Towards Forecasting Bus Arrival Thorough A Model Based On GNN+LSTM Using GTFS and Real-time Data

The city’s Public Transportation Network (PTN) organizes daily mobility services for millions of people. Some cities worldwide have used urban computing toolkits to handle acquisition, integration, and data analysis, which translates to improving mobility services, such as mitigating and notifying delays. Regarding the PTN, many urban computing approaches rely on two specifications: General Transit Feed Specification (GTFS), which represents the static schedule information, and General Transit Feed Specification Real-Time (GTFS-RT), which introduces real-time updates from trips, services, and vehicle positions. Despite the qualitative leap with the GTFS-RT specification, we perceive that it has not been as well-adopted as the GTFS, probably because of the difficulty of matching identifiers between the static and real-time data. In this context, our paper presents a new approach that combines Graph Neural Networks (GNNs) and Long Short-Term Memory (LSTM) networks to improve the prediction of bus arrival times at official stops using available real-time data. Our model leverages static schedule information from GTFS and real-time updates (from GTFS-RT or not) to create a comprehensive dataset. The GNN component captures the static characteristics of the PTN, which are then combined with temporal data and fed into an LSTM network to forecast bus arrival times. Our integrated GNN+LSTM model offers superior performance by utilizing both spatial and temporal data, overcoming static baselines and the performance of the standalone LSTM model.