Title: Reevaluating Neural Network Width as a Proxy for Alternating Current Optimal Power Flow

Abstract: Current deep learning approaches for Alternating Current Optimal Power Flow (ACOPF) suffer from a lack of systematic protocols for defining architectural scale. This study addresses the fundamental question of the minimum neural network width required to approximate the ACOPF manifold with high fidelity through a constructive thought experiment. To this end, we propose Loss-Guided Neural Densification (LG-ND), an algorithm that incrementally identifies the necessary model capacity, expanding the network only when existing topologies cease to yield performance improvements. Experimental evaluations on various IEEE test systems demonstrate that LG-ND matches the performance of established baselines while utilizing up to ten times fewer neurons per layer. This reduction in architectural complexity is essential for enabling the formal verification processes necessary for safety-critical grid operations.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC