Priya Donti

Carnegie Mellon University

Position: PhD Candidate
Rising Stars year of participation: 2021
Bio

Priya Donti is a Ph.D. student in Computer Science and Public Policy at Carnegie Mellon University co-advised by Zico Kolter and Ines Azevedo and a U.S. Department of Energy Computational Science Graduate Fellow. She is also a co-founder and chair of Climate Change AI an initiative to catalyze impactful work in climate change and machine learning. Her work focuses on machine learning for forecasting optimization and control in high-renewables power grids. Specifically her research explores methods to incorporate the physics and hard constraints associated with electric power systems into deep learning models. Priya is a recipient of the MIT Technology Review Innovators Under 35 award and best paper awards at ICML (honorable mention) ACM e-Energy (runner-up) PECI the Duke Energy Data Analytics Symposium and the NeurIPS workshop on AI for Social Good.

Incorporating Power System Physics into Deep Learning via Implicit Layers

Incorporating Power System Physics into Deep Learning via Implicit Layers
Addressing climate change will require deep cuts in greenhouse gas emissions over the next several decades including and especially within the electric power sector. Many strategies to decarbonize the power sector rely on integrating large amounts of time-varying renewable energy (such as solar and wind) which increases the speed and scale at which power systems must be managed. While this has prompted the use of machine learning (ML) methods most ML methods struggle to enforce the physics or hard constraints associated with the systems in which they operate — a critical failure mode that in the context of power systems can lead to blackouts. In my work I seek to design ML methods for power systems forecasting optimization and control that indeed respect the physics of the underlying systems in which they operate. In particular much of my work to date has focused on the design and use of implicit layers in the context of deep learning. Implicit layers are neural network layers representing implicit functions such as optimization problems or physical equations. My work has leveraged implicit layers to unlock several new paradigms in the context of electric power systems. These include (a) constructing provably robust control methods based on deep reinforcement learning (b) learning fast feasibility-preserving neural approximators for continuous optimization problems (c) creating end-to-end forecasting models that are well-tuned for the decision-making processes that employ them and (d) embedding physics-informed inductive biases into neural networks. Together these represent a new set of methodological tools for merging data-driven methods with physical knowledge with applicability both in the electric power sector and across scientific and physical domains more broadly.