A TRANSFORMER-BASED APPROACH TO REAL-TIME URBAN MOBILITY KPI PREDICTION

Accurate short-term prediction of urban mobility KPIs, such as bus arrival delays, is essential for passenger information and transit operations. This thesis proposes a Transformer-based approach that fuses static GTFS schedules with archived GTFS-RT feeds into stop-level event sequences, representing each trip as an ordered sequence of stops enriched with schedule, geometric, and observed features. A Transformer encoder is trained to predict per-stop delay along the trip, leveraging self-attention to capture delay propagation across consecutive stops. The model is evaluated on real-world data from AMTAB service (local public transport of Bari, IT) and compared against schedule-adherence and recurrent neural baselines. The work contributes a reusable GTFS-to-tensor pipeline and empirical evidence that attention-based models are well suited to sequence-structured transit forecasting.