Hurricane damage to electrical infrastructure creates immediate challenges for utilities: assessing damage scope across thousands of network components and scheduling repair crews optimally to restore service. Traditional approaches rely on computationally intensive simulations and optimization algorithms that may take hours to complete—time utilities cannot afford to lose during emergency response.
Researchers have developed an integrated two-stage deep-learning tool designed to accelerate this critical process. The framework combines damaged-line identification with repair-schedule optimization using neural network architectures tailored to each task.
The system was trained and validated using a comprehensive synthetic dataset based on the IEEE 9500-node test feeder, representing a realistic distribution network. The dataset encompasses 1,700 hurricane scenarios, each characterized by exposure features, grid metadata, fragility parameters, and detailed power-flow simulation outputs. This breadth ensures the models generalize across diverse damage patterns and network configurations.
Stage 1 focuses on identifying which lines sustained damage. The researchers benchmarked three architectures—multilayer perceptron (MLP), residual MLP, and GraphSAGE—ultimately selecting ResMLP for its superior performance. Stage 2 handles repair scheduling, comparing MLP, DeepSets, and Set Transformer approaches. The final pipeline couples ResMLP with Set Transformer, achieving a damaged-job F1-score of 0.920, indicating excellent precision and recall in damage identification.
Scheduling accuracy metrics are equally impressive: pairwise order agreement of 0.854 demonstrates the model correctly sequences repair priorities, while start- and end-time predictions deviate only 4.3 to 4.5 minutes from reference solutions generated by advanced optimization algorithms.
Though the tool inherits some Stage 1 classification errors, performance remains practical for decision support. Utilities can deploy this framework to generate rapid initial repair assessments, prioritizing crews to highest-impact areas while optimization continues in parallel. This hybrid approach balances speed with accuracy, enabling faster grid restoration during critical recovery windows.
