FALCON-UC: Feasibility-Aware Learning for Network-Constrained Unit Commitment

High commitment-prediction accuracy does not ensure that a learned unit-commitment schedule admits an operationally useful dispatch. We present FALCON-UC, which jointly predicts commitment and dispatch and incorporates direct-current (DC) network constraints through two distinct feasibility paths. During training, soft commitment probabilities define dispatch bounds, and a differentiable repair alternates power-balance correction, a power-transfer distribution factor (PTDF)-based branch-flow step, and box projection before loss evaluation. During validation and testing, deterministic commitment and dispatch heuristics reduce residual inconsistencies associated with minimum up/down-time, capacity, ramping, balance, and branch-flow constraints. On the evaluated 2383-bus instances, the complete FALCON-UC pipeline attains 99.43\% commitment accuracy, a 1.3417 MW dispatch mean absolute error, and a 0.7241\% mean absolute operating-cost difference relative to the solver reference. No dispatch-bound, branch-flow, or minimum up/down-time violations are observed at the stated tolerances and reporting precision; the maximum balance and ramping residuals are 0.0023 and 0 MW, respectively. End-to-end prediction and repair averages 0.0299 s per instance. In comparison, the unrepaired label-only endpoint exhibits severe worst-case violations despite 98.13\% commitment accuracy. These results evaluate the complete pipeline and do not isolate the individual effects of training-time and evaluation-time repair.