PDFormerFlow is an improved traffic flow prediction model that builds upon PDFormer and incorporates Flow-by-Flow modeling principles. PDFormerFlow inherits PDFormer's advantages in capturing dynamic spatial dependencies and long-range spatial dependencies and introduces new features to optimize traffic flow prediction performance. At the same time, PDFormerFlow draws inspiration from flow-by-flow models to better model the temporal dynamics of traffic flow. This implementation adapts the core strengths of PDFormer to describe the intricate relationships in flow-by-flow traffic matrices. The model can capture both spatial relationships (based on physical proximity of OD endpoints) and temporal relationships to effectively predict traffic flows.
🚀 Flow-centric architecture: Models traffic as origin-destination (OD) pairs rather than individual nodes
🌐 Dual relationship modeling: Captures both geometric (spatial) and semantic (pattern-based) flow relationships
⏱️ Temporal attention: Learns complex time-dependent patterns in traffic flows
📈 Curriculum learning: Gradually increases prediction horizon during training
⚡ Efficient implementation: Optimized for GPU acceleration
PDFormerFlow builds upon the original PDFormer architecture with these key enhancements:
Flow-based representation:
Treats each origin-destination pair as a separate flow
Models flow-to-flow relationships instead of node-to-node
Flow positional encoding:
Combines origin and destination node features
Creates unique embeddings for each OD pair
Flow relationship masks:
Geometric mask based on physical distance between flows
Semantic mask based on traffic pattern similarity
Efficient attention mechanisms:
Factorized attention over flows
Pattern-enhanced geometric attention
Using publicly available datasets to validate the proposed prediction method, such as the Abilene and GÉANT datasets. They provide the statistical traffic volume data of the real network traffic trace from the American Research and Education Network (Abilene) and the Europe Research and Education Network (GÉANT) .
| Topology | Nodes | Flows | Links | Interval | Horizon | Records |
|---|---|---|---|---|---|---|
| Abilene | 12 | 144 | 15 | 5 min | 6 months | 48046 |
| GÉANT | 23 | 529 | 38 | 15 min | 4 months | 10772 |
python=3.7.9 torch==1.7.0 tsai==0.3.0 numpy==1.19.2
for example for Abilene dataset
python SPtransf-train.py --model PDFormerFlow --dataset abilene --epochs 100 --batch_size 16 --pre_len 1 --rounds 10
[1] Jiang, J., Han, C., Zhao, W. X., & Cao, Z. (2023). PDFormer: Propagation Delay-aware Dynamic Long-range Transformer for Traffic Flow Prediction. AAAI. https://github.com/BUAABIGSCity/PDFormer
[2] Zhang, Y., et al. (2022). Bayesian Graph Convolutional Network for Traffic Prediction. IEEE Transactions on Intelligent Transportation Systems.
[3] Li, Y., et al. (2021). Dynamic Graph Convolutional Network for Traffic Prediction. IEEE Transactions on Intelligent Transportation Systems. [4] Weiping Zheng, et al. (2022). Flow-by-flow traffic matrix prediction methods: Achieving accurate, adaptable, low cost results, Computer Communications. https://github.com/FreeeBird/Flow-By-Flow-Prediction.