Skip to content
Open access

A Unified PINN Framework: Evaluating the Impact of Pre-Physics Projection Sub-Layers and Loss-Weighting Strategies on the Feasibility and Optimality of ACOPF Solutions

2026 · IEEE Access · Vol 14, pp. 157511-157538 · 0 citations · 48 references

Abstract

The Alternating Current Optimal Power Flow (ACOPF) problem is fundamental to secure and economical power-system operation, but its non-convex and nonlinear structure makes repeated real-time solution computationally demanding. Physics-Informed Neural Networks (PINNs) provide fast inference, yet soft physics penalties alone do not guarantee satisfaction of the AC power-flow equations and operational limits. This work develops a unified PINN framework based on an exact Nodal Injection Rectangular Admittance Rectangular Voltage (NIRARV) formulation that expresses ACOPF as a Quadratically Constrained Quadratic Program (QCQP). Using a common MLP backbone and physics layer, we benchmark multiple constraint-handling architectures, including a PINN baseline, Deep Constraint Completion and Correction (DC3), the Feasibility-Seeking Network (FSNet), a KKT-based variant, and Rahul’s PINN-KKT formulation, on 14-, 57-, 162-, and 300-bus PGLIB-OPF systems. We further conduct a controlled ablation comprising 160 independently trained models to isolate the effects of the Pre-Physics Projection and composite-loss weighting. The projection ablation shows that bounding the generation variables substantially reduces operational inequality violations in several test cases, although the improvement does not extend monotonically to equality residuals. The loss-weighting ablation demonstrates that emphasizing different loss components redistributes residuals among power-balance, operational-limit, and economic-objective terms rather than uniformly improving all criteria. None of the evaluated neural architectures or controlled-ablation configurations achieves complete ACOPF feasibility at the adopted $10^{-4}$ p.u. tolerance. Nevertheless, neural predictions can provide useful initializations for numerical feasibility restoration, with the largest observed mean end-to-end speedup reaching $2.225\times $ on the 162-bus system while recovering objective values numerically consistent with the corresponding cold-start IPOPT solutions. These results demonstrate that architectural bounding, loss weighting, objective quality, and residual feasibility must be evaluated jointly when assessing neural ACOPF methods.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.