AI-Integrated Traffic Information System Using Physics-Informed Neural Networks and GPT-4

Abstract

Urban traffic congestion has emerged as one of the defining infrastructure challenges of the twenty-first century. Existing traffic management solutions largely depend on rule-based systems and shallow machine learning models that fail to generalise across heterogeneous road networks and dynamic demand conditions. In this paper we present an AI-Integrated Traffic Information System (AI-TIS) that fuses Physics-Informed Neural Networks (PINNs) with the generative reasoning capabilities of GPT-4. The PINN component embeds the Lighthill- Whitham-Richards (LWR) continuum flow equations directly into the neural network loss function, thereby constraining predictions to be physically consistent even in the presence of sparse sensor data. GPT-4 acts as a high-level reasoning and natural-language interface layer that translates raw model outputs into actionable incident reports, route advisories, and operator-facing explanations. The integrated architecture is validated on real-world traffic datasets from three metropolitan corridors in India and China covering 680 km of roadway and 14 months of observations. Experiments demonstrate a 23.7 % reduction in mean absolute error for speed prediction compared to a LSTM baseline, a 31.4% improvement in congestion onset prediction lead time, and an average journey-time saving of 8.2 minutes per commuter per day when advisory outputs are acted upon. The qualitative evaluation of GPT-4-generated incident summaries achieves a human- judge coherence score of 4.41 out of 5.00.