Residual Correction Models For Ac Optimal Power Flow Using Dc Optimal Power Flow Solutions

This paper proposes a residual learning framework that bridges the gap between DC and AC Optimal Power Flow (OPF) formulations. While DC-OPF provides a tractable and simplified linear approximation, it neglects nonlinear AC effects such as voltage magnitude, reactive power, and losses, resulting in infeasible operating points. To overcome this limitation, we introduce a topology-aware local-attention Graph Neural Network that learns the residual corrections required to map DC-OPF solutions to their AC-feasible counterparts. The framework integrates DC features at both the local and global levels and employs a physics-aware loss function enforcing AC power-flow feasibility and operational limits. Using the OPFData benchmark across the IEEE 57, 118, and GOC 2000-bus systems, the proposed model achieves up to 25% reduction in MSE, 3× improvement in AC-feasibility distance, and order-of-magnitude runtime gains compared to AC-OPF solvers, while maintaining robustness under N−1 contingencies. The results highlight residual learning as an effective paradigm for accelerating AC-feasible optimal power flow computation and enabling near real time operational decision making.