Scalable Transit Delay Prediction at City Scale: A Systematic Approach with Multi-Resolution Feature Engineering and Deep Learning

TL;DR

The paper tackles scalable, network-wide bus delay prediction for city-scale transit systems by building a reproducible pipeline that combines exhaustive multi-resolution feature engineering, topology-aware clustering, and a comparison of five deep-learning architectures. It introduces a hybrid H3+topology clustering approach to mitigate the giant-cluster problem and demonstrates that a global LSTM with cluster-aware features achieves the best trade-off between accuracy and efficiency, outpacing transformers by up to 52% while using far fewer parameters. A 6-month STM Montréal dataset is processed into 1,683 features, reduced to 83 components with Adaptive PCA, and evaluated through walk-forward temporal cross-validation across elementary, segment, and trip levels, revealing systematic error cancellation in aggregation. The framework is designed for real-time deployment and reuse across networks with limited adaptation, offering a practical path toward reliable, city-scale delay predictions for passenger information and operations control.

Abstract

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 extra vehicles, and manage disruptions. Although real-time feeds such as GTFS-Realtime (GTFS-RT) are now widely available, most existing delay prediction systems handle only a few routes, depend on hand-crafted features, and offer little guidance on how to design a scalable, reusable architecture. We present a city-scale prediction pipeline that combines multi-resolution feature engineering, dimensionality reduction, and deep learning. The framework generates 1,683 spatiotemporal features by exploring 23 aggregation combinations over H3 cells, routes, segments, and temporal patterns, and compresses them into 83 components 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 H3 region, we introduce a hybrid H3+topology clustering method that yields 12 balanced route clusters (coefficient of variation 0.608) and enables 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, outperforming transformer models by 18 to 52% while using 275 times fewer parameters. We also report multi-level evaluation at the elementary segment, segment, and trip level with 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.

Figures (9)

• Figure 1: Overview of the transit delay prediction architecture, showing the five main pipelines and their interactions with external data sources and data storage.
• Figure 2: Data collection pipeline, from raw GTFS/GTFS-RT and weather feeds to validated trip-level records.
• Figure 3: Feature engineering pipeline, from trip-level records to a spatiotemporal feature matrix for modelling.
• Figure 4: Dimensionality reduction pipeline, from high-dimensional engineered features to a compact set of components used by all models.
• Figure 5: Predictive modelling pipeline, showing global model training on reduced features with cluster-aware inputs.