Limi TFP Citywide Traffic Flow Prediction With Limited Road Status Information

Citywide traffic flow prediction is a basic but challenging task in the field of Intelligent Transportation Systems (ITS). Most existing traffic flow prediction methods require historical traffic data of all the target roads in both training and application stages. However, it is not easy, and usually costly, to collect real-time traffic data across the entire road network. Therefore, many existing methods have to narrow down the scope of their prediction on the parts of road networks. In this paper, we propose a deep-learning model, named Limi-TFP, which has the ability to identify a limited number of monitored roads, and achieves citywide traffic flow prediction by using the historical traffic data of these selected roads. Specifically, an embedding module is proposed to capture the spatial context (road topology structure) and the attributes (e.g., road type, length, lanes, etc.) of each road by embedding all road segments into vectors. A road ranking method is then developed for selecting a limited number of roads to be monitored. A group of multi-head attention mechanisms is exploited for capturing the dynamic correlations with each monitored road. Meanwhile, the fusion with external factors, including the point of interests (POIs) and weather data, is also taken into account to further improve the prediction accuracy. Extensive experiments on the real-world datasets demonstrate that the proposed method achieves the superior performance and is robust with the presence of noise compared to the state-of-the-art baselines, with only 5% roads being monitored.