within Tokyo is shown in Fig. 3 (a). Observation of the Road 1 and Road 2 segments in terms of travel speed (b), traffic flow (c), and queue length (d) indicates that shortly after congestion occurred in Road 1, congestion also occurred in Road 2. As shown by Fig. 4, average speed, traffic flow, and queue length can also be obtained numerically.

Using this data, as described above, an experiment was conducted to predict the queue length one hour ahead at 1,098 road locations within Tokyo over a span of two months. The results demonstrated highly accurate prediction in both cases of congestion and of non-congestion, achieving an error of 40 m or less and a remarkable 12.6% reduction in prediction errors compared to predictions using other deep learning technologies (Fig. 1).

Fig. 5 shows plots of QTNN predictions (red lines) and observed congestion (black dotted lines). This indicates that predictions of travel speed, traffic flow, and queue length were generally close to the observed values in three different road segments (a, b, and c).

QTNN is being considered for utilization in the Metropolitan Police Department’s project that aims to enhance traffic control systems by utilizing AI and big data. Going forward, the research team plans to conduct evaluation tests on certain roads to assess QTNN’s reliability as a step towards full-scale implementation in real-world scenarios. The team also aims to establish a spatiotemporal AI technology that will provide the foundation for urban infrastructure by flexibly utilizing information related to signal control, road construction, and accident occurrences to predict queue length.

Fig. 6 shows a conceptual drawing of a future traffic control system and road infrastructure utilizing QTNN. In this future world, QTNN predicts the queue length based on the measured traffic data, enabling the traffic control system to adjust traffic signal timings at intersections and thus prevent queue formation. The predicted congestion information can also be utilized by onboard navigation systems and map applications, which guide drivers to less congested routes so as to prevent or reduce traffic congestion.