Climate change is one of the most pressing issues of our time, requiring the rapid mobilization of many tools and approaches from across society. Machine learning has been proposed as one such tool, with the potential to supplement and strengthen existing climate change efforts. In this thesis, we provide several directions for the principled design and use of machine-learning-based methods (with a particular focus on deep learning) to address climate-relevant problems in the electric power sector.

In the first part of this thesis, we present statistical and optimization-based approaches to estimate critical quantities on power grids. Specifically, we employ regression-based tools to assess the climate- and health-related emissions factors that are used to evaluate power system interventions. We also propose a matrix completion-based method for estimating voltages on power distribution systems, to enable the integration of distributed solar power.

Motivated by insights from this work, in the second part of this thesis, we focus on the design of deep learning methods that explicitly capture the physics, hard constraints, and domain knowledge relevant to the settings in which they are employed. In particular, we leverage the toolkit of implicit layers in deep learning to design forecasting methods that are cognizant of the downstream (stochastic) decision-making processes for which a model's outputs will be used. We additionally design fast, feasibility-preserving neural approximators for optimization problems with hard constraints, as well as deep learning-based controllers that provably enforce the stability criteria or operational constraints associated with the systems in which they are deployed. These methods are directly applicable to problems in electric power systems, as well as being more broadly relevant for other physical and safety-critical domains.

While part two demonstrates how power systems can yield fruitful directions for deep learning research, in the last part of this thesis, we demonstrate vice versa how insights from deep learning can yield fruitful directions for research in power systems. Specifically, we show how methods inspired by the implicit layers literature can be used to assess policy-relevant inverse problems on the power grid. We further show how combining insights from implicit layers and adversarially robust deep learning can allow us to provide scalable heuristic solutions to two central problems in power systems – N-k security-constrained optimal power flow and stochastic optimal power flow – that have seldom been addressed at realistic scale due to their computational intractability.

Overall, this thesis demonstrates how bridging insights from deep learning and electric power systems can help significantly advance methods in both fields, in addition to addressing high-impact problems of relevance to climate action.

254 pages

Thesis Committee:
J. Zico Kolter (Co-chair, Carnegie Mellon University)
Inês Azevedo (Co-chair, Stanford University)
Jeff Schneider
M. Granger Morgan
Yoshua Bengio (Université de Montréal)

Srinivasan Seshan, Head, Computer Science Department
Martial Hebert, Dean, School of Computer Science

Return to: SCS Technical Report Collection
School of Computer Science

This page maintained by reports@cs.cmu.edu