Jean-Christophe Raymond-Bertrand
← Back to research

In-progress · 2026

Dispatch of Mobile Power Sources for Enhanced Grid Resilience using Ensemble Hurricane Forecasts

M. Goutham, J.-C. Raymond-Bertrand, D. Deka, H. Nagarajan, R. Bent, R. Kannan — in collaboration with Los Alamos National Laboratory, MIT, and Illinois Tech.

The problem

Hurricanes have caused roughly 1,755 U.S. power outages between 2000 and 2023, accounting for 80% of all outages. Hardening the grid with permanent infrastructure is slow and expensive. A faster, cheaper lever is operational resilience: making smart, real-time decisions about where to position mobile power assets before and during a storm. Mobile power sources (MPS), which include batteries on wheels (MESS) and mobile diesel generators (MG), can be trucked to threatened areas to keep critical loads energized. The question is: given a hurricane approaching with uncertain track and intensity, where should you send them, and when?

The main idea

NOAA's Global Ensemble Forecast System (GEFS) publishes multiple simultaneous hurricane track predictions — informally called a "spaghetti plot" of possible futures — every six hours as a storm approaches. Rather than picking one track and optimizing for it, we build a multi-stage stochastic program (MSSP) whose scenario tree is directly constructed from these ensemble forecasts. Each branch of the tree corresponds to a different forecast model. Terminal nodes are damage scenarios derived from component fragility curves, modeled as Weibull distributions mapping local wind speed to transmission line failure probability.

The framework has three stages:

  • Stage 1, immediate action. Adjust generator set-points, begin repositioning MPS before the next forecast update.
  • Stage 2, recourse. When the next forecast is issued, identify which model best predicted it and enact its policy
  • Stage 3, restoration. After landfall, with actual damage observed, dispatch repair crews and inject MPS power to minimize load shed.

Methods

  • Stochastic Dual Dynamic Programming (SDDP). The MSSP is solved via SDDP, a decomposition algorithm that approximates the future cost function with piecewise-linear cuts, making large scenario trees tractable. Implementation uses SDDP.jl in Julia.
  • DC power flow. Grid physics are modeled with a linear approximation of the true power-flow equations, enabling the SDDP decomposition while still capturing generator ramping limits, line thermal limits, and nodal power balance.
  • MPS routing constraints. Binary reachability matrices encode road travel times and permit delays, ensuring MPS repositioning decisions respect real-world logistics between forecast updates.
  • Data-driven scenario generation. Failure probabilities are computed from NOAA wind fields via Weibull fragility curves, then reduced to a tractable scenario set using a Monte Carlo sampling and MILP-based medoid selection procedure.
  • Validation. The framework is demonstrated on IEEE 39-bus and 118-bus systems relocated to the Gulf Coast, using real GEFS forecast data from Hurricane Laura (Category 4, August 2020).

Results

  • The MSSP policy reduces total load shed by approximately 3% on average over a wait-and-see baseline across 5,000 out-of-sample damage scenarios.
  • The advantage is even more pronounced in high damage scenarios. When the storm hits hardest, proactive positioning and pre-dispatch matter most, which is exactly when you want the policy to be best.
  • Using ensemble forecasts outperforms planning against only the mean forecast, which represents averaging out the possible trajectories. The mean track misses the spread of possible paths, which can threaten very different parts of the grid.