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

초록

The transition toward larger and more dynamic power systems motivates alternating current optimal power flow (AC-OPF) methods that can provide computationally efficient solutions for large networks. Conventional AC-OPF solvers may be too slow for high-frequency operational decision-making, whereas existing learning-based methods often face trade-offs among label requirements, constraint satisfaction, and scalability. This paper proposes AltOPF, a machine learning (ML) framework that alternates supervised updates based on an auxiliary thermal-limit-relaxed AC-OPF formulation with constraint-embedded unsupervised updates for the original nonconvex AC-OPF problem. Both stages update the same compact neural network, and augmented-Lagrangian multipliers adaptively emphasize persistent constraint violations. We further develop an optional AC power-flow projection that improves the physical consistency of the predictions. An idealized convergence analysis establishes ergodic convergence to Clarke stationarity for the corresponding stochastic subgradient procedure. The study considers the base AC-OPF problem under a fixed network topology; topology changes and contingency analysis are beyond its scope. AltOPF is evaluated on the 162-, 300-, and 1354-bus benchmark systems and a real 4492-bus Korea Power Exchange (KPX) network. On the in-distribution test sets, it achieves a cost difference below 1% relative to MATPOWER reference solutions and constraint satisfaction above 99%, outperforming six established baselines. The raw neural-network forward pass requires microsecond-level inference, while the compact architecture reduces checkpoint size by 75% and peak GPU memory by approximately 40% under identical batch settings. These results indicate that AltOPF provides a label-and resource-efficient approach for rapid AC-OPF candidate generation in large-scale fixed-topology systems.

키워드