A hard-constrained NN learning framework for rapidly restoring AC-OPF from DC-OPF

This paper proposes a hard-constrained unsupervised learning framework for rapidly solving the non-linear and non-convex AC optimal power flow (AC-OPF) problem in real-time operation. Without requiring ground-truth AC-OPF solutions, feasibility and optimality are ensured through a properly designed learning environment and training loss. Inspired by residual learning, the neural network (NN) learns the correction mapping from the DC-OPF solution to the active power setpoints of the generators through re-dispatch. A subsequent optimization model is utilized to restore the optimal AC-OPF solution, and the resulting projection difference is employed as the training loss. A replay buffer is utilized to enhance learning efficiency by fully leveraging past data pairs. The optimization model is cast as a differentiable optimization layer, where the gradient is derived by applying the implicit function theorem to the KKT conditions at the optimal solution. Tested on IEEE-118 and PEGASE-9241 bus systems, numerical results demonstrate that the proposed NN can obtain strictly feasible and near-optimal solutions with reduced computational time compared to conventional optimization solvers. In addition, aided by the updated DC-OPF solution under varying topologies, the trained NN, together with the PF solver, can rapidly find the corresponding AC solution. The proposed method achieves a $40\times$ time speedup, while maintaining an average constraint violation on the order of $10^{-4}$ and an optimization gap below $1%$.

💡 Research Summary

The paper tackles the long‑standing challenge of solving the alternating‑current optimal power flow (AC‑OPF) problem in real‑time. AC‑OPF is a highly nonlinear, non‑convex optimization problem whose solution requires satisfying a large set of equality and inequality constraints (power‑flow equations, voltage limits, branch‑flow limits, generator limits, etc.). Conventional nonlinear programming solvers (e.g., Ipopt) can find locally optimal solutions, but their computational burden grows dramatically with system size, making them unsuitable for the sub‑second decision cycles required in modern electricity markets. In practice, system operators therefore rely on the much simpler direct‑current OPF (DC‑OPF) model for real‑time dispatch. However, the DC‑OPF solution is generally infeasible for the true AC physics, prompting a need for fast post‑processing methods that can restore AC feasibility and near‑optimality.

Key contributions

1. Hard‑constrained, unsupervised learning framework – The authors propose a hybrid architecture that does not require ground‑truth AC‑OPF labels. Instead, the DC‑OPF solution is used as a reference point, and a neural network (a fully‑connected feed‑forward network) learns only the residual correction of generator active powers, together with generator voltage magnitudes and the reference‑bus voltage. By focusing on the small difference between DC and AC solutions, the learning task becomes considerably easier than learning the full AC‑OPF mapping.
2. Optimization‑based feasibility restoration layer – The neural‑network output is treated as a fixed parameter in a secondary optimization problem that enforces all AC‑OPF constraints as hard constraints. The objective of this problem combines (i) the original generation cost and (ii) a projection loss (ℓ₂ distance between the NN output and the restored decision variables). Solving this problem yields a feasible AC‑OPF point that is as close as possible to the NN prediction.
3. Differentiable optimization layer via implicit function theorem – To back‑propagate through the restoration layer, the authors derive gradients of the optimal solution with respect to the NN inputs by applying the implicit function theorem to the Karush‑Kuhn‑Tucker (KKT) conditions of the inner optimization. This provides exact, efficient gradients that guide the NN toward regions of the input space that lead to feasible, low‑cost AC solutions.
4. Replay buffer for data efficiency – Because generating new DC‑OPF/AC‑OPF pairs is costly, a replay buffer stores past (load, DC‑OPF, AC‑OPF) tuples. During training, mini‑batches are sampled from this buffer, allowing the network to learn from a diverse set of operating points without repeatedly solving the DC‑OPF.
5. Extensive validation – Experiments on the IEEE‑118 bus system and the large‑scale PEGASE‑9241 test case demonstrate that the proposed method achieves an average speed‑up of roughly 40× compared with a state‑of‑the‑art AC‑OPF solver, while maintaining average constraint violations on the order of 10⁻⁴ and an optimality gap below 1 %. Moreover, when network topology changes (e.g., line outages), the updated DC‑OPF solution automatically reflects the new configuration, and the same trained NN can still produce high‑quality AC‑OPF solutions without retraining.

Technical details

• Neural network architecture: Input vector consists of the load vector (