Global-Decision-Focused Neural ODEs for Proactive Generator Deployment and Power Grid Resilience

Shuyi Chen, Ferdinando Fioretto, Feng Qiu, Shixiang Zhu

Keywords: Outage forecasting, decision-focused learning, neural ODEs, generator deployment optimization, disaster recovery logistics

Abstract:

Extreme weather events such as wildfires and hurricanes increasingly disrupt power infrastructure, leading to widespread outages and straining supply chains for emergency response. Traditional predict-then-optimize (PTO) frameworks sequentially forecast disruptions and then use these predictions to guide logistics and resource allocation. However, such two-stage approaches often suffer from misaligned objectives, resulting in suboptimal or delayed supply chain interventions. We propose a unified, decision-aware framework—Global-Decision-Focused (GDF) Neural ODEs—that integrates outage forecasting with proactive grid resilience planning. By modeling the spatiotemporal dynamics of power outages using neural ordinary differential equations, our approach embeds optimization objectives directly into the learning process, enabling strategic deployment of mobile generators. This predict-all-then-optimize-globally (PATOG) paradigm ensures system-wide consistency in both prediction and decision-making.
Through experiments on real-world outage data and synthetic hazard scenarios, GDF demonstrates significant improvements in forecast quality, decision robustness, and recovery efficiency. Our results underscore the promise of integrated AI methods for resilient supply chain operations in power systems.

Submission Number: 27