EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Graph-Based Spatio-Temporal Modelling for Complex Network Behaviour Prediction Using Attention-Driven Architectures

Bharat B. Pandey

Precision Journal of Information Technology & Scientific Innovation · 2026

DOI: 10.67887/pd2601202006 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Precision Journal of Information Technology & Scientific Innovation, 2026. Verfasst von Bharat B. Pandey. Die Arbeit nennt 70 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Introduction

Accurate traffic forecasting is a central requirement for intelligent transportation systems because urban road networks exhibit complex spatial interactions, non-stationary temporal patterns and dynamic congestion propagation. Conventional recurrent models can represent temporal continuity but usually ignore explicit road-network structure, while fixed-topology graph models may not fully capture changing sensor relationships.

Methods

This study presents a METR-LA evaluation of a static-adaptive graph attention Transformer architecture for traffic-speed prediction. The framework combines static-adaptive graph fusion, dual-branch spatial encoding, Transformer-based temporal modelling and sparsity-regularised graph learning. The static graph preserves stable sensor relationships, whereas the adaptive graph learns hidden data-driven dependencies through trainable node embeddings. Local graph aggregation captures neighbourhood-level traffic propagation, while global self-attention models non-local sensor interactions across the wider network. The temporal Transformer encoder then models the sequence-level traffic dynamics through multi-head self-attention, enabling multi-step forecasting over short- and medium-term horizons. The methodology follows a chronological METR-LA benchmarking protocol using training-only normalisation, sliding-window sample generation, fixed random seeds, saved-checkpoint evaluation and horizon-aware reporting at 15-, 30- and 60-minute horizons.

Results

Comparative analysis against persistence, recurrent, temporal convolutional and graph-based baselines is reported using repeated-run summaries, ablation analysis and cautious seed-level statistical testing.

Conclusion

The study presents a technically integrated and reproducibility-oriented framework for adaptive, graph-aware traffic forecasting in intelligent urban mobility systems.

Zitieren

APA

Pandey, B. B. (2026). Graph-Based Spatio-Temporal Modelling for Complex Network Behaviour Prediction Using Attention-Driven Architectures. Precision Journal of Information Technology & Scientific Innovation. https://doi.org/10.67887/pd2601202006

Vancouver

Pandey BB. Graph-Based Spatio-Temporal Modelling for Complex Network Behaviour Prediction Using Attention-Driven Architectures. Precision Journal of Information Technology & Scientific Innovation. 2026. doi:10.67887/pd2601202006

BibTeX
@article{Pandey2026,
  author = {Pandey, Bharat B.},
  title = {Graph-Based Spatio-Temporal Modelling for Complex Network Behaviour Prediction Using Attention-Driven Architectures},
  journal = {Precision Journal of Information Technology & Scientific Innovation},
  year = {2026},
  doi = {10.67887/pd2601202006}
}
Kostenlos starten →