Eco-Driving with Deep Reinforcement Learning at Signalized Intersections Considering On-the-fly Queue Dissipation Estimation and Lane-Merging Disturbances

Eco-driving research has grown significantly over the past decade, increasingly incorporating real-world traffic and road conditions such as road gradients, lane changes, and queue effects. However, most existing studies that account for queue effects are limited to single-lane scenarios, without considering lane-merging disturbances, and can only estimate queue length or discharge time within restricted regions. To address these limitations, this paper proposes a novel deep reinforcement learning (DRL) based eco-driving algorithm that simultaneously considers on-the-fly queue dissipation time estimation and lanemerging disturbances. The approach integrates a practical and cost-effective navigation-app-based traffic data sharing framework with a data-driven dissipation time estimation model, enabling the reinforcement learning agent to continuously receive accurate modified reference speeds that reflect both queueing and merging vehicle effects. Five comprehensive case studies, benchmarked against conventional and state-of-the-art ecodriving methods, were conducted to evaluate the effectiveness of the proposed approach. Simulation results demonstrate that the proposed method consistently achieves the best energy performance across all scenarios, reducing energy consumption by an average of 37.5% compared with the Intelligent Driver Model (IDM) baseline.