TY - GEN
T1 - LLM-Agent-Assisted PE-GNN for Accelerated MISOCP Power Flow in Hybrid AC/DC Distribution Networks
AU - Ding, Yushuai
AU - Zhang, Yueqian
AU - Tian, Xuzhe
AU - Liu, Chang
AU - Zuo, Jiasheng
AU - Zhao, Tianyang
N1 - Publisher Copyright:
© 2026 IEEE.
PY - 2026
Y1 - 2026
N2 - The integration of power electronics into hybrid AC/DC distribution networks challenges traditional power flow calculations with non-convex convergence bottlenecks, while pure data-driven methods lack physical feasibility guarantees. To overcome this, we propose a model-data hybrid-driven framework. First, a mixed-integer second-order cone programming (MISOCP) model guarantees exact physically feasible solutions. Second, a physics-encoder graph neural network (PE-GNN) rapidly infers high-quality initial states, significantly accelerating the MISOCP solver while adhering to hard physical limits. Finally, a ReAct-based LLM agent translates natural language instructions into API calls for autonomous dispatch, visualization, and diagnosis. The proposed framework ensures efficient, strictly feasible, and highly interactive power flow analysis.
AB - The integration of power electronics into hybrid AC/DC distribution networks challenges traditional power flow calculations with non-convex convergence bottlenecks, while pure data-driven methods lack physical feasibility guarantees. To overcome this, we propose a model-data hybrid-driven framework. First, a mixed-integer second-order cone programming (MISOCP) model guarantees exact physically feasible solutions. Second, a physics-encoder graph neural network (PE-GNN) rapidly infers high-quality initial states, significantly accelerating the MISOCP solver while adhering to hard physical limits. Finally, a ReAct-based LLM agent translates natural language instructions into API calls for autonomous dispatch, visualization, and diagnosis. The proposed framework ensures efficient, strictly feasible, and highly interactive power flow analysis.
KW - hybrid AC/DC distribution networks
KW - large language model agent
KW - mixed-integer second-order cone programming
KW - physics-encoder graph neural network
UR - https://www.scopus.com/pages/publications/105045477441
U2 - 10.1109/EPSIC70071.2026.11590437
DO - 10.1109/EPSIC70071.2026.11590437
M3 - 会议稿件
AN - SCOPUS:105045477441
T3 - 2026 IEEE 3rd International Conference on Electrical Power Systems and Intelligent Control, EPSIC 2026
BT - 2026 IEEE 3rd International Conference on Electrical Power Systems and Intelligent Control, EPSIC 2026
PB - Institute of Electrical and Electronics Engineers Inc.
T2 - 3rd IEEE International Conference on Electrical Power Systems and Intelligent Control, EPSIC 2026
Y2 - 22 May 2026 through 24 May 2026
ER -