Learning and Control for Energy Systems under Uncertainty

Thesis (Ph.D.)--University of Washington, 2020Future energy system operation faces various sources of uncertainties, including both supply-side uncertainties associated with the variability and intermittency of renewables, and demand-side uncertainties related to active load management and demand response. Due to the increasing uncertainty, power system operation is facing a growing risk of potential safety issues and economic loss. Research efforts on mitigating these uncertainties, therefore, are much-needed and of great importance. The goal of this thesis is to take a step towards future energy system design under significant uncertainties, from the algorithmic perspective. In particular, we leverage tools that interface between machine learning, optimization, and control to address three types of uncertainties: environmental uncertainty, model uncertainty, and uncertainty from users' interactions. The first part of the dissertation considers energy system control under environmental uncertainty, by focusing on battery control in providing pay-for-performance services, such as frequency regulation and renewable integration. We derive an optimal real-time control algorithm with accurate battery degradation cost modeling, that achieves near-optimal performance compared to the offline optima which have complete future information. The second part of the dissertation addresses energy system control under model uncertainty. Specifically, we propose a novel type of neural network architecture called input convex neural networks (ICNNs), for both system identification and controller design. We show that ICNNs significantly outperform existing purely data-driven or linear control methods with applications in building energy management and distribution system voltage control. Finally, we consider uncertainties from user interactions. We use the Cournot competition model to characterize the collective behavior of learning agents in energy markets. We prove the convergence of policy gradient dynamics to the Nash equilibrium and provide insights on the effect of pricing functions and information feedback to the convergence behavior