RESILIENCE ASSESSMENT OF MODERN POWER GRID INFRASTRUCTURE AGAINST EXTREME WEATHER USING PHYSICS-INFORMED NEURAL NETWORKS

Abstract

The frequency and severity of extreme weather events have grown noticeably over the past two decades, and modern power grids have found themselves increasingly exposed to hurricanes, ice storms, heat waves, and wildfires that stress infrastructure well beyond its historical design envelope. Traditional resilience assessment methods based on deterministic simulation or purely statistical models struggle to capture the coupled physical and stochastic nature of these events, and they scale poorly to the large interconnected networks that define contemporary electricity systems. This paper investigates the use of physics-informed neural networks (PINNs) for resilience assessment of power grid infrastructure against extreme weather. The premise is that embedding physical laws directly into the loss function of a neural network combines the data efficiency of physics-based simulation with the speed and scalability of data-driven modelling, producing an assessment tool that can evaluate thousands of hazard scenarios in the time a conventional simulator would need for a handful. We designed a PINN framework that couples power flow equations with weather-driven component failure models and cascading outage dynamics, and evaluated it on a synthetic transmission network calibrated against a large regional grid with 4,120 buses and 5,830 transmission lines. Experiments were conducted across four hazard categories including hurricane, ice storm, heat wave, and wildfire. Results show that the PINN produced resilience metrics within 3.4 percent of a high-fidelity time-domain simulator while running roughly 82 times faster, enabling large-scale scenario ensembles that were previously impractical. Findings support PINN-based approaches as a promising foundation for planning-grade resilience assessment and for near-real-time operational decision support during evolving weather events.