Research Highlights

Hongseok Kim's research team, published a paper in an international journal in the field of artificial intelligence application IFAC EAAI (JCR top 3.3%) accepted for publication 

Large-scale national power grid optimization AI technology AltOPF development

A research team led by Hongseok Kim, a professor in the Department of Electronic Engineering at Sogang University, has developed an AI-based optimization technology to aid in the economic operation of large-scale power grids. AltOPFThe technology was developed to improve the efficiency of power grid operations.. The research was conducted in collaboration with the Massachusetts Institute of Technology.(MIT), Argonne National Laboratory., The Korea Power Exchange.(KPX)and was jointly carried out with these institutions. IFAC Engineering Applications of Artificial Intelligence (Impact Factor 9.0, JCR a top international academic journal in the field of AI applications. 3.3%)and has been confirmed for publication in the journal..

To operate the power grid efficiently, we need to calculate the generation amount that satisfies various operating conditions, such as voltage and transmission line capacity, while meeting power demand and reducing generation costs, and this is a challenge. This is called the AC-OPF (Alternating Current-Optimal Power Flow) problem. As the expansion of renewable energy and the enlargement of the power grid increase the computational burden, existing AI technologies also face difficulties in securing a large amount of correct answer data and complying with the physical constraints of the power grid.

AltOPF relies on alternating between supervised learning to learn the answers to auxiliary problems with reduced computational burden and, unsupervised learning that reflects the physical laws and operational constraints of actual power grids. When the effect of auxiliary learning decreases, it switches to focus on the original problem and, gives greater weights to constraints that are repeatedly violated. This allows it to learn both economic efficiency and constraint compliance together with a small amount of labeled data.

The research team used the IEEE 162, 300, 1,354 standard test systems and domestic KPX actual systems(4,492lines) to verify performance. In the basic test conditions of the domestic system, AIdirectly predicted the results, and the generation cost showed an average difference of MATPOWER's benchmark solution and average 0.934% difference. Additionally, in parallel processing experiments including power flow correction, the average calculation time per sample was 58.63 msec, which is , compared to the existing optimization calculation of 4.10 seconds, approximately 70 times faster processing performance was recorded.

It also reduced the resources needed for learning. 1,354In the power grid system, the time it took to generate answer data was reduced from 18,432seconds to 848seconds, a decrease of about 95.4%. In a comparison of neural network structures in the same system, the capacity required to store the model and learning state was reduced by 75%, and the maximum GPU memory usage during learning was reduced by about 40%, thereby increasing efficiency for application to large power grids.

This study is significant in that it verified the speed and data efficiency of AI-based optimization together in a real domestic power grid scale. The research team applied constraint verification and power flow correction to the AI prediction results and also presented an operation plan that connects to existing optimization calculations if additional verification is not passed. This approach can be used as a power operation decision support technology that quickly reviews various operation candidates in a fixed power grid structure.

‣ learning strategy that alternates between auxiliary problems and original problems before focusing on the original problems AltOPFapproach. 

Research Information

Paper Title Alternative learning with semi-supervised relaxation for alternating-current optimal power flow

Journal Engineering Applications of Artificial Intelligence

Authors Hongseok Kim, Hien Thanh Doan, Keunju Song, Sungho Shin, Kibaek Kim, Youngmin Choi

Affiliation Sogang University, Massachusetts Institute of Technology, Argonne National Laboratory, Korea Power Exchange

Corresponding author Hongseok Kim, Professor | Department of Electronic Engineering, Sogang University

Publication date 2026 11 1

DOI 10.1016/j.engappai.2026.116099