Transit Signal Priority Strategy With Heterogeneous Graph-Based Deep Reinforcement Learning for Autonomous Public Transit Vehicles

Rapid advancements in vehicle communication and autonomous driving technologies have led to the emergence of Autonomous Public Transit Vehicles (APTVs), playing a pivotal role in enhancing the efficiency of on-demand flexible-route transit services. One promising approach to further optimize these services and alleviate urban traffic congestion is the development of smarter Transit Signal Priority (TSP) strategies. This paper proposes a decentralized intelligent traffic signal control algorithm based on Deep Reinforcement Learning (DRL), tailored for the TSP strategy supporting flexible-route transit services. Our algorithm can accommodate various road network structures and APTV penetration rates, ensuring extensive applicability. Specifically, it employs a heterogeneous graph model to capture diverse information, including network topologies and dynamic characteristics of APTVs. Through extensive testing in multiple scenarios across varied road networks and traffic conditions, our algorithm has consistently outperformed both traditional traffic control methods and state-of-the-art DRL-based methods. Furthermore, it demonstrates effective zero-shot transferability, adapting to real-world scenarios without additional training.