TSpace

Abstract (summary): Whether for environmental, conservation, efficiency, or economic reasons, developing next generation electric power infrastructure is critical. Temporally relevant, granular data from smart meters provide new opportunities for data-driven management of the power grid. New developments—for example, electricity markets with multiple suppliers, the integration of renewable power sources into the system, and spikier demand patterns due to, say, electric vehicles—create new challenges for efficient grid operation. Computer science is uniquely positioned to assist with increasingly sophisticated techniques for handling and learning from large amounts of data. The methods of game theory and multi-agent systems provide a natural framework for modeling the competing incentives of electricity market participants. This thesis focuses on the use of learning, optimization, mechanism design, and preference elicitation methods to coordinate electricity demand and supply while respecting the incentives of market participants. Specifically, we propose an approach where an autonomous agent acts on behalf of each household, coordinating with inhabitants to relay information and make decisions on their behalf about electricity consumption. We focus on three problems that arise in developing such agents: (i) how to coordinate consumers' electricity use, (ii) how to share the costs of consumption among households (via their agents), and (iii) how to gather consumption preference data from consumers. Chapters 3 and 4 focus on different aspects of the first two problems. Both use a matching markets approach. In Chapter 3, we focus on the impact of demand smoothness and peaks on the supplier’s cost, and in Chapter 4, on the impact of predictability. In both chapters, we develop new cost sharing schemes that are resilient to certain forms of strategic behavior on the part of the agents and that achieve strong performance in experiments. Chapter 5 studies the third problem. Motivated by control of heating and cooling systems, we present a new approach to preference elicitation, where the cost and accuracy of query responses is dependent on the user’s familiarity with the conditions specified in the query. We show that despite the theoretical difficulty in this setting, we can build solvers that perform well in practice.