Tuesday, June 8, 2021 – 10:00am to 11:00am
Virtual Presentation – ET Remote Access – Zoom
PRIYA L. DONTI, Ph.D. Student https://priyadonti.com/
Bridging Deep Learning and Electric Power Systems
Addressing climate change will require deep cuts in greenhouse gas emissions over the next several decades, in which electric power systems will play a key role. In my thesis, I provide several directions for the principled use of machine learning for power systems forecasting, optimization, and control.
In the first part of my thesis, I present work on problems in the electric power sector for which I employ “classical” techniques outside of machine learning. Specifically, I focus on the estimation of marginal emissions factors, as well as of voltages on the electric power distribution system.
Motivated by insights from these projects, in the second part of my thesis, I focus on the design of deep learning methods that incorporate physics and domain knowledge, both from the power sector and more broadly. In particular, I explore how the design and use of implicit layers in neural networks can enable the construction of (a) decision-driven forecasts in the context of stochastic optimization, (b) provably robust deep reinforcement learning methods, and (c) fast, feasibility-preserving neural approximators for optimization problems. I also propose to explore the use of complex-valued graph neural networks, potentially in conjunction with implicit layers, for approximating certain optimization problems.
While part two demonstrates how power systems can yield interesting directions for machine learning, in the last part of my thesis, I aim to demonstrate vice versa how insights from machine learning can yield interesting directions for power systems research. Specifically, I show how combining insights from the areas of adversarially robust deep learning and implicit layers can allow us to address the problem of N-k security-constrained optimal power flow, which has seldom been addressed to date due to its computational intractability. I also propose further work that aims to use methods inspired by the implicit layers literature to conduct data vulnerability assessments for the power grid.
J. Zico Kolter (Co-Chair)
Inês Azevedo (Co-Chair, Stanford University)
M. Granger Morgan
Yoshua Bengio (Université de Montréal)
Zoom Participation. See announcement.