ABSTRACT Power systems with rapidly growing variable renewables increasingly strain bulk transmission interfaces, yet operations still rely on static corridor ratings that ignore how secure transfer capability changes with operating conditions and dispatch decisions. This study investigates embedding a compact, reliability‐calibrated neural surrogate of Total Transfer Capability directly into mixed‐integer dispatch, enabling time‐varying and decision‐aware interface limits while preserving computational tractability. We generate alternating‐current secure labels via continuation or repeated power‐flow screening with contingency checks, construct features that remain affine in the dispatch variables, train a small rectified‐linear‐unit network, and encode it exactly in a mixed‐integer linear program with bound tightening and a quantile‐based calibration to enforce one‐sided reliability. On a provincial‐scale system with coastal renewables, pumped storage, and batteries serving inland load, the embedded surrogate increases feasible transfers, cuts renewable curtailment by about 31%, and lowers the total cost index by 1.8 points relative to static ratings, while maintaining sub‐gigawatt prediction errors and roughly 95% no‐overestimation coverage; alternating‐current back‐checks report no violations. These results show that replacing fixed interface limits with calibrated, decision‐aware ones closes a meaningful portion of the gap between offline transfer assessment and market‐time scheduling, improving renewable utilization and economic efficiency while keeping solution effort stable. The contribution is a conservative, operationally compatible learning‐to‐optimization pipeline together with system‐level evidence of value on realistic data.
Hong et al. (Wed,) studied this question.