Scalable transit delay prediction at city scale: A systematic approach with multi-resolution feature engineering and deep learning

Urban bus transit agencies need reliable, network-wide delay predictions to provide accurate arrival information to passengers and support real-time operational control. Accurate predictions help passengers plan their trips, reduce waiting time, and allow operations staff to adjust headways, dispatch additional vehicles, and manage disruptions. Although real-time feeds such as GTFS-RT are now widely available, most existing delay prediction systems handle only a few routes, rely on hand-crafted features, and offer little guidance on designing a scalable, reusable architecture. We present a city-scale prediction pipeline that combines multi-resolution feature engineering, dimensionality reduction, and deep learning. The framework systematically generates spatiotemporal features by exploring aggregation combinations over spatial regions (using hexagonal hierarchical indexing), routes, segments, and temporal patterns, then compresses them using Adaptive PCA while preserving 95 % of the variance. To avoid the “giant cluster” problem that occurs when dense urban areas fall into a single spatial region, we introduce a hybrid clustering method that combines geographic and network topology information to yield balanced route clusters and enable efficient distributed training. We compare five model architectures on six months of bus operations from the Société de transport de Montréal (STM) network in Montréal. A global LSTM with cluster-aware features achieves the best trade-off between accuracy and efficiency ( R 2 = 0 . 7121 at the elementary level), outperforming transformer models by 18 % to 43 % while using 275 × fewer parameters. LSTM’s compact architecture (31,000 parameters) effectively captures short-term temporal dependencies in the compressed feature space, making it more suitable than transformer models, which are overparameterized for this task. We also report multi-level evaluation at the elementary segment, segment, and trip level using walk-forward validation and latency analysis, showing that the proposed pipeline is suitable for real-time, city-scale deployment and can be reused for other networks with limited adaptation. • Developed a scalable, systematic framework for multi-resolution feature engineering, leveraging Uber’s H3 spatial indexing system to generate spatial–temporal features across diverse aggregation levels. • Introduced a hybrid clustering method combining H3 indexing and topological structure to mitigate the issue of disproportionately large clusters, a challenge we refer to as the “giant cluster” problem, which arises in dense urban mobility data. • Demonstrated that a global LSTM architecture surpasses transformer-based models in trip-level delay prediction, achieving higher accuracy with significantly fewer parameters. • Built a complete end-to-end prediction pipeline incorporating hierarchical feature aggregation, walk-forward validation, and thorough computational efficiency analysis. • Validated the system’s robustness and generalizability through extensive evaluations across model types, spatial resolutions, and challenging conditions such as extreme weather and special events.