Degree

Doctor of Philosophy (PhD)

Department

civil and environmental engineering

Document Type

Dissertation

Abstract

Researchers continue to seek numerous techniques for making the transportation sector more sustainable in terms of fuel consumption and greenhouse gas emissions. Among the most effective techniques is Eco-driving at signalized intersections. Eco-driving is a complex control problem where drivers approaching the intersections are guided, over a period of time, to optimize fuel consumption. Eco-driving control systems reduce fuel consumption by optimizing vehicle trajectories near signalized intersections based on information of the SpaT (Signal Phase and Timing). Developing Eco-driving applications for semi-actuated signals, unlike pre-timed, is more challenging due to variations in cycle length resulting from fluctuations in traffic demand. Reinforcement learning (RL) is a machine learning paradigm that mimics the human learning behavior where an agent attempts to solve a given control problem by interacting with the environment and developing an optimal policy. Unlike the methods implemented in previous studies for solving the Eco-driving problem, RL does not necessitate prior knowledge of the environment being learned and processed. Therefore, the aim of this study is twofold: (1) Develop a novel brute force Eco-driving algorithm (ECO-SEMI-Q) for CAV (Connected/Autonomous Vehicles) passing through semi-actuated signalized intersections; and (2) Develop a novel Deep Reinforcement Learning (DRL) Eco-driving algorithm for CAV passing through fixed-time signalized intersections.

The developed algorithms are tested at both microscopic and macroscopic levels. For the microscopic level, results indicate that the fuel consumption for vehicles controlled by the ECO-SEMI-Q and DRL models is 29.2% and 23% less than that for the case with no control, respectively. For the macroscopic level, a sensitivity analysis for the impact of MPR (Market Penetration Rate) shows that the savings in fuel consumption increase with higher MPR. Furthermore, when MPR is greater than 50%, the ECO-SEMI-Q algorithm provides appreciable savings in travel times. The sensitivity analysis indicates savings in the network fuel consumption when the MPR of the DRL algorithm is higher than 35%. At MPR less than 35%, the DRL algorithm has an adverse impact on fuel consumption due to aggressive lane change and passing maneuvers. These reductions in fuel consumption demonstrate the ability of the algorithms to provide more environmentally sustainable signalized intersections.

Date

10-16-2018

Committee Chair

Wilmot, Chester

Available for download on Tuesday, December 31, 2019

Share

COinS