Urban Transport Systems Optimization Through Real-Time Data Analytics

Abstract

Urban traffic congestion continues to increase travel time, fuel consumption, and environmental impact in modern cities, and Tbilisi is no exception: traffic intensity at signalized intersections varies sharply depending on time of day, direction, and the workday–weekend cycle. This study proposes an integrated, AI-assisted framework that connects real-time route-level traffic observation, short-term forecasting, and adaptive signal control within a single decision-support pipeline. Traffic data were collected at five-minute intervals for twelve origin–destination route pairs surrounding one critical intersection in Tbilisi using the Google Routes API, with automated acquisition and storage implemented through AWS Lambda, EventBridge, and Amazon S3, producing a structured dataset of 23,556 observations and 75 variables. A LightGBM regression model forecasts short-term delay at two horizons (t+15 and t+30 minutes), while a two-stage multi-class classifier separates normal from congested traffic states and, within the congested subset, distinguishes slow traffic from jam conditions. Forecasts are translated into operational signal-control actions through a rule-based decision layer and evaluated in a SUMO/TraCI microscopic simulation environment against a fixed-time baseline, a max-pressure-style pressure-adaptive controller, a bridge-model hybrid controller, and an experimental Deep Q-Network (DQN) reinforcement-learning controller. Results show that LightGBM achieved R² = 0.82 and R² = 0.77 at the two forecasting horizons, while the two-stage classifier reached 90.6% overall accuracy. In simulation, the pressure-adaptive controller reduced mean waiting time by approximately 33.6% relative to the fixed-time baseline, while the DQN controller achieved an average reduction of approximately 23.3% across multiple random seeds. The proposed framework demonstrates a reproducible, data-driven pathway from real-time observation to adaptive traffic-signal decision-making, without removing engineering judgment from the control loop.