A Hybrid GNN-IZR Framework for Fast and Empirically Robust AC Power Flow Analysis in Radial Distribution Systems

TL;DR

This work tackles the speed-reliability dilemma in AC power flow analysis for radial distribution systems by hybridizing a physics-informed Graph Neural Network with the non-iterative Implicit Z-Bus Recursive solver. A two-stage robustness trigger enables near-real-time predictions in the fast path while guaranteeing accuracy by routing stressed cases to the IZR fallback, achieving $0.00\%$ empirical failures on a stressed IEEE 33-bus test set. Ablation results demonstrate that combining physics-informed loss with Z-bus sensitivity features is critical, reducing failures from $98.72\%$ (data-only) to $13.11\%$ and enabling bimodal runtimes that balance speed and reliability. The framework offers a pragmatic pathway to scalable, trustworthy power-flow analysis in real-time operations, with potential extensions to uncertainty quantification and more complex networks.

Abstract

The Alternating Current Power Flow (ACPF) problem forces a trade-off between the speed of data-driven models and the reliability of analytical solvers. This paper introduces a hybrid framework that synergizes a Graph Neural Network (GNN) with the Implicit Z-Bus Recursive (IZR) method, a robust, non-iterative solver for radial distribution networks. The framework employs a physics-informed GNN for rapid initial predictions and invokes the IZR solver as a failsafe for stressed cases identified by a two-stage trigger. A failure is defined as any solution with a maximum power mismatch exceeding 0.1 p.u., a significant operational deviation. On a challenging test set of 7,500 stressed scenarios for the IEEE 33-bus system, the GNN-only model failed on 13.11 % of cases. In contrast, the hybrid framework identified all potential failures, delegating them to the IZR solver to achieve a 0.00 % failure rate, empirically matching the 100 % success rate of the analytical solver on this specific test set. An expanded ablation study confirms that both physics-informed training and Z-bus sensitivity features are critical, collaboratively reducing the GNN's failure rate from 98.72 % (data-only) to 13.11 %. The hybrid approach demonstrates a pragmatic path to achieving the empirical reliability of an analytical solver while leveraging GNN speed, enabling a significant increase in the number of scenarios analyzable in near real-time.

Figures (7)

• Figure 1: Hybrid GNN-IZR inference framework. The system defaults to a fast path (GNN+d-LSE) unless a robustness trigger detects a potential failure, in which case it invokes the robust IZR solver.
• Figure 2: Diagrams of the offline training phase. (a) The streamlined data generation flowchart using stressed cases. (b) The GNN architecture, consisting of three GATv2Conv layers with a residual connection and Z-bus sensitivity features.
• Figure 3: Training dynamics showing the raw validation loss components and the physics loss annealing weight ($w_{PQ}$). The physics loss (green) is introduced after a data-only pre-training phase (epoch 40) and its weight is gradually increased, leading to a stable convergence.
• Figure 4: Histogram of IZR solver convergence (number of series terms) for the 50,000 generated training scenarios.
• Figure 5: GNN performance analysis. Violin plot comparing the maximum power mismatch distribution for the purely data-driven GNN versus the proposed PINN + Z-bus model, showing a significant reduction in large mismatches.