ST-LLC: A Large-Scale and Lightweight Model Collaboration Framework for Spatio-Temporal Traffic Prediction

Sensor-cloud technology is a key enabler for transportation of smart cities, providing large-scale traffic data for analysis in increasingly complex traffic systems. However, spatio-temporal traffic data poses significant computational challenges, making it critical to fully exploit such data and explore cloud–edge collaborative intelligence in data-driven scenarios. To address this, we propose a spatio-temporal large-scale and lightweight model collaboration framework (ST-LLC), which distributes computational workloads across the cloud and edge to achieve accurate traffic prediction. In this framework, a lightweight graph network deployed at the edge captures short-term dynamics, while dual auto-encoders in the cloud extract long-term traffic priors. These heterogeneous features are then integrated through a cross-feed attention (CFA) mechanism, and a large language model (LLM) is designed to transfer domain knowledge and enhance adaptability. We evaluate the proposed framework on three real-world public datasets. Extensive experimental results demonstrate that ST-LLC effectively integrates cloud–edge traffic sequence features and achieves superior prediction accuracy compared with state-of-the-art methods. Finally, we discuss practical challenges such as communication latency and data security, and outline future directions.