A framework for continual learning in real-time traffic forecasting utilizing spatial–temporal graph convolutional recurrent networks
Abstract Traffic flow prediction is essential for enhancing urban mobility and facilitating effective transportation systems. The rapid increase in traffic data, along with the inherently dynamic characteristics of urban traffic, poses considerable challenges for traditional Machine Learning (ML) mo...
Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Springer
2025-08-01
|
| Series: | Complex & Intelligent Systems |
| Subjects: | |
| Online Access: | https://doi.org/10.1007/s40747-025-02049-7 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Abstract Traffic flow prediction is essential for enhancing urban mobility and facilitating effective transportation systems. The rapid increase in traffic data, along with the inherently dynamic characteristics of urban traffic, poses considerable challenges for traditional Machine Learning (ML) models, which often find it difficult to efficiently handle large-scale datasets. Although Deep Learning (DL) models demonstrate potential, their significant computational requirements and susceptibility to catastrophic forgetting limit their effectiveness in dynamic and real-time contexts, including traffic emergencies or evolving road networks. To address these challenges, this research presents an innovative framework known as the Continual Learning-based Spatial–Temporal Graph Convolutional Recurrent Neural Network (STGNN-CL) for persistent and accurate long-term traffic flow prediction. By utilizing techniques such as Elastic Weight Consolidation (EWC), Memory Aware Synapses (MAS), and Synaptic Intelligence (SI), the proposed model effectively addresses the issue of catastrophic forgetting while simultaneously enhancing its capacity to incrementally assimilate new traffic data streams. An advanced traffic pattern fusion strategy is introduced, utilizing the Kullback–Leibler Divergence (KLD) metric to measure traffic divergence across different scenarios. This approach improves the efficiency of the Continual Learning (CL) process by enabling the model to adapt to new traffic patterns more effectively over time. Extensive experiments conducted on the PeMSD3, PeMSD4, PeMSD7, and PeMSD8 datasets reveal the superiority of the proposed models, STGCN-EWC, STGCN-MAS, and STGCN-SI models achieve significant reductions in error rates compared to baseline methodologies. These results highlight the potential of continual learning in developing efficient, scalable, and adaptive traffic flow prediction systems, paving the way for advancements in transportation management and autonomous driving technologies. |
|---|---|
| ISSN: | 2199-4536 2198-6053 |