Statistics Seminar: Learning under Constraints and Extremes: Methods and Applications in Energy Systems
Presenter: Yu Zhang, Associate Professor, ECE Department of UC, Santa Cruz
Description: Modern cyber-physical systems present statistical learning problems that deviate significantly from standard i.i.d. supervised settings. In particular, two challenges frequently arise: (i) learning under hard structural constraints, and (ii) learning under severe distributional imbalance and rare events. In this talk, I present two case studies from energy systems that illustrate these challenges and motivate new learning paradigms. First, I consider the problem of approximating the solution map of the AC optimal power flow (AC-OPF), a nonlinear and nonconvex optimization problem governing power grid operations. Rather than relying solely on labeled optimal solutions, we develop both unsupervised and semi-supervised physics-informed learning frameworks that incorporate equality constraints directly into the training objective via augmented Lagrangian formulations and implicit gradient estimation. These approaches enable data-efficient learning while maintaining physical feasibility, and can be interpreted as constrained function estimation where physical laws provide structural supervision. Second, I discuss short-term power outage forecasting under extreme weather conditions, where the data exhibit zero inflation, heavy tails, and strong temporal dependence. We propose a two-stage modeling framework that separates event occurrence and magnitude, combining calibrated classification with Tweedie-based regression to better capture rare but high-impact events. Together, these examples highlight a unifying theme: modern applications often require learning methods that effectively integrate domain structure while remaining robust to challenging data characteristics such as sparsity and extreme events. I conclude with a discussion of broader implications for scalable learning, uncertainty handling, and decision-making in complex systems.
About the speaker: Dr. Yu Zhang is an Associate Professor in the Department of Electrical and Computer Engineering at the University of California, Santa Cruz. He received his Ph.D. in Electrical and Computer Engineering from the University of Minnesota, followed by postdoctoral appointments at the University of California, Berkeley and the Lawrence Berkeley National Laboratory. Dr. Zhang’s research advances the resilience, efficiency, and sustainability of modern electric power systems through innovations in AI-driven optimization, machine learning, and dynamic decision-making. His work develops physics-aware learning methods, stochastic and robust optimization techniques, and cyber-physical coordination frameworks to support reliable grid operations under uncertainty. Recent projects include learning-augmented outage forecasting, planning for weather-driven grid hardening, and integrating large flexible loads such as data centers into market and operational strategies. Dr. Zhang has been recognized with multiple awards, including the 2025 Outstanding Young Investigator Award from the Energy Systems Division of the Institute of Industrial and Systems Engineers (IISE), the 2021 Early Career Best Paper Award from the INFORMS Energy, Natural Resources, and the Environment (ENRE) Section, and the 2019 Hellman Fellowship.
This seminar is hosted by Professor Allen Kei.