Ruben Glatt | 19-FS-030
The adoption of electric vehicles (EVs) by transportation network company (TNC) drivers has the potential to reduce emissions and increase energy efficiency in the transportation sector, but the limited range of EVs creates usability challenges for drivers. When deciding when and where to charge their vehicles, drivers must consider many evolving and uncertain factors, including current traffic conditions, the destinations of future trips, and the availability of vehicle-charging stations.
This study explored the feasibility of using deep reinforcement learning (DRL) techniques to optimize the vehicle charging behavior of TNC drivers who use electric vehicles. We used DRL to learn to estimate an optimal strategy in the form of an artificial neural network, which can be used as an onboard decision-support tool that recommends when and where to charge an EV based on the multidimensional state of local transportation and energy systems, as well as driver objectives. We showed that such a support tool could help reduce the energy costs and emissions of a TNC EV while providing a similar level of transportation services. We developed an agent-based simulation environment that incorporates models of transportation systems and power grids for training and testing EV driving and charging strategies, and demonstrated the feasibility of using a trained EV charging strategy to reduce the carbon footprint of commercial mobility services.
This work is relevant to Lawrence Livermore National Laboratory's energy and resource security mission research challenge, including applying high-performance computing (HPC) to achieve energy innovation. The transportation sector accounts for 28 percent of all greenhouse gas emissions in the U.S., and tools like the one we studied offer the potential to improve vehicle charging and driving behavior, thereby reducing carbon emissions in the transportation sector. Our work also explored using HPC capabilities to conduct large-scale simulations of intelligent and coordinated agents in transportation networks, which supports the Laboratory's core competency in high-performance computing, simulation, and data science.
In addition, this study aligns to the goals of the DOE Vehicle Technologies Office, including its Energy Efficient Mobility Systems Program, by demonstrating a technology that can increase the energy efficiency of new types of mobility services. The methods developed in this project also have the potential to increase the resiliency of the transportation system and the electric grid by enabling EVs to adapt and respond to systemwide and local disruptions in the electric grid.
Publications, Presentations, and Patents
Glatt, R., et al. 2019. "DMA (DriveMind AI)." Software/code in development. LLNL-CODE-797971
Pettit, J. F., et al. 2019. "Increasing performance of electric vehicles in ride-hailing services using deep reinforcement learning." Presentation at Workshop on Tackling Climate Change with Machine Learning, December 2019. LLNL-PRES-798189