Integrating Data-Driven and Model-Driven Approaches for Traffic-State Estimation in Data-Deficient Areas

Abstract

Missing or incomplete traffic data caused by sensor malfunctions and the absence of detectors create data-deficient areas that hinder the efficient and safe operation of road networks. This issue is particularly acute on highly congested road segments where a high-resolution traffic state is essential to mitigate congestion and enhance safety. This study presents a traffic-state estimation model designed to function effectively in such data-deficient environments. The proposed attention-based model integrates model-driven and data-driven approaches, combining the former’s ability to infer unobserved states with the latter’s adaptability to diverse traffic scenarios. Microscopic traffic simulation was employed to generate physically coherent and high-resolution training data, incorporating realistic variations in origin–destination patterns and driving behaviors. The model learns both the temporal dependencies of traffic evolution and the spatial correlations among detectors through gated recurrent units (GRU) and attention mechanisms. Validation was conducted using detector and drone data collected from the Gyeongbu Expressway, one of South Korea’s most heavily traveled corridors. The model achieved a mean absolute error within 17 vehicles per lane for volume, 10 km/h for speed, and 6% for occupancy, successfully reproducing fine-scale traffic dynamics even where no detectors were installed. This research contributes to improving traffic-state estimation practices by demonstrating how a substantial amount of simulated data with simple calibrations offers a versatile model, particularly in areas where data are deficient.