AI China

Shanghai Tests Graph Recurrent Neural Networks for Traffic Prediction

A new study from Shanghai's Transportation Information Center (STIC) and Shanghai Jiao Tong University uses Graph Recurrent Neural Networks (GRNN) for high-accuracy traffic prediction and city traffic control.

The largest city in China by population, Shanghai is also a global financial center and a major transportation hub. An efficient traffic prediction system can be a practical tool to help ease congestion and avoid traffic accidents in such a bustling metropolis. A new study from Shanghai’s Transportation Information Center (STIC) and Shanghai Jiao Tong University uses Graph Recurrent Neural Networks (GRNN) for high-accuracy traffic prediction and city traffic control.

Traffic prediction is an essential component of an Intelligent Transportation System (ITS). The challenge is obtaining high accuracy while keeping computational complexity low due to the spatiotemporal characteristics of traffic flow, especially in a supercity such as Shanghai.

The outstanding effectiveness of Graph Recurrent Neural Networks (GRNN) elevated the STIC study above others and above most current approaches, which treat traffic prediction as a time series problem by utilizing common time series analysis and statistical learning. However, traffic conditions of one road segment are actually highly correlated with other segments. Previous studies underestimated the influence of global information on the traffic network. When such studies used extra spatiotemporal data to examine how global information affected the traffic network, the extra data turned out to be costing additional computation.

Shanghai Image1.png

The STIC and Jiao Tong researchers linked two solutions to tackle the challenge. They first initiated a topological framework called Linkage Network to perform modeling of road networks and illustrate the propagation patterns of traffic flow. They then introduced the Graph Recurrent Neural Network as an online predictor to mine and learn the propagation patterns in the graph globally and synchronously. Combining the two components enabled simultaneous traffic flow prediction from information collected from the whole graph. Researchers found this could reduce computational complexity while maintaining high accuracy.

The STIC researchers designed the Linkage Network to show propagation patterns. When a vehicle chooses a specific linkage, researchers will define the proportions as a “propagation pattern.” This is a critical contribution to traffic variation. Taking an everyday traffic jam as an example, when a huge number of vehicles are on a road, the congestion they cause will eventually move to downstream segments. Propagation patterns became key to understanding traffic variations for example during Shanghai’s busy rush hours.

Researchers trained the online GRNN based on Graph Neural Networks (GNN) to interpret propagation patterns and predict traffic conditions. The GRNN was able to simultaneously predict traffic flow for all road segments based on the whole graph provided by the Linkage Network. The mechanism of the propagation module in GNNs elaborates from the static relationships among vertexes on an expanded time axis. The movement of vehicles via various road segments determined the propagation patterns are time-variant. The GRNN in the STIC study is a sequence-to-sequence model which can cope with streaming data.

Shanghai Image2.png

Researchers used a 310GB raw taxi trajectory dataset which contains 65,836 Shanghai road segments. The dataset also includes important features such as geopositioning, upload timestamp, speed, and orientation of the vehicle. However, most of the segments showed very sparse samples. Researchers believe that the GRNN will be able to achieve higher performance as additional data becomes available in the future.

Shanghai is among a growing number of Chinese cities to adopt Smart City strategies. Last month, internet giant Baidu signed a strategic cooperation initiative to develop the city’s Baoshan District into a smart city model with the help of AI, big data, IoT, blockchain, autonomous vehicles and so on. Three months ago, Waymo quietly opened a new Shanghai office.

The Shanghai municipal government also offers residents a number of digital tools designed to make their city experience more streamlined and intelligent. The cloud-based Citizen Cloud platform for example enables seven million users to access over 200 government services online. Alongside ongoing academic research such as the new traffic study, Shanghai is determined to ramp up its urban modernization efforts with the help of AI.

The paper Efficient Metropolitan Traffic Prediction Based on Graph Recurrent Neural Network is on arXiv.


Journalist: Fangyu Cai | Editor: Michael Sarazen

 

0 comments on “Shanghai Tests Graph Recurrent Neural Networks for Traffic Prediction

Leave a Reply

Your email address will not be published.

%d bloggers like this: