Learned Surrogates for AC Feasibility in Stochastic Unit Commitment: A Comparative Study With Conformal Guarantees

Stochastic unit commitment with AC network constraints is intractable for realistic systems. We benchmark machine learning surrogates for AC feasibility classification—Kolmogorov–Arnold networks (KANs), multilayer perceptrons (MLPs), gradient‐boosted trees and support vector machines (SVMs)—on three power systems (118–2848 buses) under two feature tracks: pre‐solve (load multipliers, renewable factors and contingency identity) and dispatch‐aware (adding generator setpoints). KAN, MLP and tree‐based models all achieve area under the ROC curve (AUROC) above 0.99 on the pre‐solve track. On the dispatch‐aware track, gradient‐boosted trees lead with AUROC 0.997, while KAN with weight decay reaches AUROC 0.992, a residual gap of approximately 0.005. Violation‐type error decomposition identifies which constraint categories drive false‐feasible predictions. Split‐conformal prediction with group‐conditional calibration provides distribution‐free coverage guarantees and flags uncertain points for full power flow verification. Speedups of three to five orders of magnitude over Newton–Raphson enable real‐time screening.