A Spatio-Temporal Neural Network for Bus Travel Time Forecasting

Bus travel time forecasting is a critical challenge in analysing and optimizing Public Transportation Networks. In this thesis, we propose a Spatio-Temporal Neural Network that captures both spatial and temporal dependencies by combining Graph Neural Networks (GNNs) with Recurrent Neural Networks (RNNs). The model is applied to a large and heterogeneous set of real-world bus routes from Belo Horizonte, Brazil, obtained from the integration of scheduled and real-time GTFS data. The proposed model significantly outperforms standard baselines, especially in detecting irregular and less frequent temporal patterns.