SE PhD Final Defense: Alexander Wasilkoff

SE PhD Final Defense: Alexander Wasilkoff

TITLE: Ellipsoidal Uncertainty Sets for Robust Unit Commitment Using Conformal Prediction

ADVISOR: Michael Caramanis (ME, SE)

CHAIR: John Baillieul (ME, SE, ECE)

COMMITTEE: Panagiotis Andrianesis (SE); Pablo Ruiz (SE); William Hogan, Harvard University

ABSTRACT: Increasing penetration of weather‑dependent renewable resources introduces substantial uncertainty into bulk power system operations, challenging traditional reliability practices based on deterministic forecasts. Robust optimization, and particularly the adaptive robust unit commitment (ARUC) problem, offers a structured approach to ensure feasibility under worst-case conditions, but its effectiveness depends critically on the construction of the uncertainty set. In this work, a data-driven method for forming ellipsoidal uncertainty sets tailored to renewable-rich systems is proposed. First, a covariance matrix from historical observations whose ex‑ante forecasts exhibit feature-based similarity is estimated. Then the ellipsoid is sized by determining its radius leveraging conformal prediction adapted to varying forecast magnitudes that provides system‑level statistical guarantees. The approach is evaluated on a sequential process (currently used in many electricity markets) of a day-ahead deterministic unit commitment followed by a reliability process, which, instead of using deterministic forecasts, is based on the robust counterpart of the ARUC problem with the derived ellipsoidal sets with linear decision rules. Numerical experiments using the RTS‑GMLC system with ensemble wind forecasts demonstrate the effectiveness of the approach and provide useful insights on the applicability of linear decision rules and the locational aspect of the system response to wind uncertainty. Additional work provides approximate methods to scale the methodology to large-scale systems. The uncertainty sets are used to find worst-case dispatch scenarios for renewable resources. These can be modeled as additional contingencies or robust constraints in the reliability process. The feasibility of the methods at a real-world scale was demonstrated in collaboration with a large-area power system operator.