Retrieval-Augmented Generation with Knowledge Graphs for Traffic Control Strategy in Highway Networks

Abstract Traffic control strategy generation is a core operational task for maintaining safety and efficiency in highway networks; decisions made by traffic control systems should be grounded in traceable evidence while meeting stringent requirements on timeliness and accuracy. However, traditional practice often remains inefficient and costly to maintain when integrating multisource, heterogeneous operational data. To address these challenges, this paper proposes a retrieval-augmented generation (RAG) framework based on a knowledge graph (KG), aiming to improve the accuracy and explainability of highway network control decisions by leveraging structured knowledge representation and a dynamic retrieval mechanism. This study makes two key contributions. First, it introduces an innovative, lightweight knowledge graph storage architecture based on attribute triples, enhancing retrieval efficiency through structured semantic associations. Second, it designs a chain-of-thought prompting mechanism that enhances the model’s reasoning capabilities, enabling an end-to-end mapping from demand input to control scheme generation. Experimental results showed that the proposed framework achieved high accuracy and reliability in real-world road network control scenarios, with a faithfulness score of 0.933, answer relevance of 0.848, numerical accuracy of 0.935, and scores of 1.000 for both recall and strategy accuracy. Additionally, ablation studies demonstrated the critical roles of the chain-of-thought and traffic knowledge modules. Removing these components respectively resulted in reductions of 38.2% and 6.5% in faithfulness, 66.7% and 27.8% in recall, and 66.7% and 33.3% in strategy accuracy. The framework’s modular design provides scalability by allowing flexible adjustment of the knowledge graph’s granularity and retrieval scope to accommodate varying management and control needs.