Data-Driven Scalable AI Techniques for Addressing Problems in the Study of Smart Grids
The wave of grid modernization and climate change is rapidly changing the landscape of residential energy demands. For example, hotter summers imply increased use of A/C units, use of electric vehicles implies increase energy demands and use of rooftop solar implies local generation. A central question thus is to understand how energy is consumed at granular social, spatial, and temporal resolutions. Such an understanding can lead to better solutions to demand-response events, pricing strategies, examining equity issues in energy and so on. To answer these critical questions, several `Modeling & Simulation' solutions are appearing in the literature at a noteworthy rate. However, we observe some critical problems that need to be addressed, especially in the areas of modeling details & approaches, data quality, and robust validation. Due to these drawbacks, many public policy and social impact questions requiring detailed knowledge of the domain remain unexplored. In my work, I address each of these research gaps.
First, I develop comprehensive data-driven scalable AI framework for modeling energy demand at occupant/household level for a wide range of appliances. The resulting dis-aggregated energy dataset is the first comprehensive synthetic dataset for energy demand in the residential sector at national scale and household resolution for the U.S. We make extensive efforts to evaluate multiple facets of energy demand by considering external variables such as climate and intrinsic attributes of the energy profile such as load shape and magnitude. Second, we observe that, many models in the smart grid literature (specifically, agent-based models) are developed for solving the same problem but may vary in design or data. This raises a general question of how to compare models with similar goals. In this work, we develop a novel methodology for comparing models that are developed for the same domain (e.g. solar adoption), but may differ in structure of the model and/or the data sets (e.g. geographic region). This is achieved by learning response surfaces and using active learning to facilitate efficient comparisons of parameter spaces in high dimension. Finally, we demonstrate the value of our work by studying sustainability, policy, and social good questions in the energy domain (e.g. issues related to energy equity and climate change).
- Anil Vullikanti, Committee Chair, (CS, Biocomplexity/SEAS/UVA)
- Madhav Marathe, Advisor, (CS, Biocomplexity /SEAS/UVA)
- Samarth Swarup, Co-advisor, ( Biocomplexity/SEAS/UVA)
- Henning Mortveit, (ESE, Biocomplexity/SEAS/UVA)
- Madhur Behl , (Cs/SEAS/UVA)
- Sonia Yeh, (Chalmers University)