Akademik Veri Yönetim Sistemi
Journal of Advanced Transportation, cilt.2025, sa.1, 2025 (SCI-Expanded, Scopus)
This study was conducted to ensure traffic continuity at an adaptive signalized intersection by developing a SUMO-based digital twin of the Heybe Intersection in Antalya, using real traffic data obtained from the Antalya Traffic Control Center (covering 165 days of observations). To address potential sensor failure scenarios, a solution integrating traffic forecasting and reinforcement learning was developed. After applying data cleaning techniques, multiple deep learning models were trained to forecast traffic volumes, and their outputs were used to generate an origin-destination (O/D) matrix that served as input to a Deep Q-Learning (DQL) control model. Three scenarios were evaluated in the simulation: (i) baseline adaptive signal control under normal operating conditions, (ii) the existing system under sensor failure reverting to a fixed-time plan, and (iii) the proposed DQL-based intersection management. Results demonstrated that, under sensor failure conditions, the DQL-based system achieved substantial improvements compared to the fixed-time baseline: the average delay was reduced by 61.3%, the average speed increased by 134.6%, and the level of service improved from E to B. These findings highlight the potential of integrating forecasting models with DQL to enhance the resilience of smart intersections against sensor malfunctions.