Environmental Impacts of Autonomous Electric Vehicles from Different Charging Patterns

The shift to autonomous electric vehicles (AEVs) has the potential to dramatically reduce air pollution and greenhouse gas (GHG) emissions, largely due to a switch from petroleum fuels to low-carbon electricity. The extent to... [ view full abstract ]

The shift to autonomous electric vehicles (AEVs) has the potential to dramatically reduce air pollution and greenhouse gas (GHG) emissions, largely due to a switch from petroleum fuels to low-carbon electricity. The extent to which these benefits are realized could depend on how AEVs are adopted. Specifically, AEVs may be individually owned and used as personal vehicles are today, or personal vehicles could be replaced by publicly or privately-owned, on-demand AEV fleets. Variance in AEV charging and travel behavior in these two different models could results in different environmental profiles.

Most dispatchable electricity - electricity that can be produced on-demand - in the U.S. is generated at fossil power plants. The plant, or plants, that generate more electricity in response to increased demand are called marginal generating units (MGUs). Models that predict MGUs generally fall into one of two categories: regression-based or unit-commitment economic dispatch. The first category can account for effects (e.g. imperfect information) that are ignored in the second by examining real-world behavior. However, models that simply regress on historical behavior are ill-suited to making predictions about future grid conditions. Applying machine learning techniques to energy sector analyses represents a pathway for potential improvements in the prediction of MGU behavior, and understanding the environmental impacts of energy transitions. This area of research is especially important as researchers and policy makers try to predict the economic and environmental impacts of vehicle electrification, demand-response management, and large scale deployment of variable renewable generating sources like wind and solar.

Using a combination of supervised and unsupervised learning algorithms with nine years of hourly generation data from Texas, we train a model that predicts how much a group of power plants will increase their generation in response to changes in demand and wind generation at different grid conditions. We group similar power plants into clusters to increase the accuracy of our predictions, and assess the effect of this clustering on uncertainty in air emissions. Outputs from this model inform the environmental impacts of different AEV use and charging strategies.