Schedule deviations in public transit networks—discrepancies between planned and actual vehicle arrival times—are a significant source of user dissatisfaction and operational inefficiency. The growing availability of real-time operational data through the GTFS-RT standard offers new opportunities to analyze and predict disruptions. Yet, a significant gap remains between data availability and practical use to improve service quality.
This thesis proposes an integrated, two-component approach to address this gap. The first component develops a diagnostic platform that integrates automatic ingestion of large-scale GTFS-RT streams, systematic classification of disruptions into their recurrent or one-off categories, use of H3 hierarchical spatial indexing for multi-resolution aggregation, and high-performance interactive visualization with KeplerGL. Applied to Montréal’s STM network, this open-source platform reveals significant spatio-temporal patterns and supports the identification of critical locations requiring intervention.
The second component develops a systematic approach for predicting schedule deviations at the scale of a city-wide transit network. The proposed method combines (1) a structured framework for multi-resolution feature engineering based on H3, including hierarchical spatial caractéristiques (features), cyclic temporal encodings, and learned embeddings; (2) an LSTM architecture optimized to capture complex temporal dependencies while satisfying scalability constraints through an appropriate clustering strategy; and (3) rigorous empirical evaluation using a spatio-temporal validation protocol on independent time periods. The results show predictive performance superior to traditional baselines, with a mean absolute error below two minutes for short-term prediction horizons.
The distinctive contribution of this thesis lies in the coherent integration of diagnosis and prediction within a unified methodological framework, in which insights from exploratory analysis directly inform the design of the predictive system. This synergy, combined with an emphasis on reproducibility, operational scalability, and an open-source implementation, aims to narrow the gap between academic research and practical deployment in intelligent transportation systems.
Boudabbous, E. (Author),
Montecinos (Supervisor) &
Sboui (Co-supervisor),
11 Feb 2026Student thesis: Master's thesis › Master in Engineering: Automated Manufacturing Engineering