The uninterrupted growth in transportation activities has been exerting significant pressure on our socio-economics and environment in recent years. However, emerging technologies such as connected and automated vehicles (CAVs), transportation electrification, and edge computing have been stimulating increasingly dedicated efforts by engineers, researchers and policymakers to tackle these transportation-related problems, including those that are focused on energy and the environment. With the advancement towards vehicle connectivity and automation, vehicles can reduce the energy consumption, emissions and improve urban mobility and safety through environmentally-friendly eco-driving strategies, vehicle electrification, and driving coordination.
In this dissertation, we developed predictive models and trajectory planning algorithms using machine learning and optimization techniques to address four key challenge: 1) Driving in real-world scenarios with constrains and interaction from downstream vehicle's trajectory and traffic; 2) Extracting essential traffic information from sparse vehicle trajectory data in a connected vehicle environment; 3) Evaluating and quantifying the behavior of a complex Vehicle-Powertrain Eco-Operation System; and 4) Optimal scheduling and coordinating automated vehicles in terms of mobility benefits and energy savings considering the tradeoff between solution optimality and computational efficiency for online performance.
This research first starts with developing an electric vehicle energy consumption model based on real world data and integrating it into eco-driving algorithms considering the regenerative braking effect. By introducing a hybrid modeling approach which provides variables with actual physical meaning instead of the exhaustive method used in conventional data-driven approaches, we feature knowledge-driven variable selection and data-driven statistical synthesis together to further improve the estimation accuracy.
Many of the existing eco-driving algorithms, including the eco-approach and departure (EAD) algorithm, are not flexible enough to effectively handle customized powertrain characteristics, interaction with other traffic, road grade, and traveling with the presence of the downstream vehicles. Therefore, when considering the real-world deployment of the EAD application, it is beneficial to further explore the dynamic states from downstream vehicles and incorporate these into the trajectory planning process. A machine learning technique has been applied to the snippet of downstream vehicle trajectory (which may be obtained from onboard sensors, such as radar) to predict trajectories using real-world data. By integrating the prediction on future states of the preceding vehicle into the trajectory planner, the resulting enhanced EAD algorithm provides an eco-friendly speed trajectory in the presence of preceding traffic and queues at intersections with an additional 2 - 34 % energy savings and emission reduction compared to the EAD algorithms without prediction.
This dissertation also describes a technique to integrate vehicle dynamics and powertrain operations, using a comprehensive simulation study that was designed and tested for both electric buses and plug-in hybrid electric buses driving across total 11 signalized intersections in a test corridor. The overall simulation framework incorporates a two-layer vehicle optimal trajectory planning module that seamlessly integrates a graph-based trajectory planning algorithm and a deep learning-based trajectory planning algorithm while interacting with the environment calibrated using real-world data. It was found that deep learning-based EAD algorithms can achieve a good balance between solution optimality and computational efficiency. Besides, a dynamic queue prediction in the connected vehicle environment has been developed to better plan the bus trajectory for greater energy consumption. An average of 21.0% energy savings can be achieved across various traffic conditions when co-optimizing vehicle dynamics and powertrain operations. In addition, an extra around 7% energy efficiency improvement for a test PHEV bus was shown when introducing 20% connected vehicles in the network.
With a partially connected vehicle network, we not only improve the longitudinal control of the vehicle to obtain better energy efficiency but also extract essential traffic condition information such as lane-level traffic information that can be used for better lane selection. We developed a Lane-Hazard Prediction (LHP) application that can detect lane-level hazards effectively and efficiently. A machine learning approach was developed with feature extraction from the spatial-temporal domain to achieve sustainable high accurate lane-level prediction of a downstream hazard within tenths of seconds after it occurred by crowdsourcing sparse connected vehicle trajectories. The LHP application then guides the application-equipped vehicles with suggestions for proper lateral maneuvers far ahead of the hazard to avoid traffic jams. Results demonstrate that LHP-equipped vehicles may gain significant mobility and safety benefits without compromising the mobility and safety performance of the overall traffic under various traffic conditions and penetration rate of connectivity in the vehicle network.
Finally, a Bi-level Optimal Edge Computing (BOEC) methodology was developed under the fully connectivity vehicle network to maximize both the vehicle mobility benefit and energy saving by optimizing the vehicle coordination and motion planning. For the on-ramp scenario, the first-level edge computing is conducted in the roadside unit (RSU) that collects connected vehicle data, dynamically assigning each vehicle into an associated cluster group based on its state and potential merging conflict and periodically solved for the clustered vehicles with their optimal scheduling sequence and arrival time at the merge bottleneck point. Once the clustered vehicles have their assigned arrival time at the merge point, the second-level edge computing determines the optimal vehicle trajectory to guarantee vehicles meet the assigned arrival time with the minimum energy cost. It is shown that the computational cost of vehicle trajectory planning approaches can satisfy the objective of real-time performance with 63.4%-66.8% energy savings.