AC OPF learning via modular decomposition and adaptation to topology changes

We propose a novel framework for efficiently solving the AC optimal power flow (OPF) problem under varying grid topologies. Traditional learning-based approaches to OPF often retrain entire models when the network topology changes, leading to high computational cost and limited generalization. The proposed approach exploits the fact that different components of the OPF problem respond to topology changes to varying degrees: device-level operating limits remain largely unaffected, the network equations are modified locally around a tripped line, whereas the resulting optimal operating point may change throughout the network. It consists of two complementary training phases. In Phase I (modular pretraining), the OPF learning task under the intact grid topology is decomposed into three sequential subtasks (voltage setpoint prediction, power generation estimation, and generation cost computation) representing distinct functional mappings. In Phase II (adaptive learning), these pretrained modules are jointly fine-tuned for new topologies using a composite loss that combines supervised objectives with physics-informed regularization for branch-flow accuracy, nodal power balance, and operational constraint satisfaction. Extensive evaluations on benchmark grids demonstrate accurate predictions for the intact grid as well as for grids with seen and unseen contingencies. In our numerical experiments, at most 60 N − 1 contingency topologies are used for adaptation, and the largest mean prediction error among the reported variables under unseen topologies is approximately 3.1%.