As a preferential treatment at signalized intersections, Transit Signal Priority (TSP) remains a key technology for enhancing transit performance. Recently, TSP systems based on General Transit Feed Specification (GTFS) Realtime have gained traction in the market, mainly because of their low implementation and maintenance costs. However, leveraging GTFS Realtime messages for TSP presents significant challenges, particularly because of two types of message delays: (1) high latency; and (2) long update intervals. Building on previous work that introduced regression models to compensate for message latency, three new machine learning models are proposed to more accurately predict future vehicle locations while mitigating these delays. To overcome the limitations of earlier regression approaches, a long short-term memory architecture for single-step prediction was developed, a long short-term memory architecture for multistep prediction was developed, and a Transformer-based architecture for multistep time series prediction was developed, which can address interval updating issues. The experimental results show that all three proposed models significantly outperform both previous regression models and five baseline statistical methods. These advancements improve the reliability and accuracy of GTFS-based Automatic Vehicle Location, reinforcing its role as a dependable data source for cloud-based TSP systems.
Zhou et al. (Fri,) studied this question.