Benchmarking Neural Training Approaches for DCOPF: Supervised, Hybrid Physics-Informed, and Unsupervised Physics-Informed

Neural networks offer a promising approach for accelerating solutions to the optimal power flow (OPF) problem. However, achieving predictions that are both optimal and feasible remains challenging; a trade-off typically exists between optimality and feasibility. Many approaches for training neural networks for this problem have been proposed. This paper benchmarks three neural training approaches for the DCOPF problem: a) a supervised approach, b) a hybrid physics-informed approach, and c) a fully unsupervised, physics-informed approach. All three approaches are evaluated under a unified pipeline across three IEEE test cases (case5, case57, case300), using consistent architectures, datasets, and metrics. Results show that the supervised approach offers good optimality but suffers from constraint violations. The unsupervised approach delivers relatively good feasibility but suffers in terms of optimality. The hybrid approach offers the most balanced trade-off between optimality and feasibility, making it a practical option.