Jarah Alqasim

تخصص تحول الطاقة · دفعة 2026

Solver-Grade AC-OPF at Near-DC Speed: Residual Graph Neural Networks with Newton-Raphson Restoration

Optimal power flow (OPF) sets the operating point of an electrical grid. Its exact alternating-current (AC) form is nonconvex and too slow for near-real-time use. The linear direct-current (DC) approximation is fast, but it returns operating points that are not physically valid on the AC grid. We reproduce and extend a residual-learning framework that maps a DC-OPF solution to its AC counterpart with a topology-aware graph neural network (GNN), training on 270k solved instances per system from the OPFData benchmark. We then add a Newton-Raphson feasibility-restoration layer that makes the prediction satisfy the AC power-flow equations exactly. On the IEEE 118-bus system, the residual GNN reproduces the original study’s reported bus-voltage error to within 12%, and improves on our graph-convolutional and graph-attention baselines by about an order of magnitude. On two systems the original work never tested, IEEE 14 and IEEE 30, residual learning improves the voltage error by factors of 15 to 34 across the three architectures. The restoration layer then drives the average AC power-flow residual from about 10−2 down to about 10−6 p.u., which is indistinguishable from a classical solver, at roughly 3.3× the speed of an interior-point AC-OPF solve. Because we clip the predictions to generator limits, the same layer also improves the economics of the solution, lowering the dispatch error on every system and cutting the cost gap to the AC-OPF optimum from 0.60% to 0.06% on IEEE 118. Physics-restored residual learning therefore bridges DC and AC optimal power flow. It delivers solver-grade, AC-feasible operating points at near-DC speed.

اقرأ التقرير النهائي (PDF)