CLSM: A Collaborative LLM-SM Framework for Road Network Congestion Prediction

Accurate and timely prediction of road network congestion is crucial for intelligent transportation systems (ITS) and urban mobility management. Traditional prediction models often struggle with the complexity, scale, and dynamic nature of real-world road networks. This paper proposes a novel Collaborative Large-Small Model (CLSM) Framework for road network congestion prediction, which leverages the complementary strengths of a powerful, general-purpose Large Language Model (LLM) as an intelligent orchestrator and interpreter, and a specialized, efficient Small Model (SM) for real-time, high-precision time-series forecasting. The SM, built upon a two-layer Gated Recurrent Unit (GRU) architecture, integrates multi-dimensional features including time, space, and road attributes, and employs a clustering-driven approach (DTW + K-Medoids) for pattern recognition. Our framework achieves a system-wide accuracy of 73.8%, demonstrating superior performance and strong generalization capabilities. The CLSM framework introduces a new paradigm for applying large models in the transportation sector, offering a path towards more intelligent, efficient, and scalable traffic management solutions.