Team: Dr. Negar Soheili, Mahtab Danaei
Overview
This project develops machine learning models for predicting and mitigating infrastructure failures during extreme weather events. With climate change increasing the frequency and severity of natural disasters, critical infrastructure systems - power grids, water networks, transportation - must be designed to withstand and recover from disruptions.
Approach
We use graph neural networks (GNNs) to model interdependent infrastructure networks and predict cascade failure patterns. Combined with stochastic programming techniques, our framework identifies optimal pre-positioning strategies for repair crews and emergency resources. The models are trained on historical failure data and physics-based simulations of extreme weather scenarios.
Impact
Our framework has been applied to post-hurricane recovery planning, demonstrating a 30% faster restoration of critical services compared to conventional approaches. The work is being extended to address wildfire and flooding scenarios in collaboration with national laboratories.
Related Publications
Self-guided approximate linear programs: randomized multi-shot approximation of discounted Markov decision processes
P. Pakiman, S. Nadarajah, N. Soheili, and Q. Lin
Management Science
Self-adapting robustness in demand learning
B. Chen, P. Pakiman, S. Nadarajah, and S. Jasin
Manufacturing & Service Operations Management
Delayed allocation in marginalized flow models for weakly coupled Markov decision processes
S. Nadarajah and A. Cire
Management Science
Self-adapting network relaxations for weakly coupled Markov decision processes
S. Nadarajah and A. Cire
Management Science
Back to the future: Revisiting a pioneering approximation of average cost Markov decision processes using a multi-shot perspective
P. Pakiman and S. Nadarajah
Operations Research
Revisiting approximate linear programming: constraint violation learning with applications to inventory control and energy storage
Q. Lin, S. Nadarajah, and N. Soheili
Management Science