# A spatio-temporal self-supervised meta-learning network with dynamic graph learning for traffic flow forecasting

## TL;DR

This paper introduces a new deep learning model for predicting traffic flow that improves accuracy and adapts well to different urban environments.

## Contribution

The novel SSML-Net integrates self-supervised meta-learning and dynamic graph learning for traffic forecasting.

## Key findings

- SSML-Net outperforms existing models on PeMS datasets in traffic flow prediction.
- The model shows robustness and generalization in small-sample and cross-domain experiments.
- It reduces training costs while maintaining high prediction accuracy.

## Abstract

Accurate traffic flow prediction is essential for alleviating urban congestion, improving road network efficiency, and sup-porting sustainable transportation systems. Existing spatio-temporal graph neural networks, however, often struggle to capture dynamically evolving spatial dependencies, effectively integrate spatio-temporal features, and generalize across diverse traffic scenarios. To address the challenges of accurately modeling complex spatio-temporal dependencies in traffic flow forecasting, this study proposes a Spatio-Temporal Self-Supervised Meta-Learning Network (SSML-Net), the model comprises spatial- and temporal-level learners, and integrates self-supervised meta-learning with meta-learning-driven feature fusion and gated coupling mechanisms to enhance spatio-temporal interaction and generalization capabilities. Comprehensive experiments on the PeMS datasets demonstrate that SSML-Net clearly outperforms traditional statistical approaches, deep temporal models, and spatio-temporal graph-based networks. The ablation study validated the effective-ness and necessity of the model’s core components, whilst the small-sample experiments demonstrated its robustness and generalisation capabilities under extreme data conditions. Concurrently, we expanded our training and evaluation experi-ments based on METR-LR and Beijing traffic data, further validating the model’s exceptional transfer learning capability and generalisation performance in cross-domain traffic forecasting scenarios. This approach not only achieved superior prediction accuracy but also substantially reduced model training costs. These results indicate that SSML-Net adapts to varying data scales and dynamic urban scenarios, providing a robust, adaptive, and high-precision spatio-temporal traffic flow prediction framework.