Hybrid Spatio-temporal GraphWaveNet for Real-Time Traffic Forecasting on Dynamic Graphs

Kota, Karthik (2025) Hybrid Spatio-temporal GraphWaveNet for Real-Time Traffic Forecasting on Dynamic Graphs. Masters thesis, Dublin, National College of Ireland.

Preview	PDF (Master of Science) Download (2MB) | Preview
Preview	PDF (Configuration Manual) Download (543kB) | Preview

Abstract

Real-time traffic sensor networks implicitly require well-performing models that capture space-time invariant structures, such as topology of the road yet consider time-varying aspects, e.g. time-of-day congestion. Traditional spatiotemporal graph neural networks (ST-GNNs) are often limited to fixed graphs or retraining models entirely when the distribution of features changes, which are not scalable or adaptive. The work suggests a hybrid modeling architecture that comprises Static Spatial Embeddings (SE) together with Feature-Based (FB) dynamic graph construction as a GraphWaveNet (GWN) backbone. We trained 16 spatiotemporal models, and the 7 best of them were tested on the 2 tasks of 3-step one-shot (H3) and 24-step autoregressive (H24-AR). Another 24-step one-shot task with time-of-day attributes (H24-TOD) employed three models (a static no-change baseline and two SE - FB hybrids with cosine and RBF attention). The analysis was done on a 60-day, 554-node traffic dataset using RMSE, MAE, MAPE, latency, sensitivity to sensor dropout, and scalability to 500 nodes. Hybrid models performed better than the static and adaptive baselines Most notably, the RBF variant performed the best on RMSE whereas the cosine variant performed best in terms of MAPE and robustness to dropouts. Both models will be scaled linearly and generalized without retraining. whereas, complete graph construction was expensive, and peak-hour errors were an indication that it will perform better. Furthermore, this involves GPU-accelerated inference, dynamic graph updates and optimisation strategies at peak times.

Item Type:	Thesis (Masters)
Supervisors:	Name Email Khan, Sallar UNSPECIFIED
Uncontrolled Keywords:	GraphWaveNet; Spatio-Temporal Graph neural networks; Feature-based attention; Incremental Update Edge; Traffic forecast; Computation efficiency; Robustness
Subjects:	Q Science > QA Mathematics > Electronic computers. Computer science T Technology > T Technology (General) > Information Technology > Electronic computers. Computer science H Social Sciences > HE Transportation and Communications > Urban Transportation
Divisions:	School of Computing > Master of Science in Data Analytics
Depositing User:	Ciara O'Brien
Date Deposited:	01 Jul 2026 11:20
Last Modified:	01 Jul 2026 11:20
URI:	https://norma.ncirl.ie/id/eprint/9433

Actions (login required)

View Item