GPU-Accelerated Physics-Informed Digital Twins for Real-Time State Estimation and Fault Localization in Distribution Grids

Abstract

This study addresses a persistent operational problem in distribution-grid management: utilities need fast, trustworthy state estimation and fault localization to support restoration and switching decisions, yet conventional analytics are often too slow for real-time use and can be difficult to trust under changing grid conditions. The purpose of the study was to evaluate whether GPU-accelerated, physics-informed digital twins can measurably improve real-time state estimation, fault localization effectiveness, and operational decision support in cloud and enterprise deployment contexts. Using a quantitative, cross-sectional, case-based design, data were collected from a sample of N = 210 respondents drawn from cloud and enterprise distribution-grid digital-twin cases. The key independent variables were GPU Acceleration Capability, Physics-Informed Modeling Strength, Digital Twin Fidelity, and Data Integration Quality; the dependent variables were Real-Time State Estimation Performance, Fault Localization Effectiveness, and Operational Decision Support Value, with an additional trust indicator for physics-consistency confidence. The analysis plan employed reliability testing (Cronbach’s alpha), Pearson correlations, and multiple regression modeling to estimate direct effects and the enabling pathway from acceleration and physics constraints to operational value. Reliability results indicated strong internal consistency across constructs (α = .83 to .91). Correlation results showed that GPU capability was positively associated with state estimation (r = .56, p < .001) and that physics-informed strength had an even stronger association with state estimation (r = .61, p < .001); state estimation was strongly related to fault localization (r = .63, p < .001). Regression findings confirmed that state estimation performance was significantly predicted by GPU capability (β = .29, p < .001) and physics-informed strength (β = .37, p < .001), explaining 54% of variance (R² = .54). Fault localization effectiveness was significantly predicted by state estimation (β = .41, p < .001), digital twin fidelity (β = .28, p < .001), and data integration quality (β = .19, p = .002), explaining 62% of variance (R² = .62). Operational decision support value was driven primarily by fault localization (β = .46, p < .001) and state estimation (β = .21, p = .004), with GPU providing incremental contribution (β = .14, p = .030), explaining 58% of variance (R² = .58). Trust results further indicated high physics-consistency confidence (M = 4.12, SD = 0.48). Overall, the study implies that utilities should prioritize physics-consistent modeling and GPU-ready architectures, while investing in data integration quality and twin fidelity to convert computational speed into reliable, actionable grid decisions.