서강대학교 전자공학과

Sogang Research Team Develops AltOPF for Fast Optimization of National Power Grids

A research team led by Professor Hongseok Kim at Sogang University, in collaboration with MIT, Argonne National Laboratory, and the Korea Power Exchange, has developed AltOPF, with the study accepted for publication in Engineering Applications of Artificial Intelligence. AltOPF alternates supervised learning on an auxiliary relaxed problem with constraint-aware unsupervised learning on the original AC optimal power flow problem. Tests included three benchmark systems and a real Korean 4,492-bus network, where raw predictions achieved a mean cost difference of 0.934% relative to MATPOWER reference solutions. In parallel-batch tests on the Korean network, prediction with power-flow projection took 58.63 milliseconds per sample, approximately 70 times faster than the conventional solver, excluding any subsequent solver fallback. On the 1,354-bus system, the method reduced label-generation time by 95.4% relative to full supervision, while its compact architecture reduced checkpoint size by 75% and peak training GPU memory by approximately 40% relative to the compared fully connected network. These results support rapid, resource-efficient operating-point generation for large fixed-topology grids, with constraint screening and conventional solver fallback providing additional operational checks.