Global-Decision-Focused Neural ODEs for Proactive Grid Resilience Management

TL;DR

The paper addresses the challenge of aligning predictive outage forecasts with system-wide resilience decisions under extreme hazards. It introduces the predict-all-then-optimize-globally (PATOG) paradigm and a global-decision-focused (GDF) Neural ODEs framework that jointly model spatio-temporal outage dynamics and globally optimized interventions. By training with a combined global decision regret and prediction loss, and by modeling outage evolution with SIR-like Neural ODEs, the approach yields coherent, proactive resilience strategies and improved decision quality, demonstrated on real Nor’easter data and synthetic scenarios. The results show notable gains in outage-consistency, SAIDI reduction, and computational scalability, with applications to mobile generator deployment and power line undergrounding, underscoring the method's practical impact for utilities facing hazard-driven disruptions.

Abstract

Extreme hazard events such as wildfires and hurricanes increasingly threaten power systems, causing widespread outages and disrupting critical services. Recently, predict-then-optimize approaches have gained traction in grid operations, where system functionality forecasts are first generated and then used as inputs for downstream decision-making. However, this two-stage method often results in a misalignment between prediction and optimization objectives, leading to suboptimal resource allocation. To address this, we propose predict-all-then-optimize-globally (PATOG), a framework that integrates outage prediction with globally optimized interventions. At its core, our global-decision-focused (GDF) neural ODE model captures outage dynamics while optimizing resilience strategies in a decision-aware manner. Unlike conventional methods, our approach ensures spatially and temporally coherent decision-making, improving both predictive accuracy and operational efficiency. Experiments on synthetic and real-world datasets demonstrate significant improvements in outage prediction consistency and grid resilience.

Figures (11)

• Figure 1: The number of outaged customers and the meteorological factors, including wind speed and turbulent kinetic energy.
• Figure 2: Overview of the proposed GDF framework. Given covariates $\boldsymbol{z}_k$ and initial states $\mathbf{S}_k(0)$ for all $K$ service units, a model parameterized by $\theta$ predicts the system states $\hat{\mathbf{S}}_k$ for all units. These predictions inform global decision-making, where optimal actions $\boldsymbol{x}^*(\hat{\mathbf{S}})$ minimize the global decision loss $g(\boldsymbol{x},\hat{\mathbf{S}})$. The framework optimizes $\theta$ by minimizing a global-decision-focused loss, regularized by a prediction-focused loss (e.g., MSE loss) to enhance predictive interpretability. Red arrows denote the backpropagation through $\nabla \ell_\texttt{GDF}(\theta)$, ensuring that the model learns both system-level decision quality and region-specific prediction accuracy.
• Figure 3: A synthetic example of the mobile generator deployment problem for a system with three cities and five generators ($Q_w = 5$). The $y$-axis represents the number of outaged households. In this example, the uniform travel time $\delta_t = 10$, transportation cost is set to $c = 400$, the customer interruption cost is $\tau = 1$, and the operational cost is $\gamma = 2$. Four methods are compared on out‑of‑sample data: the proposed GDF framework (regret = $183.47$), a two‑stage approach (regret = $488.45$), an online baseline with observation lag of 5 (regret = $2846.13$), and the optimal solution with the ground-truth. Details of the synthetic data setup are provided in Section \ref{['sec:dataset']}.
• Figure 4: The real outage and restoration trajectories in Indianapolis, IN and their predictions. Left: actual outages and the Neural ODE prediction; the right axis shows restored and unaffected customers for both ground-truth and Neural ODE prediction. The vertical dotted line marks the train–test split. Right: comparison of test-set active-outage predictions from the Neural ODE, RNN (and extended), and LSTM (and extended).
• Figure 5: Performance Comparison for Synthetic Mobile Generator Deployment: A detailed comparison of regret outcomes for GDF, Two-Stage, and Online methods under varying customer interruption costs ($\tau$), transportation cost factors ($c$), operational costs ($\gamma$), and numbers of generators ($Q_{w}$).